AI Paper Digest

Tuesday, September 22, 2026

709 papers cs.AI · cs.LG · cs.CL ← 2026-09-212026-09-23 →

Jul Aug Sep

Highlights

Measuring the Checker: Mutation Analysis for GPU-Kernel Benchmark Oracles

Highlight HF pick · 2▲Large Language Models Mingzhe Du, Anh Tuan Luu, Dong Huang, See-Kiong Ng Benchmarks for LLM-generated GPU kernels decide correctness with a few random inputs and a loose floating-point tolerance, and their verdicts now feed leaderboards and reinforcement-learning rewards, yet there has been no way to measure whether hand-made patches to these checkers actually suffice. The authors apply mutation analysis as an adequacy metric: deterministic rules inject 10,303 compilable faults into verified CUDA implementations of 188 KernelBench problems, 7,384 with an independent kill witness, and any test protocol is scored by the fraction of faults it detects. The official check deterministically misses one in six witnessed faults (16.9%), with 78.6% of precision faults escaping versus 8.7% of arithmetic faults, and the metric explains the tolerance blind band, splits the gain of KernelBench-Verified into +4.0 points from hidden inputs and +4.5 from tighter tolerance, and exposes a published fuzzing recipe that rejects correct kernels 107 times. Optimizing test suites over the kill matrix reaches 98.0% detection with two inputs per problem, blindness grows with model scale across 48 whole architectures, and the dataset is released as KernelBench-M.

LLM-generated GPU kernels are judged correct by benchmarks like KernelBench using a handful of random inputs and a loose floating-point tolerance, and those verdicts now drive leaderboards and reinforcement-learning rewards, yet nobody could measure how much a checker actually catches. The paper adapts mutation analysis to graded numerical oracles: deterministic rules inject thousands of compilable faults into verified CUDA implementations, and any test protocol is scored by the fraction of provably detectable faults it kills.

  • 124 deterministic rewrite rules across six fault families (arithmetic, indexing, semantic, boundary, synchronization, precision) produce 10,303 mutants over 188 KernelBench problems, with NVRTC compilation at 84 ms per mutant instead of ~200 s, compiled-image hashing to drop equivalents, and a kill-witness search that admits only the 7,384 mutants some validity-gated input verifiably detects into the denominator.
  • The official five-seed check misses 16.9% of witnessed faults, and the misses are sharply skewed by family: 8.7% of arithmetic faults escape but 22.9% of boundary, 27.8% of synchronization, and 78.6% of precision faults do, driven by a tolerance blind band that widens with reduction size (an all-zeros softmax output passes at d = 393,216) and a measured validity ceiling above which harsher inputs falsely reject correct kernels through fp32 accumulation noise.
  • Auditing hardened checkers on a unified denominator shows KernelBench-Verified's +8.5-point gain splits into +4.0 from its four hidden input scalings and +4.5 from its tighter tolerance, its constant-scaling design leaves shape and indexing faults unreachable by construction, and a reconstructed Correctness Illusion-style fuzzer reaches 86.2% only by crossing the ceiling and rejecting correct kernels 107 times.
  • Greedy set cover over the kill matrix reaches 98.0% detection with two inputs per problem (94.8% on held-out mutants) versus 83.1% for the official five, and a knowledge-ladder experiment shows that handing an LLM test generator the fault taxonomy and ceiling nearly triples a fuzz baseline (21.6% to 61.0%) and beats showing it the concrete mutated source (57.3%).
  • At whole-architecture scale across 48 level-3 networks blindness grows (survival up to 1.7x operator level, near-opaque in deep homogeneous pipelines like VGG-19 at 90%), two problems prove unrefereeable because their own fp32 references violate the benchmark tolerance against fp64, and the authors caution that adequacy is relative to a fault model that omits tensor-core paths and multi-site faults, substrates are LLM-authored rather than sampled from submissions, and all results come from a single H100.

arXiv:2609.22220 PDF

Deep Persona: A Psychologically Grounded Architecture and Evaluation Framework for Role-Playing Agents and Simulations

Highlight HF pick · 9▲Agents Rotem Dror, Zohar Elyoseph, Yuval Haber, Elad Refoua, Oshrat Ayalon, Adir Solomon Persona simulation with large language models (LLMs) usually relies on shallow character descriptions that fail to keep behavior coherent over long interactions. Deep Persona organizes a character into three layers, observable expression, latent beliefs, and core motivational drives, and under principles of scripted determinism and bounded agency confines the model to a reactive engine driven by a structured internal script. A companion reference-free evaluation framework benchmarks dialogue naturalness against empirical human distributions using established clinical psychological instruments and adversarial stress-tests. LLMs achieve high pragmatic fluency but show systematic deficits in emotional expression and joint attention, and a case study of two personas finds the structured architecture yields interactions closer to human conversational behavior.

Prompt-based LLM personas drift, hallucinate, and break character over long interactions because they encode only surface traits. Deep Persona replaces flat prompts with a three-layer character specification (observable expression, latent beliefs, hidden motivational drives) elicited from domain experts, and pairs it with a reference-free, clinically inspired evaluation suite that tests whether agent dialogue is statistically distinguishable from human conversation.

  • Personas are built by interviewing a domain expert to fill an External Layer (enforceable style and tone rules), a Middle Layer (conditional rules that surface only when triggered, such as deflecting questions about the past until trust is built), and an Internal Layer (persistent motivational constraints the agent must never verbalize), with a control module that tracks turn counts and stage transitions and an optional embodied-expression module that emits bracketed nonverbal cues.
  • Evaluation adapts four dimensions from the ADOS clinical instrument into automated proxies (pragmatic fluidity via ROUGE-L echolalia and self-repetition, joint attention via tracking of newly introduced entities, affective congruence between speech and bracketed actions, and emotional diversity from lexicon counts), then compresses the profile into a Dialogue Naturalness Score (DNS) using Mahalanobis distance from a human baseline with a chi-squared test for statistical indistinguishability.
  • Applied to Role-Play, ABC-Eval, and CounselChat human-LLM datasets, existing LLMs show high pragmatic fluency (0.88 to 0.97) but weak and inconsistent joint attention (0.46 on open-domain ABC-Eval) and poorly calibrated emotional expression, with zero dialogues from the role-play and open-domain sets passing as human against the CounselChat expert baseline.
  • Two Gemini 2.5 Pro case-study personas (Sarah, a suicidal teen in a Hebrew clinical simulation, and Evelyn, a vaping teenager in a parent-training scenario) reached DNS of 0.92 and 0.85 under the combined human baseline and were statistically indistinguishable from human interaction in all dialogues, exceeding every evaluated human-LLM dataset, and both stayed in character under adversarial stress tests such as sarcastic threats and mid-conversation language switching.
  • The evidence is a narrow proof of concept: Sarah is a single 49-turn session, Evelyn's 16 dialogues were driven by the persona's own developer, there is no flat-prompt ablation to isolate the architecture's contribution, and the framework penalizes deliberate verbal-nonverbal incongruence that is often the point of a realistic simulated patient.

arXiv:2609.22255 PDF

OmniEdu: Open Foundation Models for Learning and Teaching

Highlight HF pick · 21▲Applications Hao Liang, Qihan Lin, Meiyi Qiang, Linzhuang Sun, Hengyi Feng, Mingrui Chen et al. Educational language models typically focus on either problem solving or tutoring, with training data organized by source rather than by capability. OmniEdu is an open family of K-12 foundation models whose instruction-tuning corpus combines over 100 educational resources with general instruction data, organized around four capabilities: subject competence, curriculum grounding, diagnostic reasoning, and pedagogical action and scaffolding. A pipeline of deterministic cleaning, semantic auditing and rewriting, quality scoring, token-budgeted diversity selection, and pedagogical instruction assignment yields 69,999 examples, and the authors fine-tune 4B, 9B, and 27B models. Education-oriented tuning consistently improves curriculum grounding, K-12 problem solving, and pedagogical tutoring across all model scales, with OmniEdu-27B reaching 85.89% on MathFish, 86.95% on EDUMATH, 78.74% in the Scaffold setting of MathTutorBench, and the highest Teaching average on LongTutor among evaluated models.

Educational language models tend to be trained either for exam-style problem solving or for tutoring dialogue, with data mixtures organized by source rather than by the behaviors a tutor needs. OmniEdu is an open family of 4B, 9B, and 27B K–12 models fine-tuned on a corpus deliberately balanced across four capabilities: subject competence, curriculum grounding, diagnostic reasoning, and pedagogical action and scaffolding.

  • A six-stage pipeline shrinks roughly 1.34M candidate examples from 100+ sources to 60,951 through decontamination, Qwen3.5-122B-A10B-FP8 keep/rewrite/remove auditing, GPT-5.6-Terra rubric scoring, and k-center greedy diversity selection over BGE-M3 embeddings under per-task supervised-token budgets, then adds 9,048 general examples and one of 20 pedagogical system prompts per example for a final 69,999 examples / 15.96M response tokens.
  • On curriculum grounding OmniEdu-27B reaches 63.12% EM / 76.69% F1 on K12-Bench (base 52.11/73.48), 85.89% on MathFish, and 86.95% MaC on EDUMATH, leading all evaluated models on the first two and trailing only Kimi-K3 (90.00%) on the third.
  • Tutoring shows the largest jumps: MathTutorBench Scaffold win rate rises from 20.42% to 75.79% for 4B and from 57.16% to 78.74% for 27B, LongTutor Evidence goes from 36.80% to 78.20%, and the 27B model posts the best Teaching average of 3.02, though Claude-Opus-5 still wins on Scaffold (87.89%) and Evidence (84.17%).
  • K–12 problem-solving gains are more modest, with 27B moving from 91.55% to 94.87% on GAOKAO-Bench and 46.04% to 57.76% on MDK12-Bench, while proprietary models keep a clear edge on EXAMS-V (Kimi-K3 87.29% vs 69.52%) and general benchmarks like MMMU-Pro improve slightly rather than degrade (64.97% to 67.98%).
  • The comparisons are against raw base checkpoints rather than instruct-tuned Qwen models, so a large share of the tutoring gains likely reflects generic instruction tuning, no ablations isolate the four capability categories or the pedagogical system prompts, and knowledge-state diagnosis on LongTutor remains weak at 54.04% even for the best model.

arXiv:2609.23088 PDF

One to More, More to One: Category-Aware Iterative Expert Training for Software Engineering Agents

Highlight HF pick · 13▲Multimodal Jie Zhao, Ziyu Jiang, Suhang Zheng, Minghui Shan, Xiaoxiao Xu, Lin Qu Open-vocabulary audio-visual event localization (OV-AVEL) grounds a text-queried event in time using video, audio, and language, and the available teachers differ in how reliably they mark temporal boundaries: on OV-AVEBench the configured visual teacher gives more reliable boundary cues than the configured audio teacher, which is still semantically informative. OV-OrthKD treats this as a supervision-placement problem, letting visual feature transfer shape a decision-aligned representation while audio transfer enriches an auxiliary subspace kept off the segment-logit path, with a text prototype anchoring seen and unseen category semantics and an orthogonality loss limiting overlap between the two teacher-specific projections. The student still fuses both modalities with query-aware fusion at inference, and it reaches 0.816 segment AP on OV-AVEBench, improving F1@0.5 over the official fine-tuning baseline by 2.7 points overall and 3.4 points on unseen categories, with path-assignment, role-swap, corruption, and transfer analyses supporting placement as a task-specific design axis.

Pooled reinforcement learning on repository-level software engineering tasks exhibits a "category see-saw": gains in one task category coincide with regressions in another while the aggregate resolution rate stays flat, hiding the redistribution. The authors split training by task category, iteratively develop one RL expert per category from a shared base, and distill the experts back into a single deployable policy without any external teacher model.

  • SWE Labeler assigns each task two hierarchical semantic axes (26 Task Type and 21 Repository Domain families, 119 and 108 fine-grained labels) plus three four-level scale axes, and a deterministic rule over Repository Domain collapses tasks into three routes: service/data/security, user-facing applications, and systems/tooling/runtimes.
  • Each category expert runs a Refresh–Repair–Expand loop that alternates long-horizon Agentic-miniRL (RLOO advantages, K1 reference penalty in the reward path, truncated importance sampling against the rollout policy, turn-aware loss reduction, 8 rollouts per instance) with Repair SFT on the expert's own verifier-approved trajectories, then re-probes an expanded pool to reselect the training frontier.
  • Label-routed multi-teacher on-policy distillation (MOPD) consolidates the three experts into one student on the student's own trajectories, with ReLU-gated reward extrapolation that keeps only each teacher's improving direction relative to the reference policy.
  • The final distilled policy, built on Qwen3.6-27B, reaches 58.04% on Pro-618 (an audit-filtered 618-task subset of SWE-bench Pro) and 59.00% on SWE-bench Multilingual, gains of +5.39 and +2.78 points over the base model, with progress also tracked via the minimum category gain and a "see-saw gap" metric.
  • The three-way category partition is coarse and hand-chosen, the evaluation drops 113 SWE-bench Pro instances via a public-report audit filter, and the pipeline requires three separate RL tracks plus a distillation stage, so the per-category comparison against Pooled and Balanced RL baselines is the key result to check in the full paper.

arXiv:2609.23377 PDF

Why Do Video Diffusion Models Violate Physics? Unveiling the Flaws in Attention Mechanisms

Highlight HF pick · 7▲Vision Yueyan Li, Haibo Wang, Caixia Yuan, Xiaojie Wang State-of-the-art video diffusion models produce visually convincing clips that often break physical laws, and rather than adding external priors the work looks for the cause inside the model. An interpretability study of the motion-planning process in text-to-video diffusion shows trajectories forming during early denoising in a first-shape-then-details pattern, and combining cross-attention trajectory patterns with causal head contributions isolates a subset of attention heads that drive motion planning. Self-attention analysis reveals that Rotary Position Embedding (RoPE) induces excessive spatial attention decay, causing candidate regions to lock prematurely into physically implausible positions and suppressing plausible trajectories in adjacent frames. A lightweight fix that scales the RoPE frequency across denoising steps reduces this decay, and both training-free and training-based experiments show improved physical commonsense in generated videos.

State-of-the-art text-to-video diffusion models such as Wan2.1-T2V produce visually polished clips in which objects bounce mid-air, float, or freeze, and prior fixes bolt on external physics priors or curated data without explaining why the failures happen. This work traces the "motion planning" stage of early denoising inside the model, attributes most failures to RoPE-induced spatial attention decay in self-attention, and fixes it by rescaling RoPE frequencies during the first few denoising steps.

  • Cross-attention maps from the video latent to the object token show multiple candidate positions per frame collapsing into a definite trajectory within about 5 of 50 denoising steps, and combining a convergence-speed metric with causal attribution patching isolates a small subset of heads whose zero-ablation collapses the trajectory, while heads that merely display a clear trajectory pattern can be ablated with no effect on motion.
  • Self-attention analysis reveals a "spatial anchoring effect": RoPE's long-range decay along height and width makes each region attend to the same spatial location across all frames, so once a few frames lock in early (one region reaching 0.63 mutual consistency at step 3), spatially adjacent candidates in neighbouring frames gain confidence and override physically reasonable but more distant ones, which is exactly how the basketball ends up stationary in mid-air.
  • The proposed fix multiplies the RoPE rotation angle on the height and width axes by a factor below 1 (default 0.75) for only the first 5 denoising steps, either training-free or with a LoRA fine-tune on the 80K-video WISA dataset using a timestep sampler biased toward the earliest 10% of timesteps.
  • On VideoPhy (343 prompts), the training-free modification lifts Physical Commonsense from 29.45 to 39.94 on Wan2.1-T2V-1.3B, and LoRA plus modified RoPE plus prompt refinement reaches 87.46 Semantic Adherence / 62.68 Physical Commonsense versus 80.47 / 44.31 for prompt refinement alone, with gains concentrated in the solid-solid and solid-fluid subsets.
  • The training-free variant needs manual tuning of the scaling factor, the fluid-fluid subset sees little or no benefit (semantic adherence even dips slightly training-free), the mechanistic analysis is built around a single 1.3B model and solid-object cases, and the 14B model is validated only qualitatively, with image-to-video and other architectures deferred to future work.

arXiv:2609.23658 PDF

Jev-Mem: System-One-Controlled Agentic Memory for Efficient AI Agents

Highlight HF pick · 5▲Agents Dongming Jiang, Yi Li, Bingzhe Li Agentic memory systems often put autoregressive large language model calls on the critical path for organizing, retrieving, and using memories, making memory operations expensive. Jev-Mem borrows the System-One/System-Two split from cognitive science: a fast System-One control plane governs memory typing and relational organization during construction and handles query routing, retrieval budgeting, graph traversal, candidate scoring, and adaptive stopping during retrieval, while a slower System-Two reasoning plane is invoked only for complex reasoning and answer synthesis over a structured multi-relational memory. On LoCoMo it reaches an LLM-as-a-Judge score of 0.777, an 11.0% relative improvement over the strongest baseline, while cutting memory construction time to 158 seconds, a 6.6 times speedup over the fastest competing system, and lowering average query latency to 0.93 seconds.

Agentic memory systems increasingly hand every memory decision (what type a memory is, how it relates to existing ones, which graph to search, when to stop) to an autoregressive LLM, which puts token generation on the critical path of both writes and reads. Jev-Mem splits the work along System-One/System-Two lines: a lightweight typed controller (Jev, from TypeSafe AI) answers batched yes/no and multiple-choice questions with probabilities for all high-frequency control decisions over a shared multi-relational memory graph, and the LLM is invoked only for final answer synthesis.

  • On the write path every observation is kept as a canonical node scored on four overlapping types (episodic, semantic, procedural, preference), deterministic vector, lexical, entity, and timestamp signals pick at most 10 candidate neighbours, and a single batched controller call judges semantic and causal links, while timestamps and shared identifiers create temporal and entity edges directly without inference.
  • Retrieval is a closed loop: the controller predicts per-view relevance for the semantic, temporal, causal, and entity graphs plus multi-hop depth and recency weight, splits a fixed expansion budget across active views, seeds traversal with reciprocal-rank-fused vector and keyword anchors, scores each discovered candidate on relevance, relation usefulness, novelty, and support, and stops once its own sufficiency, missing-evidence, and contradiction estimates clear thresholds or the predicted utility of further search falls below one.
  • On LoCoMo with gpt-4o-mini as the answer model, Jev-Mem scores 0.777 overall on LLM-as-a-Judge versus 0.700 for MAGMA (an 11.0% relative gain), with the biggest jumps on adversarial questions (0.962 vs 0.742) and open-domain questions (0.618 vs 0.517).
  • Memory construction takes 158 s against 1,044 s for Nemori (6.6× faster) and over 3,000 s for A-MEM and MemoryOS, and mean query latency is 0.93 s, 36.7% below MAGMA at 1.47 s and below full-context prompting at 1.74 s.
  • The evidence is thin in places: only one benchmark and one backbone are reported (the setup section mentions two benchmarks but shows one), no ablation isolates routing, scoring, or stopping, the strongest baseline MAGMA is the authors' own prior system, the controller is a commercial service whose outputs the appendix says are not calibrated probabilities, and the prose and Table 1 disagree on several per-category numbers, with the table showing Jev-Mem behind MAGMA on temporal questions (0.637 vs 0.650) where the text claims a tie.

arXiv:2609.23986 PDF

1% of Tokens Can Be Enough: On Gradient Estimation in On-Policy Distillation

Highlight HF pick · 3▲Large Language Models Huanxin Sheng, Zhiling Ye, Haonan Wang, Jian Wang, Jinjie Gu, Jian Kang Sparse on-policy distillation (OPD) sends teacher supervision to only a small subset of tokens in student-generated trajectories, but even useful guidance can produce a noisy update when its gradient is estimated from a single sampled next token. The authors analyze this estimation problem at a fixed prefix through information geometry and define an information-efficiency ratio (IER) from a signal-to-noise decomposition that characterizes relative gradient-estimation error under an optimal scalar baseline, with a candidate-set approximation that makes it usable for token selection alongside existing usefulness scores while keeping the sampled reverse-KL objective. On mathematical and medical reasoning tasks adding IER improves existing selectors, and sparse configurations using only 0.1% to 1% of tokens match or exceed full OPD without token selection.

Sparse on-policy distillation supervises only a small fraction of tokens in student rollouts, but existing selectors rank tokens by heuristic usefulness and ignore whether the one-sample reverse-KL gradient at that position is even a reliable estimate. The paper proposes the information-efficiency ratio (IER), a signal-to-noise score of the token-level gradient estimator derived in Fisher geometry, and uses it alone or fused with usefulness scores to select tokens under budgets as low as 0.1%.

  • At a fixed prefix, the reverse-KL gradient is an expectation over the vocabulary but is estimated from one sampled token; Theorem 1 shows that under the optimal scalar baseline the squared signal in the Fisher pseudoinverse norm equals the variance of the log-ratio log(p/q), and the residual noise equals a leverage-weighted second moment (leverage 1/p minus 1) minus that variance, so IER is simply signal over noise.
  • Because full-vocabulary statistics are too costly, IER is approximated on a candidate set built from the top-K logits of student and teacher plus the sampled token, then converted to a normalized per-batch rank and combined with usefulness scores via soft OR (IER-OR, either signal suffices) or soft AND (IER-AND, both must be high), while the sampled reverse-KL training objective is left unchanged.
  • The IER distribution is extremely heavy-tailed, with fewer than 0.1% of tokens scoring above 1, meaning estimated noise exceeds signal for almost every position, and as a standalone selector at a 0.1% budget it nearly matches full OPD on JustRL-Nemotron-1.5B to OpenMath-Nemotron-1.5B (58.9 vs 59.9 on AIME26) and reaches a HealthBench overall of 45.25 vs 45.77 for full OPD while selecting roughly one token per trajectory.
  • Fusing reliability with usefulness is where the headline claims land: TIP+IER-AND at 0.1% beats full OPD on all four math benchmarks for JustRL-Qwen3-4B to Qwen3-1.7B (e.g. 19.4 vs 14.4 on AIME25), Prefix+IER-OR lifts HealthBench overall/hard from 38.30/8.68 to 44.98/19.49 at 0.1%, and sweeps over 0.1% to 80% budgets show that adding more tokens is not monotonically helpful and can degrade results.
  • Gains are selector- and setting-dependent: Entropy-based combinations often get worse on math, both IER and the usefulness scores are approximations with no selection guarantee, results are on 1.5B to 4B students only, and the per-benchmark standard errors of 0.4 to 1 point mean many of the individual cell differences are within noise.

arXiv:2609.24432 PDF

RRSI: Regularized Recursive Self-Improvement of Agent Harnesses

Highlight HF pick · 65▲Agents Peng Xia, Rujun Han, Zifeng Wang, Yanfei Chen, Yufan Zhang, Yoonho Lee et al. An agent's harness, meaning the prompts, control flow, tooling, memory, and context management around a frozen backbone model, can be improved automatically by iteratively proposing and selecting edits, but this recursive self-improvement (RSI) tends to memorize the training tasks and lose its gains out of distribution. RRSI (Regularized Recursive Self-Improvement of Agent Harnesses) constrains both sides of the loop: the proposer works under a temporally annealed budget on how many edits a candidate may bundle and is steered toward unexplored trajectories, while the selector uses a critic to screen benchmark-specific proposals and a pruner to drop changes that are too small, too costly, or no longer useful. Across eight benchmarks spanning coding, agentic workspace, and engineering design tasks, RRSI gains up to 14.1 points on the evolved split and up to 4.7 points on five out-of-distribution benchmarks while the resulting harness runs on 30% fewer policy tokens than unregularized evolution.

Automated harness evolution, where an LLM proposer iteratively edits the prompts, control flow, tools, and memory around a frozen backbone and keeps whatever raises the score on a finite evolve set, is a form of recursive self-improvement that tends to overfit: large in-distribution gains shrink or reverse on held-out benchmarks. RRSI keeps every harness component editable but regularizes the search itself, borrowing L0/L1/L2 analogues to limit how much each round can change and which measured improvements are allowed to become permanent state.

  • On the proposal side, a cosine-annealed budget caps how many independent edits a candidate may bundle (shrinking from b_max to b_min over the run), the proposer conditions on a full log of past hypotheses, diffs, and outcomes so falsified ideas are not retried, and when progress stalls within the noise band part of the budget is reserved for components never yet touched.
  • On the selection side, a critic rejects diffs that encode task names, answers, or other evolve-set-specific logic before they are ever scored, candidates must clear a noise-adjusted floor (best score so far minus a δ estimated by repeatedly running the unchanged base harness), token-cost growth must satisfy ΔC ≤ β0 + β1·ΔS, and components that stop producing gains over a window are flagged for deletion.
  • Across eight benchmarks in coding, agentic workspace, and engineering design, RRSI gains up to 14.1 points on the evolve split (Terminal-Bench 2.1 with Gemini 3.5 Flash, 64.6 to 78.7) and up to 4.7 points out of distribution, with SWE-bench Verified rising 82.0 to 83.8 under Claude Opus 4.8 and Frontier-Eng gaining 4.3 Medal points (24.3% relative) despite never being scored during evolution.
  • On the agentic workspace suite the ranking inverts between splits: RRSI posts the smallest evolve-set gain of any evolved harness (90.5 on Harvey LAB versus 93.0 for Meta-Harness) but the only out-of-distribution average that clears the base harness by more than a point (43.6 versus 39.7 across JobBench, GDPval, and APEX-Agents), while AHE and TTHE finish below the harness they started from and fully unregularized evolution scores 92.8 in-distribution but only 40.3 out of distribution.
  • The final harness runs at 2.42M policy tokens per trial versus 3.80M for unregularized evolution (though still well above the base harness's 1.56M), and a harness evolved with Gemini 3.5 Flash still lifts the unseen Gemini 3.1 Flash Lite from 11.2 to 14.6 on Terminal-Bench 2.1, but the method only covers frozen-weight settings, introduces several hyperparameters (δ, β0, β1, budget schedule, stall and pruning windows) tuned on the evolve set, and the Harvey LAB evolve-set gain is a modest 1.1 points.

arXiv:2609.24972 PDF

Harness-Zero: Harness Distillation via Agent-as-Harness

Highlight HF pick · 10▲Agents Haoran Ye, Yuxing Lu, Haonan Dong, Zhaochen Su, Guojie Song Agent harnesses, the external systems mediating model-environment interaction, can boost performance substantially, but the best harness varies by domain, instance, and model, and the gains disappear whenever the harness is swapped out. Harness-Zero distills a domain- or instance-optimized harness into model weights through an agent-as-harness scheme: a harnessing agent, guided by the optimized harness, corrects the student's responses before execution within the target harness's action space, producing demonstrations that are then used for fine-tuning so the specialized harness can be dropped at deployment. Across knowledge work, tool use, and science domains, agent-as-harness outperforms code-as-harness for frontier models under the same evolved harness, and the base model's macro-average task success rises from 23.3% to 44.3% with the specialized harness removed, exceeding the 41.7% it reaches with that harness still attached. The method also recovers harness-induced behaviors absent from the base model, with 82.3% average recovery across 28 patterns.

Agent harnesses (the tools, middleware, skills, and memory wrapped around an LLM) can lift agent performance substantially, but the gains vanish when the harness is removed and the best harness differs by domain, task, and model. Harness-Zero distills an optimized harness into model weights via agent-as-harness: a harnessing agent, guided by a private reference harness adapted from the optimized one, reviews and minimally corrects the student's proposed responses before execution under a fixed minimal target harness, and the resulting trajectories are used for supervised fine-tuning so the specialized harness can be dropped at deployment.

  • The pipeline evolves a domain harness on training tasks (DeepAgents-style tools, middleware, skills, and memory), rewrites it into review guidance for a GPT-5.6 Sol harnessing agent, then runs a Qwen3.5-9B student under a mini-SWE-agent harness with a single Bash tool while the reviewer either passes each proposal or replaces it with the smallest coherent correction in the student's own action space, with no access to hidden solutions or the sandbox, before two epochs of LoRA SFT on the reviewed rollouts with reviewer-perspective reasoning masked out.
  • Without any training, agent-as-harness beats code-as-harness on frontier models across SpreadsheetBench Verified, AppWorld, and USPTO Retrosynthesis: 81.1% vs 78.1% averaged over six benchmark-model settings, against 68.6% for the bare harness, while review with an empty reference harness reaches only 69.2%, so the guidance rather than the review loop drives the gain.
  • After distillation and with the evolved harness, reference harness, and reviewer all removed, the 9B student's macro-average task success rises from 23.3% to 44.3%, exceeding the 41.7% the base model gets with the evolved harness still attached, whereas mounting generic DeepAgents or Claude Code harnesses on the base model actually hurts it (20.5% and 15.9%).
  • Ablations on USPTO show the supervision source matters more than demonstration quality: direct SFT on teacher rollouts, teacher or student rollouts under the evolved harness, empty-guidance review, or oracle-answer review yields only 3-15% pass@1 versus 30% for Harness-Zero, even though oracle access pushes collection success to 98.6%, and teacher trajectories under the evolved harness drop the student to 3% because it keeps calling tools that do not exist under the target harness.
  • The distilled model recovers 82.3% of 28 harness-exclusive behavior patterns on average, but the method needs a capable harnessing model (weaker reviewers become harmful), raises USPTO collection latency 2.4x, leaves deep domain knowledge partly untransferred (USPTO reaches 30% vs 38% with the harness attached), and cannot yet express mechanisms like context management as student responses.

arXiv:2609.24974 PDF

onPanda: Efficient Annotation of On-Policy Alignment Data for LLMs and Agents via Token-Level Correction

Highlight HF pick · 19▲Large Language Models Lei Yang, Mengyin Liu, Jia Wang, Hangyu Guo, Liang Zhao, Zheng Ge et al. Producing on-policy alignment data and agent trajectories by hand is slow because manual post-editing rewrites large spans of model output. onPanda is an interactive annotation tool built around token-level correction: the annotator reads a response, finds the first inappropriate token, either picks a substitute from the model's candidate tokens or types free-form text, and the system truncates everything after that point and regenerates from the corrected prefix, repeating until the response is satisfactory. Because nearly all tokens in the final response are still model-generated, the data stays close to the model's sampling distribution and suits on-policy supervised fine-tuning (SFT) and preference training, while the recorded corrections yield position-precise supervision and naturally paired positive and negative samples. A small controlled study suggests a 52% reduction in median annotation time over manual post-editing; the tool connects to external tools and harnesses for trajectory annotation, and the authors release the Panda-CVL dataset plus a token-level correction benchmark.

Hand-writing or post-editing model responses for supervised fine-tuning is costly and yields off-policy data, while preference ranking is cheap but coarse and useless when the model never samples a good answer. onPanda (StepFun) replaces both with a locate-correct-continue loop: the annotator finds the first bad token, swaps it for one of the model's top-20 candidates or types a replacement, and the model regenerates everything after it, so nearly all tokens in the final response remain model-generated.

  • The tool only needs an inference API that supports continuing from an assistant prefix and returning per-token logprobs, and every correction forks a node in an annotation tree, so intermediate versions become negatives and the companion onpanda library exports SFT samples, response-level preference pairs, and position-exact token-level correction triples without extra labeling.
  • Reasoning and tool-call messages are rendered through a response template into the model's native token stream, so special tokens and tool-call arguments are correctable inline, tools connect via MCP (with a harness_to_mcp adapter for Claude Code, Codex, and OpenClaw), and annotators can approve, fix, or reject calls before execution.
  • In a 3-annotator, 21-prompt Latin-square study with Qwen3.5-35B-A3B rollouts, onPanda cut median annotation time to 330 s versus 681 s for POTATO post-editing (51.5% less) and matched Argilla's 336 s, kept response perplexity within +0.86% of the rollout baseline versus +36.31% for post-editing, won 66.7% of GPT-5.5-judged pairwise comparisons, and scored lowest on NASA-TLX workload (3.1 vs 5.4 and 6.8).
  • Across roughly 132K production sessions yielding about 388K corrections, 97.0% of tokens in accepted responses were model-generated, 2.1% were picked from candidates, and 0.9% were typed, and on the released Panda-CVL benchmark (7,491 Chinese vision-language sessions) the best model, GPT-5.5, reaches only 17.09% F1 at judging acceptability and correcting the first error.
  • The controlled study is small (one rollout model, image-description prompts only, in-house annotators, LLM-as-judge with a modest human check), there is no controlled agent study, the on-policy benefit fades when the data trains a different model, and the paper does not yet show that token-level correction signals improve downstream training.

arXiv:2609.24983 PDF

Applications 209

LLM-Generated Feature Pools for Time Series Anomaly Detection

Youssef Attia El Hili, Malik Tiomoko, Corinne Ancourt A deliberately simple pipeline for univariate time series anomaly detection extracts a small pool of sliding-window statistics, scores each window with a transductive robust median-absolute-deviation (MAD) model, and selects a feature subset per domain on a held-out tuning split. On TSB-AD-U it reaches 0.529 per-series VUS-PR, above the best neural and statistical leaderboard entries and within 0.06 of the strongest pretrained foundation model. Ablations show the candidate feature pool dominates: changing the pool moves the score by 0.226 while switching selection strategy moves it by only 0.031. Prompting a multimodal LLM with in-context example windows to generate a pool per domain matches the hand-crafted pool, and selecting over the union of both lifts the pipeline to 0.588, matching the top leaderboard entry.

arXiv:2609.21801 PDF

EnterpriseVal: Quantifying the Efficacy, Reliability and Value of Generative AI in the Enterprise

Abbas Raza Ali, Muhammad Ajmal Siddiqui, Moona Zahid Frontier models match expert work on many tasks, yet most enterprise generative AI projects fail to show measurable business impact, which the authors attribute largely to measuring what a model can do rather than whether a specific workflow is fit, reliable, safe, and worth scaling under local data and controls. EnterpriseVal specifies a frozen socio-technical configuration of model, prompts, retrieval, tools, guardrails, and human oversight with an autonomy level and consequence tier, a metric catalogue spanning fidelity, utility, efficiency, reliability, assurance, and oversight, blinded expert grading scaled with calibrated LLM-as-judge scoring through prediction-powered inference, and an executable two-tier gate that maps metrics with confidence bounds to REJECT, CONDITIONAL, or SCALE decisions. In a pilot at a global bank, credit-memo drafting reached 88% citation precision and a 1.6% hallucination rate against gates of 70% and 5%, and procedure transformation cut analyst refinement effort from about 27.4 to 2.9 hours per document. The authors explicitly separate established results, pilot evidence, the proposed system, and open hypotheses.

arXiv:2609.21841 PDF

When Does Learning Beat Heuristics? A Case Study in Kubernetes Scheduler Score Plugins

Wang Xuying, Zhibek Sarypbekova cross-listed Production Kubernetes scheduler plugins that score candidate nodes are hand-tuned heuristics, and the study asks whether a scoring function learned from real placement decisions in a large production cluster trace can match or beat them. An external HTTP-backed scoring plugin for a widely used scheduler simulator serves two learned models, a Random Forest over engineered features and a graph neural network encoder over per-job task-dependency graphs, whose regression fit improves modestly but monotonically over four feature-engineering rounds to an R^2 of about 0.042. On Top-1 ranking accuracy, whether the model ranks the node the production scheduler actually chose highest, a trivial rank-by-free-CPU heuristic scores 74 to 84% against 65 to 66% for either learned model. The gap is attributed to objective mismatch, since both models were trained with pointwise mean-squared-error regression rather than a ranking objective, echoing the learning-to-rank literature and prior evidence that reinforcement-learning schedulers beat heuristics only when the training objective matches deployment; the paper also ablates the occupancy reconstruction needed to make offline trace data usable, analyses inference latency and serving memory constraints, and releases code and data pipelines.

arXiv:2609.22142 PDF

OpenBlock: Constructive and Verified Content Generation for Adaptive Tile-Matching Games

Jiang Jun Tile-matching puzzle games reach hundreds of millions of players, but the algorithms that choose which pieces to offer each turn are proprietary and no open platform exists for studying adaptive difficulty in the genre. OpenBlock pairs an always-available deterministic rule-based generator with an optional learned generator, both passed through a verification gate that uses exhaustive sequential-placement search to guarantee every delivered piece set is fully placeable, so the learned track can never break the feasibility guarantee. A self-play reinforcement-learning placement agent with auxiliary per-shape placeability supervision reaches a 35.6% win rate over 234,000+ episodes and serves as a diagnostic tool: at 70 to 75% board fill, 33 to 56% of long-bar pieces have no legal placement, while spawn difficulty is statistically indistinguishable between won and lost games, pointing to board-state degeneration rather than content difficulty as the cause of late-game failure. A 14-day online rollout with 48,000 players lifted day-1 retention by 1.8 percentage points and session duration by 7% over the rule track alone.

arXiv:2609.22177 PDF

WildfireSpreadBench: The Metric Decides the Model in Wildfire Spread Prediction

Arin Gopakumar, Marco Pannozzo Machine-learning models that forecast next-day wildfire spread are usually ranked by Average Precision (AP), which summarizes performance over all decision thresholds even though acting on a forecast requires choosing one. Five discriminative architectures and one generative model were benchmarked on WildfireSpreadTS with a shared evaluation pipeline and two input configurations, comparing AP against threshold-dependent F1 and IoU. Model rankings changed with the metric: the highest-AP model flagged 4 to 5 times the area that actually burned and ranked fifth of six on F1 and IoU, while models with more usable predictions scored 24 to 37 AP points lower. Three prediction profiles emerged (over-predicting, balanced, and under-predicting) that AP alone cannot distinguish, and expanding inputs from 7 to 23 channels shifted AP by only 0.03 on average against a 0.21 to 0.24 spread across architectures.

arXiv:2609.22191 PDF

The Role of AI in Online Reviews

Valeria Lerman, Oren Rigbi, Yaniv Dover Strategic use of large language models (LLMs) to write manipulative reviews is hard to measure because AI-generated text is rarely directly observable. The authors treat abrupt changes in LLM prices and capabilities as supply shocks and compare verified against unverified reviews across more than 13 million Trustpilot reviews to identify changes in platform activity tied to those shocks. After LLM supply shocks, unverified reviews shift toward greater negativity, with more 1-star and fewer 5-star ratings, an effect driven mainly by new model releases and concentrated among firms with the lowest and highest review volumes. Supply shocks also trigger short, concentrated bursts of review activity, which the authors read as evidence that generative AI is already reshaping reputation and competition on online platforms.

arXiv:2609.22198 PDF

H2LooP Telecom Model v1: From Telecom Comprehension to Autonomous Issue and PR Resolution

Amit Singh, Vedant Nipane, Mayank Goel, Pulkit Agrawal, Sairanjan Mishra H2LooP Telecom Model v1 is a pair of large language models fine-tuned for telecommunications: a comprehension variant for domain question answering and reasoning, and an agentic variant for telecom code generation, pull request resolution, and commits on production repositories. Training draws on curated corpora spanning 3GPP standards, O-RAN specifications, network telemetry, and real repository commits. The 31B-parameter comprehension variant scores 81.8% weighted average on GSMA Open Telecom Lite Pass@3 and ranks 5th on the community-run Open Telco AI Leaderboard, ahead of Claude Opus 4.6, GPT-5, Gemini 3 Flash, Grok-4-fast, and Kimi K2.5 on that leaderboard. The agentic variant improves AST similarity by 8.8% and location IoU by 20.0% relative to its base model on a proprietary telecom code benchmark while holding MMLU at 74.0% and BFCL v3 multi-turn function calling at 79.0%, which the authors present as evidence of no catastrophic forgetting.

arXiv:2609.22241 PDF

Connected Content Retriever: Dense Graph Edge Features Powering Pre-Ranking at LinkedIn

Akhilesh Gupta, Sudarshan Srinivasa Ramanujam, Chirag Bhanuprasad Mehta, Reshma Asharaf Beena, Dhritiman Das, Birjodh Singh Tiwana et al. Content from a member's connections and follows drives over 70% of impressions and engagement on the LinkedIn Feed, and the second-degree fan-out of posts that connections reacted to pushes the candidate index past one billion items, tens of thousands of which must be scored within a 120 ms p99 latency budget. Connected Content Retriever (CC Retriever) is a pre-ranking system that scores these candidates with a full deep ranking model on GPUs, built around a sorted-search GPU primitive that joins dense viewer-to-author graph affinity features with document-level features resident on the GPU in 5 to 10 ms. Moving scoring to GPUs enabled a 50x increase in ranking model parameters and a 2.5% lift in content time spent in online experiments, well above typical gains for Feed experiments. The paper details the economic-graph feature set, the scoring architecture, and the online serving stack.

arXiv:2609.22441 PDF

Apollo Restore: A Foundation LLM for Historical Greek Optimized for Fill-in-the-Middle Restoration of Ancient Greek Texts

Hope McGovern, Anna Dolganov, Samuel Belkadi, Guillaume Kunsch, Dimitris Vlitas, David A. Smith Fragmentary Ancient Greek papyri and inscriptions contain physical gaps, or lacunae, that scholars must reconstruct. Apollo Restore is a 24-billion-parameter model fine-tuned from Mistral Small with a fill-in-the-middle objective that restores missing spans without knowing their length, and is described as the first large-scale decoder model for any ancient Mediterranean language. On gaps of up to ten characters it places the correct restoration among its top twenty candidates for 80.6%, 54.6%, and 61.0% of documentary-papyrus, literary-papyrus, and stone-inscription lacunae, exceeding the strongest published models by 1.6x, 2.6x, and 1.4x, and the margin grows under a length-balanced metric that removes prior protocols' bias toward trivially short gaps. In a blind study, 20 experts preferred it to the strongest baseline and judged it at least as good as human restorations in 77% of cases, and it improved the published reading of a Herculaneum papyrus roll digitally unrolled after its training data was compiled.

arXiv:2609.22455 PDF

Autonomous Model Lifecycle Management for Digital Twin-Based Manufacturing Control

Zhengyang (Cissy), Gu, Thomas Cook, Fredaljohn Rohrbaugh, Joseph E. Hernandez, Chris Couch Manufacturing AI systems face continuous distribution shift from raw-material variability, ambient changes, and equipment aging, while model failures can cause physical damage and operators must trust the controller. The described closed-loop Cyber-Physical System, deployed in automotive manufacturing since 2023, manages per-product pairs of a sequence-to-sequence physics model (LPP) serving as a digital twin and a deep Reinforcement Learning control policy (LCP) trained against it; each retraining cycle pits multiple architecture families and RL algorithms against one another and promotes only the best-scoring candidate, with a Conductor orchestrator handling dependency-aware retraining and a Proportional-Integral-Derivative (PID) fallback. The policy score embeds an operator-trust gate penalizing deviation from established practice, and without it 23% of policies are rejected by operators despite passing accuracy thresholds. Across multiple facilities, controlled processes show 28 to 45% stability improvements over uncontrolled baselines with zero safety incidents.

arXiv:2609.22701 PDF

Analyzing Public Discourse on Urbanism: Topic Clustering, Sentiment Analysis and Retrieval-Augmented Generation using YouTube Comments

Jakob Morales, Monica Hegde, Fayeq Jeelani Syed Online discussion of walkability, transit, housing density, and street safety is abundant but unstructured, and existing city-evaluation tools ignore it. The authors build a pipeline combining geographic entity resolution, topic modeling, sentiment analysis, and Retrieval-Augmented Generation (RAG) over 22,788 chunks of YouTube transcripts and comments covering 309 North American cities, and report what happens when standard NLP components meet short, informal, geographically ambiguous text. A Twitter-tuned RoBERTa classifier beats a VADER lexicon baseline by 12.6 macro-F1 points but both collapse on the dominant neutral class, where annotators themselves reach only a Cohen's kappa of 0.53; dense retrieval beats TF-IDF at every cutoff, and video-level relevance proxies badly understate chunk-level precision. For groundedness, BERTScore proves unusable when a multi-sentence summary is scored against a single short comment, with scores nearly flat regardless of relevance, and ROUGE-1 groundedness is shown to be a paraphrase-driven lower bound rather than a hallucination rate.

arXiv:2609.22705 PDF

Scout: Open-World Species Recognition on the Edge

Mohammad Mehdi Rastikerdar, Hui Guan, Deepak Ganesan cross-listed Large vision-language models (VLMs) can recognize objects beyond a fixed class set but are too heavy for many edge devices, while offloading every image to the cloud consumes bandwidth and communication energy. Scout is an autonomous open-world recognition system for wildlife camera traps that intermittently calls a cloud VLM to identify new species and turns each one into a persistent, site-conditioned class in a compact edge model, starting from only the deployment location and empty site frames with no species list, human labels, or manual tuning. Across 30 camera-trap deployments in three regions on an NVIDIA Jetson Orin Nano, its accuracy stays within 0.1 to 2.5 percent of a model given a predefined species list, and on species outside its initial class set it reaches 53.7 to 59.1 percent accuracy versus 56.5 to 65.1 percent for full cloud offload while using 59 to 71 percent less deployment energy.

arXiv:2609.22897 PDF

LLMs Anchor on Chief Complaint and Fail to Integrate Evidence in Sequential Clinical Triage

Dipankar Srirag, Haokai Zhao, Ashutosh Kumar, Eleanor Hopper, Michael Dalton, Quoc Dung Nguyen et al. Emergency department triage unfolds turn by turn, yet existing evaluations of large language models (LLMs) for triage use completed retrospective records and report near-physician performance. The authors evaluate six LLMs at five sequential checkpoints on growing prefixes of nurse-patient conversations, using 425 LLM-generated and 50 physician-authored conversations labeled under the Emergency Severity Index (ESI), measuring agreement with quadratic weighted kappa (QWK). Every model drops from moderate-to-substantial agreement on complete records to fair-to-moderate agreement at every checkpoint, and controlled perturbations show predictions are anchored on the chief-complaint exchanges, with prompting interventions failing to help. Three expert clinicians reach a QWK of 0.887 to 0.929 on the same conversations while the best model reaches 0.295, models agree with each other more than with the ground truth so ensembling makes things worse, and the surprisal of the true label rises across checkpoints even though models extract relevant content from later turns.

arXiv:2609.22904 PDF

Enforcing Narrative Reliability and Epistemic Pacing in LLM-Driven Detective Games via Structured Knowledge Trees

Parsa Rahmati, Richard Zhao cross-listed Large language models (LLMs) enable open-ended dialogue in games, but their non-deterministic output makes it hard to preserve authorial control, factual consistency, and the intended order of information disclosure, which matters most in detective games where premature reveals or fabricated details break player progression. The proposed architecture pairs a Structured Knowledge Tree with a tri-agent LLM pipeline that separates knowledge retrieval, dialogue generation, and response verification so a virtual suspect discloses only what the current narrative state permits. Evaluated through a playable testbed, The Interrogation of Adrian Gale, and a user study, the system reduces critical hallucinations by 64.78% and entirely prevents premature narrative disclosure. The strict constraints introduce usability trade-offs around forced conversational reveals, but players report a clear sense of progression toward solving the case.

arXiv:2609.23043 PDF

CSC: Calibrated Simplicity for Conflict-Aware Social Bot Detection in the LLM Era

Yipeng Qian, Pengjie Zhao, Chaoxi Niu Large language models make social bots cheap to camouflage in text, producing modality conflict where an account reads as human in semantics but remains suspicious in graph structure, profile attributes, or cross-modal consistency, and recent detectors respond with graph-side complexity that the authors find is not the most reliable fix. CSC instead combines a simplified prototype-guided graph expert that drops unstable graph-side heuristics, simplex-constrained fusion that calibrates heterogeneous confidence spaces before late fusion, and a lightweight inconsistency expert that models disagreement between modalities. On TwiBot-22, TwiBot-20, and MGStBot-large the framework improves calibrated operating-point decision quality while staying competitive on external benchmarks, and a semantic-camouflage stress test that swaps selected bot text for matched human text sharply degrades the standalone text expert while graph and fused evidence stay stable.

arXiv:2609.23320 PDF

Co-occurrence Patterns of LoRA Adapters in Production Diffusion Model Inference Services

Tao Zhang, Bin Liao, Tao Zhou, Yanping Liu cross-listed Low-rank adaptation (LoRA) adapters are central to serving personalized diffusion models in the cloud, yet how adapters co-occur, contend for resources, and churn under real workloads had not been quantified. Using GenTD26, Alibaba's production diffusion inference dataset, the authors build an adapter co-occurrence network and characterize both its static structure and its temporal evolution. They find an extremely sparse network with heavy-tailed adapter usage, a 66.1 percent execution-latency overhead for the first adapter with diminishing marginal cost afterward, co-occurrence driven by shared base models (90.6 percent of multi-adapter requests use adapters from the same dominant base model), and a core-periphery structure with a 54.5 percent churn rate among the top-10 hottest models within a 12-hour window. A preloading strategy built on top-k co-occurrence statistics covers 81.0 percent of test-set co-occurrence pairs at k=3, giving a data-driven basis for cache preloading, adaptive scheduling, and GPU memory management.

arXiv:2609.23321 PDF

Rethinking Class Imbalance for Single-Cell Foundation Models: A Systematic Benchmark Across Architectures and Long-Tail Loss Functions

Zeyu Dong, Jiahui Zhong Single-cell foundation models such as scGPT, scBERT, and Geneformer reach up to 97.5 percent cell-type classification accuracy, but that aggregate can hide systematic failure on rare, often disease-relevant cell populations. The authors benchmark six long-tail loss functions (cross-entropy, weighted cross-entropy, class-balanced loss, focal loss, LDAM, and logit-adjusted softmax) across three backbones and three datasets (Multiple Sclerosis, Zheng68K, and human Pancreas) in 162 controlled training runs. Rare-class failure splits into two regimes visible in embedding geometry before any loss is chosen: classes a suitable loss can recover, and classes that remain linearly separable yet are absorbed into unrelated classes' neighborhoods under every loss and architecture. Among recoverable classes, reweighting efficacy is predicted by absolute training-set size rather than dataset share or imbalance ratio, class-balanced loss and LDAM are the most consistent choices, and logit adjustment trades rare-class precision for recall rather than improving both.

arXiv:2609.23325 PDF

Low-Rank Frequency Convolution and Noise-Range Augmentation for Real-Time Pitch Estimation on Edge Devices

Venkat Suprabath Bitra, Homayoon Beigi cross-listed Pitch estimation on an edge device must be small, robust to noisy input, and fast enough to emit one frame within the frame period. The authors factor their earlier Frequency Convolution Network (FrCN) into a low-rank form, cutting parameters 35.9 percent from 17,787 to 11,397 with no accuracy loss in domain or on two unseen corpora, and find that lowering the training noise floor from +6.02 dB to -20 dB raises RPA50 at -20 dB from 2.44 to 22.41, an effect about two orders of magnitude larger than any architectural effect measured, at a cost of 0.35 points of clean accuracy. The semi-orthogonal constraint from TDNN-F proves redundant with the bottleneck normalization and hurts when both are applied, and the factorization is 45 percent slower in eager PyTorch but 15 percent faster in a compiled kernel. A standalone 83 kB C kernel depending only on libc and libm beats OpenBLAS on all four CPUs tested, runs 3.8 times faster than ONNX Runtime and 13 times faster than PyTorch, and selects the same pitch bin as PyTorch on all 271,893 held-out frames per model.

arXiv:2609.23340 PDF

ARID: A Deployable Edge AI System for Structured Information Extraction from Industrial Maintenance Work Orders

Kuanlin Chen, Chen-Wei Kuo Industrial maintenance work orders often have to be processed offline on embedded hardware while downstream software needs predictable structured output. ARID (Aviation-inspired Routing for Industrial Deployment) extracts component, failure mode, symptom, and maintenance action into fixed-schema JSON on an 8 GB NVIDIA Jetson Orin NX, combining conservative dual-teacher filtering, noise-aware synthesis, a single routing decision per work order, 4-bit inference, and grammar-constrained decoding. From 2,326 unlabeled OMIn records it retains 716 training pairs plus 99 synthesized records targeting action extraction, and on 300 human-labeled records it reaches 84.8% token-F1 on the reference stack and 82.9% on the deployed Jetson, serving at roughly 5.3 to 5.7 seconds P50/P99 latency at 12.5 W. On zero-shot transfer to MaintNet, semantic F1 drops to 46.4% while parser success stays at least 99.8%, showing that output validity transfers but field semantics do not.

arXiv:2609.23582 PDF

ETH-TraceBench: A Large-Scale Event-Stream Benchmark for Ethereum DeFi under Temporal, Protocol, and Contract Shift

Kemal Kirtac, Carsten Maple Ethereum decentralized finance (DeFi) exposes a public, timestamped stream of transaction events, but repeated public symbols such as contract addresses and event topics invite machine-learning shortcuts. ETH-TraceBench evaluates DeFi event-stream representations under temporal, protocol, pool-infrastructure, and symbolic shift, drawing on a raw universe of 1.35 billion transactions and 5.01 billion log rows from January 2021 to December 2025 and a fixed 911,267-instance supervised sample trained on 2021-2024, selected on 2025H1, and tested on 2025H2. Simple models score up to 0.959 macro-F1 on the canonical decentralized-exchange test, but macro-F1 drops to roughly 0.74 to 0.79 under protocol novelty, with Uniswap v4 and Ekubo v1 markedly harder, and masking emitter and topic identity lowers liquidation macro-F1 to 0.774. A standard Transformer over log-ordered events does no better than a deterministic shuffle of the same events, indicating that high aggregate scores can arise without real chronological modeling, so the benchmark treats hard transfer and controlled-input conditions as the main target.

arXiv:2609.23659 PDF

TriFleetRCA: On-Premise LLM Root Cause Analysis for Kubernetes

Rohit Patel, Susil Kumar Mohanty, Jeenal Chaudhary cross-listed Diagnosing faults in a remote Kubernetes cluster means sifting pod logs, events, and cluster objects, and many operators cannot ship production logs to a hosted model. TriFleetRCA runs entirely on one workstation GPU with Qwen2.5-14B-Instruct, gathers evidence at pod, namespace, or cluster scope, ranks it by template de-duplication followed by BM25, screens runbooks through an ingest guard, and returns a root cause with supporting evidence lines. On a live cluster with four injected faults over 100 analyses, hit rates were 0.85, 0.90, and 0.95 across the three scopes, de-duplication before ranking raised the hit rate from 0.75 to 0.90 at equal token cost, a poisoned runbook instructing namespace deletion was rejected by the guard every run, and one fault was diagnosed correctly but cited incorrectly in every trial, a failure that accuracy alone would hide.

arXiv:2609.23766 PDF

From UNDRR Reports to Event Records: Schema-Constrained LLM Extraction of Georeferenced Disasters

Camilla Andreozzi, Phuong-Anh Nguyen-Le, Zhijing Jin, Revati Mani Disaster-risk-reduction archives describe hazard events in prose that structured databases such as EM-DAT cannot ingest directly. The authors build a large language model (LLM) pipeline that emits candidate georeferenced event records under a controlled hazard vocabulary and fixed schema while retaining supporting evidence, and run it over 10,000 documents from PreventionWeb, the knowledge hub managed by the UN Office for Disaster Risk Reduction (UNDRR), yielding 3,572 records across 24 hazard types with 81% of location mentions resolved to OpenStreetMap geometries. On a stratified reference set, GPT-5 reached 86.0% pooled attribute F1 versus 44.2% for a spaCy gazetteer baseline, and GPT-5.4 ranked highest of ten LLMs at 86.6%, though verbatim evidence occurrence fell from 72.0% to 47.2% between the two. The evaluation pools hazard families, locations, and years within documents without checking assignment to individual events, and the paper documents production failure modes and automated compliance checks.

arXiv:2609.23853 PDF

Matched-Input Estimates Differ in Sign Across Architectures: Auditing EEG Foundation Models on Motor Imagery

Kevin Zhou, Sparsh Roy Pretrained electroencephalography (EEG) foundation models are promoted as general-purpose encoders for brain-computer interfaces, yet benchmarks disagree about when their representations transfer. The authors audit LaBraM and CBraMod on motor imagery under a validation-locked protocol in which every choice, from preprocessing and freeze depth to temperature and method selection, is fixed using training-session data only. On four-class BCI Competition IV-2a, every supervised comparator outperforms every foundation-model configuration, including validation-selected fine-tuning. Retraining supervised architectures on the broadband inputs consumed by the foundation models yields matched-input differences of opposite sign, improving ATCNet by 0.078 accuracy while reducing EEG Conformer by 0.088, though none is significant after multiple-comparison correction at n = 9; the deficit does not reproduce on two-class BNCI2014-004, and validation-fitted temperature scaling restores foundation-model calibration to the supervised range.

arXiv:2609.23924 PDF

Open-Jev Judgments on CallScreenBench: Calibrated One-Pass Scam Screening with a Small Language Model

Simiao Ren, Kidus Zewde, Xingyu Shen, Yuchen Zhou, Dennis Ng, Ankit Raj et al. Screening a phone call for fraud requires a trustworthy probability after every caller turn within milliseconds. The authors test JevLite, an open implementation of typed decision readouts, by LoRA-tuning Qwen3-4B so that a temperature-scaled softmax over two answer-label logits directly yields the probability of a scam in a single forward pass with no generated text. On 41 held-out CallScreenBench scenarios a three-seed ensemble reaches an AUROC of 0.974 with calibration error of 0.052, non-inferior to a MiniMax-M3 judge at a pre-registered margin, with no false alarms on legitimate calls, decisions 1.14 turns earlier, and 64.5 ms per decision on one consumer GPU, 4.9 times faster than the same backbone generating its answer. The authors stress that the gain lies in readout and calibration rather than accuracy, since a fine-tuned ModernBERT encoder is not significantly worse, the recipe was selected with test-set exposure, and all callers are synthetic.

arXiv:2609.23959 PDF

UniK: Universal Knowledge Perception for Digital and Physical AI

Nirmit Desai, Kunal Sawarkar, Aditya Mahakali, Dongkon Lee, Kevin Park, Eric Song cross-listed Both digital AI systems that reason over enterprise knowledge and physical AI systems that learn control from video and sensor telemetry are bottlenecked by heterogeneous private corpora that existing infrastructure cannot access reliably. Universal Knowledge Perception (UniK) is proposed as a shared platform covering ingestion, enrichment, indexing, retrieval, and continuous evaluation across modalities, built on Polymath Retrieval, a multi-index fusion over automatically enriched indices with no task-specific fine-tuning. Across five digital AI domains spanning medical literature, open-domain question answering, chemistry, legal video proceedings, and government open data, UniK paired with an open-source 70-billion-parameter model matches or beats much larger proprietary models, reaching 76% retrieval-augmented generation accuracy on government data versus 47% for GPT-5 and 77.9% on medical question answering without fine-tuning. The authors argue the same infrastructure addresses the data curation and retrieval challenges of training world models for physical AI.

arXiv:2609.23971 PDF

Temporal Generalization and Explanation Stability of Control Flow Graph Neural Networks for Malware Detection

Md. Asif Sajeed, Md. Nazrul Islam Mondal, Md Ashraful Hossen Akash cross-listed Graph neural networks over control flow graphs perform well at malware detection, but they are typically scored on random splits of a corpus collected in a single period, which reveals nothing about later samples. Here every model trains on 459 control flow graphs from 2024-2025 and is scored once on 223 graphs from 2026, statically extracted from 1,989 Windows portable executables with 37 features per node. The ranking reverses across the time boundary: a flat-feature control that ignores topology is the best in-distribution model yet among the worst on the later corpus, so a conventional benchmark would have discarded message passing outright. The message-passing operator, not recalibration or ensembling, governs robustness, and the operator that generalizes best is also the hardest to explain.

arXiv:2609.24280 PDF

Evaluating Decision Models for Text Annotation in Computational Social Science

Hazem Ibrahim, Yasir Zaki Computational social science increasingly rests on language-model annotations, and a new class of decision models answers typed categorical questions with a label, a distribution over the label set, and a confidence score at a small fraction of frontier inference cost, but their accuracy and calibration on social science constructs were unknown. The authors mirror the Ziems et al. (2024) evaluation across 18 classification tasks (7,977 items), comparing the first commercial decision model and two open-weight counterparts against 19 frontier and open-weight LLMs under one zero-shot protocol. The decision model trails the per-task best LLM on 14 of 15 tasks by a median 11.6 macro-F1 points, at a median 44 times lower measured cost, and its confidence is better calibrated than the verbalized confidence of 16 of the 19 LLMs, though on an empathy task it reports high confidence while performing near chance; routing low-confidence items to an LLM matches or beats the LLM alone at a quarter to half the cost.

arXiv:2609.24574 PDF

Assessing Readability with LLMs: The Role of Reasoning and Few-Shot Prompting

Rapha\"el Thieffry, Matej Martinc Traditional readability formulas generalize poorly across genres and languages, and supervised models depend on scarce annotated corpora, so the authors benchmark open-source large language models (LLMs) as training-free readability assessors that predict the discrete levels required by educational frameworks. They evaluate in English and in Slovenian as a less-resourced language, testing Chain-of-Thought (CoT) prompting, reasoning-oriented models, and few-shot in-context learning against traditional unsupervised metrics and state-of-the-art supervised baselines. Explicit reasoning gives significant gains over direct answering, and a single labelled example per category substantially improves over zero-shot prompting, with additional examples offering diminishing returns.

arXiv:2609.24650 PDF

Decomposing Error and Style in Automated Clinical Coding

Han-Chin Shing, Jack Moriarty, Ryan Ware, Afton Marchbanks, Carlyn Canvasser, Stefanie Higgins et al. Automated clinical coding models are scored against a single gold set of diagnosis and procedure codes, so any deviation counts as error, yet two teams coding the same 110 ACI-Bench encounters agreed on only 73% of codes by Jaccard similarity, rising to just 77% after an independent clinical audit removed mistakes. The remaining gap is modeled as coding style, a coder- or site-specific policy over what to code and how much to document, estimated with a 10-dimension rubric and supplied as a conditioning variable so that coding becomes predicting a code given both the note and the style. Across five datasets, conditioning on a data-matched style raises International Classification of Diseases (ICD) F1 by up to 26 points, while an extreme mismatched style lowers it by up to 21, and four prompt-based coding methods that spanned 39 to 49 F1 converge to 52 to 56 once style is provided. Much of what single-gold evaluation attributes to model error is therefore recoverable, unmodeled style.

arXiv:2609.24877 PDF

180 more specialized papers

Large Language Models 123

RBS-Attention: Radius-Bounded Sparse Prefill for Long-Context Large Language Models

Chuxu Song, Jiuqi Wei, Zhencan Peng Long-context inference is bottlenecked by prefill, where dense attention must process the whole prompt before the first token; block-sparse selection helps, but a block's centroid can hide a single highly relevant key among many irrelevant ones, a failure the authors call mean dilution. RBS-Attention is a training-free sparse-prefill method that combines a centroid-based branch capturing average relevance with a rescue branch that uses each key block's maximum radius, thresholded per prompt, layer, and head, to recover blocks at risk of underestimation, while keeping standard block-sparse FlashAttention execution. On H100 GPUs at 128K context it reaches a 5.97x end-to-end time-to-first-token speedup on Qwen3-30B-A3B-Instruct-2507-FP8 (20.65x on standalone prefill attention), and on dense Qwen3-32B it scores 88.65 on RULER versus 89.52 for dense attention, with further quality checks on LongBench-v2, InfiniteBench, and Video-MME.

arXiv:2609.20971 PDF

Attention-Aware Routing: Coupling Routing and Attention in MoEs

Despoina Kosmopoulou, Anastasios Tsetsilas, Efthymios Georgiou, Giannis Karamanolakis, Swastik Roy, Alexandros Potamianos In Mixture-of-Experts language models the router picks experts from a token's hidden state alone, with little contextual information. Attention-Aware Routing (AAR) augments the router with temporal and spectral features summarizing a sliding window of attention weights, and the authors train only the routing parameters on a frozen OLMoE backbone so that routing is the sole variable. AAR improves GSM8K by 3.37 percentage points over a routing-only fine-tuning baseline, shortens incorrect answers without changing correct ones, and reveals that routing changes at one layer propagate through the residual stream to amplify attention sinks in the next; applying it at shallow layers hurts factual retrieval while deeper application preserves the math gains.

arXiv:2609.20974 PDF

How Much of a Real Workload Can LLM-Generated GPU Kernels Actually Reach?

Gaurav Agarwal, Ashish Garg, Isha Singhal cross-listed Language models can now write GPU kernels that beat PyTorch, but it is unclear how much of a real workload such kernels can actually speed up. Evaluating five model configurations on KernelBench level 1, a frontier model produces correct kernels for 91.1% of problems with verified speedups on 22 of 56 at a 1.235x median, while the best open-weights model reaches only 30.4% correct and solves no convolutions. Profiling seven workloads shows the addressable fraction of runtime ranges from 8.9% to 58.2%: on transformers 80-86% of time sits in cuBLAS GEMM and FlashAttention, capping realistic end-to-end gains near 1%, whereas recommenders expose 58.2% in a single embedding kernel, motivating the new DLRM-Bench where generated kernels project an 8.63% end-to-end gain. The authors also show that KernelBench's tolerance-based correctness check is satisfied by all-zero outputs on 4 of 60 problems, which two of their own kernels exploited, and propose scale-invariant replacements.

arXiv:2609.21058 PDF

Detecting Hallucination in LLMs: Tracing the Topological Signatures of Impaired Context Sharing

Amir Jalilifard, Anderson Rocha, Eric Wong, Marcos Medeiros Raimundo Detecting hallucinated responses from a model's internal attention patterns in a single pass remains difficult. The authors analyze Forman-Ricci curvature on attention graphs to locate information bottlenecks and build a detector that captures both semi-local and global information-flow features of the attention heads associated with hallucination. Across several LLMs and two hallucination-detection benchmarks the method consistently outperforms attention-based and multi-response baselines, and analysis links hallucinations to impaired context sharing: over-reliance on self-attention, diffuse retrieval from earlier tokens, or information over-squashing, especially in the final transformer layer.

arXiv:2609.21096 PDF

CogGym: Towards Large-Scale Comparative Evaluation of Human and Machine Cognition

Lance Ying, Jinzhou Wu, Yingshan Susan Wang, Shivam Aarya, Luca M. Schulze Buschoff, Harry Chen et al. Comparing how AI models and humans respond across the huge diversity of tasks that cognitive science studies is hard to do rigorously at scale. CogGym is a unified framework that uses a semi-automated, human-in-the-loop pipeline to standardize diverse experimental paradigms into a task-agnostic Experiment Markup Language (EML), enabling matched-trial comparison between models and human participants. The initial release standardizes 258 commonsense-reasoning experiments from 100 papers and evaluates 50 large language models against human responses, finding that larger and more recent models fit human judgments better, but that progress on these tasks is considerably slower than on formal benchmarks like math and coding. The best models reach R² of 0.59 on text, 0.58 on image, and 0.43 on video experiments, well below human split-half reliability of 0.92 to 0.95.

arXiv:2609.21259 PDF

CESBench: Benchmarking Large Language Models on Cryptographic Engineering Security for IoT Devices

Wenquan Zhou, An Wang, Jing Liang, Peien Feng, Jingqi Zhang, Yaoling Ding et al. cross-listed For Internet of Things (IoT) devices a secure cryptographic algorithm is not enough, since an attacker with physical access can target the implementation directly, and large language models (LLMs) are increasingly used to build and analyze such implementations without any benchmark covering cryptographic engineering. CESBench contains 380 expert-written items across side-channel, fault injection, implementation, countermeasures, evaluation, and integration, in four task types: 209 multiple-choice recall items, 67 judgment items requiring a security verdict and justification, 63 scenario items requiring an engineering diagnosis, and 41 code tasks graded by 572 test cases. Eleven open-weight and proprietary LLMs were scored automatically on multiple-choice and code, and by an LLM judge cross-checked against a second judge family and human re-scoring on judgment and scenario items, yielding composite scores from 54.4% to 83.6%. Models get 88.5% of security verdicts right but earn only 53.4% of the rubric marks for their justifications, so while multiple choice is near ceiling and most code tasks are solved, justifying a verdict remains the weakest competence.

arXiv:2609.21344 PDF

Talking Past the Machine: Morality, Politeness, and Alignment in Human-AI Dialogue

Marina Mitiaeva, Lu Xiao cross-listed Conversational AI produces fluent, socially appropriate replies, but it is unclear whether it participates in cooperative communication or merely simulates its surface forms. The study analyzes 15,881 human-ChatGPT and 10,784 human-human multi-turn dialogues with mixed-effects models to see which features of morality, politeness, and alignment predict turn-to-turn linguistic alignment. AI reproduces the surface features of cooperation without the underlying social architecture: moral output appears preconfigured rather than negotiated, warmth is generated without face sensitivity, and linguistic convergence declines persistently. Most strikingly, mechanisms that increase accommodation between humans run in reverse with AI: hedging and softening are associated with reduced alignment when produced by the model, and purity framing, which drives humans apart, coincides with users converging toward the AI, while giving users agency to shape the exchange is the most consistent predictor of alignment in both settings.

arXiv:2609.21401 PDF

The Communication Bottleneck: A Round-Trip Study of Tree-Structured Expression Serialization in Language Models

Xavier Suau, Alex Ferrando de las Morenas, Luca Zappella, Samy Bengio When language models reason in chain-of-thought or pass free-text intermediates to one another, they must serialize hierarchical structure into natural language, and it is unclear how much of it survives. The authors propose a round-trip protocol in which a generator turns a procedurally generated arithmetic expression tree into a word problem, a separate extractor recovers the expression from the text alone, and symbolic equivalence serves as an exact oracle, evaluated over all pairwise combinations of sixteen models to separate generation quality from extraction quality. The channel is lossy and asymmetric: swapping which model generates and which extracts shifts accuracy by up to 60.4 points, the best pair reaches 92.9% by using different models on each end, and at least 73.6% of failures originate at generation and track tree structure rather than model family. Roughly 3,600 fine-tuning examples lift every open-weight model above untrained Gemini-3.1-Pro, and the gains persist in a disjoint-domain regime with new operators and vocabulary.

arXiv:2609.21509 PDF

One Prompt Does Not Fit All: Self-Meta-Evolve for Personalized Information Extraction

Hongliang Li, Lu Wang, Yong Xu, Hanyang Chen, Zhitao Hou, Xiaoting Qin et al. Enterprise information extraction (IE) with large language models has to reorganize the same document differently for each user, but existing prompt-optimization methods tune a single prompt against a global objective. Self-Meta-Evolve keeps a dedicated prompt per user and refines it in two loops: an inner loop edits structured prompts from persona-conditioned feedback, and an outer loop evolves the meta-prompt itself by distilling successful editing patterns. The authors also release a benchmark of 292 simulated enterprise users generated from O*NET occupational taxonomies, on which the method reaches a 74.58% success rate, 13.56 points above the strongest prompt-optimization baseline, and a double-blind study with twenty professionals preferred its adapted prompts over static baselines in 71% of pairwise comparisons.

arXiv:2609.21626 PDF

Accelerating Dense LLMs via L0-regularized Mixture-of-Experts

Zhenyu Zhang, Jiudong Yang, Zhaowen Tao, Meng Chen Large language models are slow and costly at inference, existing acceleration methods usually degrade quality noticeably, and training Mixture-of-Experts (MoE) models from scratch demands extensive compute. L0-MoE converts a dense LLM into a lightweight MoE using L0 regularization, adding a cluster confusion matrix for domain-aware dataset curation and dynamic batching for efficient training. The authors report up to a 2.5x speedup over the dense model with nearly no performance loss, outperforming existing LLM acceleration baselines.

arXiv:2609.21672 PDF

Detecting Pretraining Data in Large Language Models from a Free-Energy Perspective

Chenye Ke, Zirui Liu, Qi Liu, Yan Zhuang, Jintao Zhang, Zhenya Huang et al. cross-listed Detecting whether a text was in an LLM's pretraining data is hard because high likelihood can reflect either memorization or ordinary generalization, so a likelihood-only detector draws a horizontal boundary in loss-entropy space and mistakes predictable non-members for members. The authors instead use an inclined boundary that scores prediction loss relative to predictive entropy, showing analytically that this correction preserves the expected membership signal while reducing its variance, and extending the analysis to a nonzero mean entropy gap. The adjusted score admits a Helmholtz free-energy interpretation, giving Energy Transfer Detection (ETD). ETD achieves the best average detection performance, improving AUROC by up to 3.5% and TPR at 5% FPR by up to 5.1% while staying robust across diverse settings.

arXiv:2609.21888 PDF

Recognition, Simulation, and Refusal: A Contamination-Aware Study of Classic Psychological Effects in LLM Agents

Joy Bose An LLM reproducing the response pattern of a human psychological effect is not the same as possessing that bias, and the two are easily conflated when the test paradigm sits in the training data. PsyAgentBench re-runs classic psychology experiments on LLM agents in a factorial design that crosses explicit naming of the paradigm versus framing it as a routine task, the canonical textbook version versus a structurally matched counterfactual with reduced lexical overlap, and a persona manipulation, releasing 41,904 trials across five paradigms and up to three open-weight model families. The apparently human-like effects arise through qualitatively different routes: Asch conformity jumped from 0% blind to 83.3% when the paradigm was named on gpt-oss-120B, anchoring was absent on grounded facts but near total on invented quantities, framing was amplified on novel content under labeling, sunk cost was robustly absent, and minimal-group allocation was dominated by refusals. A one-sentence agreeableness persona eliminated, dampened, or reversed effects depending on the paradigm, and the authors argue for reporting replication profiles rather than scalar bias-susceptibility scores.

arXiv:2609.22090 PDF

Token Signatures of Code: Comparing Coding Behaviors Across Large Language Models

Junpeng Wang, Yuzhong Chen, Menghai Pan, Uday Singh Saini, Yiwei Cai As more large language models (LLMs) clear baseline pass@k thresholds on coding tasks, performance metrics alone say little about how their coding behaviour differs. CLIC (Code Learning for Identification and Comparison) represents each generated program as a token-frequency vector and trains an interpretable decision tree to separate two models' code sets, adding a robustness metric that tracks whether the pair stays distinguishable as the most discriminative tokens are removed and a concentration metric that captures whether the difference rests on a few dominant tokens or many. An interactive visual analytics system lets users navigate pairwise comparisons across model pairs, tasks, and tokenization levels and drill into discriminative tokens in context, and case studies comparing 10 LLMs across 22 Kaggle machine-learning tasks surface insights for model selection and prompt engineering.

arXiv:2609.22097 PDF

TreeSpark: Calibrated, Load-Adaptive Draft Trees for Semi-Autoregressive Speculative Decoding

Huapeng Zhou, Huayu Wang, Xinyu Wang Speculative decoding speeds up inference by having a cheap drafter propose tokens that the target model verifies in parallel, and block drafters now emit a whole block per backbone pass, but existing draft trees rank candidates by per-position marginals that ignore which parent a candidate extends, so wider trees on semi-autoregressive drafters mostly add mis-ranked nodes, and fixed-size trees ignore how much speculation a given round or serving load can absorb. TreeSpark reads a parent-conditioned distribution from the drafter's existing Markov head at negligible cost, calibrates it into an edge-acceptance estimate, and lets path survival drive best-first expansion, per-round stopping, and a load-adaptive serving policy, while sampling siblings without replacement with matching residuals keeps decoding lossless at any temperature. Against a tuned chain on the same drafter it accepts 15-25% more draft tokens per round and decodes 8-14% faster in single-request wall-clock, and under rising load it shrinks the tree back to a chain.

arXiv:2609.22098 PDF

AdaMem: Adaptive Memory Token Allocation for Soft Compression in Retrieval-Augmented Generation

Artem Sakhno, Grigorii Davydenko, Omar Zoloev, Julia Belikova, Andrey Savchenko, Maksim Makarenko Retrieval-augmented generation (RAG) becomes expensive and noisy when many long passages are fed to the generator, and soft compression methods that encode passages as continuous memory embeddings currently give every passage the same number of embeddings regardless of relevance. AdaMem uses a shared query-conditioned compressor that produces both passage memories and relevance scores in one pass, then applies a deterministic rule that assigns more of a fixed memory-token budget to higher-scoring passages and can drop low-scoring ones entirely. Across six open-domain QA benchmarks it beats OSCAR, the matched uniform-allocation baseline, and other soft-compression methods at equal budgets, with an average relative gain of 3.4% at 16x compression that grows to 14.6% at 64x compression, peaking at 9.8 points on PopQA, while matching uncompressed answer quality at up to 4x lower inference latency than full-context inference.

arXiv:2609.22100 PDF

When Who You Are Can Change the Code You Get: A Study of Persona-Induced Bias in LLM Code Generation

Anubhav Gupta, Mayara Costa Figueiredo, Leticia Santos Machado, Tanner Wright, Ivan Beschastnikh, Cleidson R. B. de Souza et al. cross-listed Large language models (LLMs) are widely used as programming assistants, but it is unclear whether a user's demographic information changes the technical quality of the code they get. The study compares 18 personas spanning nationality, gender, and experience level against a neutral baseline on Gemini 2.5 Pro and GPT-OSS-120B, analysing more than 35,000 generated programs for demographic marker leakage plus functional correctness, maintainability, style, and security. Demographic markers appeared in up to 65% of responses and 70% of reasoning traces despite being irrelevant to the tasks; on LiveCodeBench persona prompts lowered Gemini's correctness by 1.54 percentage points on average, with one persona dropping 3.6% (odds ratio 0.51), while GPT-OSS accuracy rose 3.4-5.7% across all personas. Maintainability and style differences were statistically significant but negligible in effect size, and security vulnerabilities showed no systematic persona-specific pattern.

arXiv:2609.22102 PDF

PRQuant: Permutation Residual Quantization for Low-Overhead Inference

Peiran Wang, Anqi Wang, Jiaying Zhao, Huiwen Yang, Zhenyu Ming, Rongqian Wang et al. Low-bit quantization of linear layers is often undermined by a handful of outlier channels, and existing remedies such as smoothing, rotation, or residual correction either shift error onto the weights or add online overhead at inference time. PRQuant (Permutation Residual Quantization) is a training-free post-training method that, after AWQ-style scaling, identifies the input channels contributing most to weight quantization error, permutes them into contiguous tail blocks, and precomputes residual weight sub-tensors offline. Because the compensating channels are contiguous, the residual correction becomes a regular tail-augmented matrix multiply with no dynamic gather at runtime. Ablations show smoothing and residual compensation drive most of the accuracy gain while permutation mainly buys the hardware-friendly layout, and across five downstream benchmarks PRQuant beats default MXFP4 by 1.24 points on Qwen3-4B-Instruct-2507 and 0.55 on Qwen3-30B-A3B-Instruct-2507.

arXiv:2609.22106 PDF

A Shared Learning Rate Is Not a Neutral Control in Selective On-Policy Distillation

Chencheng Zhu Selective on-policy distillation trains a student only on the token positions a selector ranks highest, and published comparisons between selectors fix a single shared learning rate on the assumption that it is a neutral control. Sweeping an 8x learning-rate grid with LoRA on GSM8K (a Qwen2.5-1.5B student, 7B teacher), dense supervision is statistically flat while every selective variant moves with the rate, up to 17.7 percentage points for a teachability selector, so that the dense-versus-selective gap reads 10.1 points at one rate but 5.1 at the adjacent one. The authors name this selector-rate entanglement and, via a preregistered frozen-scoring ablation, show it stems from selection itself, with live scoring aggravating rather than causing it; under full fine-tuning at the rates the literature actually uses, the swings grow to nearly 50 points. On MATH-500 the rate dependence does not reproduce under LoRA, but the roughly 10-point cost of selective training does, and the prescribed fix is to report a full arm-by-rate matrix rather than a single shared-rate column.

arXiv:2609.22109 PDF

Type-Driven Tokenization for Brahmic Scripts

Sai Hemanth Kapila, Rakshika Bagavathy Standard LLM tokenizers produce malformed text on Brahmic scripts such as Devanagari, Telugu, Tamil, and Kannada, a family of abugidas in which consonants carry an inherent vowel that dependent marks modify. The root cause is that these tokenizers assume any two valid tokens can be concatenated, which holds for English (a semigroup) but not for Brahmic orthography, where only some concatenations yield valid strings (a partial semigroup). The authors formalize this distinction in Agda, model valid tokens as chains in a transition system, and derive a provably correct fixToken function that extends any candidate token to respect orthographic boundaries. That derivation becomes a practical patch for SentencePiece and a standalone Rust pre-tokenizer library, which eliminates the observed errors across Indic scripts.

arXiv:2609.22125 PDF

Correlation-Aware Structured Pruning for Large Language Models

Sicheng Xu, Hao Shi, Wei Zhang, Haoran Pang, Zhenyu Ming, Hao Wu et al. Structured pruning cuts large language model (LLM) inference cost while staying hardware friendly, but most methods score channels or heads independently, implicitly assuming pruning errors add up, an assumption broken by non-orthogonal weights and strongly correlated activations. The method casts pruning as a cardinality-constrained binary quadratic program that explicitly models cross-unit dependencies in the reconstruction error, then solves it approximately with a greedy interaction algorithm driven by dependency-aware marginal costs, since the exact problem is NP-hard. A gradient-based strategy allocates sparsity adaptively across layers of the whole model. Across mainstream LLMs, incorporating correlation information yields competitive accuracy-efficiency trade-offs against representative structured pruning baselines.

arXiv:2609.22131 PDF

Observational Equivalence of LLM and Human Annotation

Kentaro Nakamura, Jing Ling Tan, George Yean Whether large language models (LLMs) can stand in for human coders in text annotation is tested by replicating text-classification tasks from 14 peer-reviewed political science studies, with ten LLMs, three human experts, and 165 crowdsourced workers classifying the same texts under identical codebooks. Recent LLMs agree with expert coders at rates comparable to the agreement observed among the experts themselves, and the residual disagreement is driven by ambiguity in the texts and coding rules: where LLMs diverge from experts, experts also diverge from one another, and clarifying the rules reduces disagreement for both. The authors conclude that annotation quality alone gives little reason to prefer human coding, given the speed and cost advantages of LLMs, and that the real challenge is writing codebooks that minimise ambiguity. They propose using disagreement across LLMs to flag hard cases and refine codebooks, and derive ambiguity-aware bounds for downstream inference when no unique label can be defined for every text.

arXiv:2609.22133 PDF

Does the Truthfulness Signal Survive Code-Mixing? Probing Hidden States for Hallucination Detection in Hinglish

Tanveer Singh (Plaksha University) Hidden-state hallucination probes, linear classifiers trained on a large language model's activations to judge whether an answer is faithful to the input, report AUROC of 0.90 to 1.00 across several benchmarks and languages, but none have been tested on Hindi-English code-mixed text (Hinglish) despite its very large user base. The authors build a 5,674-item Hindi/English/Hinglish question-answering benchmark, generate and label 17,022 responses from Qwen2.5-7B, Mistral-7B, and Llama-3.1-8B, extract per-layer hidden states at two token positions, and train linear and MLP probes for in-distribution detection and cross-lingual transfer. The hallucination signal survives code-mixing, with transfer AUROC between 0.88 and 0.99 and gaps mostly under 0.05 relative to in-distribution probes, and Hindi-trained probes transfer to Hinglish more reliably than English-trained ones. As a separate finding, all three models hallucinate substantially more on Hindi and Hinglish than on English for matched facts; code and the synthetic Hinglish dataset are released.

arXiv:2609.22138 PDF

GRRR: The Geometry of Reshaping, Rotation, and Routing in Decoder LLM post-training

Jianing Qi, Hao Tang, Zhigang Zhu How supervised fine-tuning (SFT) and reinforcement learning (RL) actually change a large language model's weights is examined across 12 post-training chains by expressing each weight update in the singular value decomposition (SVD) frame of the corresponding pretrained matrix. This splits every update into three geometrically distinct parts: diagonal entries that reshape singular values, off-diagonal entries that rotate the coupling between pretrained input and output directions, and null-space entries that route signal outside the matrix's original nonzero core. On a math evaluation suite, removing the diagonal component usually preserves most of the post-training gains, suggesting that post-training works mainly by reconfiguring and extending pretrained pathways rather than by substantially altering singular values.

arXiv:2609.22146 PDF

Is Imagination Derived from Hallucination? A Cross-Taxonomy Evaluation of Imagination and Hallucination in Large Language Models

Zixuan Tang, Hongzong Li, Shuxin Zhuang, Dapeng Wu, Zi Liang Creativity benchmarks measure how far an output departs from common answers but never check whether the departure is licensed by the prompt, and hallucination is always scored in a separate pipeline, so the claim that imagination and hallucination share a generative mechanism has not been directly testable. Whiteboard adapts seven mechanism-grounded imagination subtypes from cognitive instruments for measuring human imagination, crosses them with ten support-boundary hallucination subtypes, and scores both axes on the same generation through a deterministic, auditable atom matrix with no LLM judge on the primary path. The item bank holds 1,660 prompts; on an 80-item anchor set the authors evaluate 79 LLMs and validate the instrument against 13,280 human judgments. Most subtype couplings between hallucination and imagination are negative, and every anchor item reproduces the negative coupling on its own, contradicting the intuition that imagination is derived from hallucination.

arXiv:2609.22152 PDF

The Limits of Speculation: Bounding Speculative Decoding in Mixture-of-Experts

Aidar Amankulov, Denis Mamatin Speculative decoding in Mixture-of-Experts (MoE) models suffers from unstable verification cost because which experts get loaded depends on the input. The authors cast speculation-budget selection as an offline Stochastic Shortest Path (SSP) problem over reference sequences and build a diagnostic Oracle that uses counterfactual simulation to account for MoE verification cost. Analyzing the Oracle's decisions on a Qwen3-Coder and EAGLE-3 pairing in the space of marginal deltas, rejected candidates form a strict linear boundary, meaning the global optimization is locally governed by a necessary condition balancing marginal expected cost against marginal expected progress, which the authors offer as a reference point for designing adaptive online heuristics.

arXiv:2609.22156 PDF

PAGE: Partition-Aware Gated KV-Cache Eviction

Pankaj Kumar, Subhankar Mishra KV-cache eviction methods choose which tokens to keep but never ask whether to evict at all, so a benchmark average can hide inputs where compression drives accuracy from 99% to 0%. The authors reframe eviction as a per-input admission decision, showing inputs split into a capacity-bound class where eviction is catastrophic at any budget and a dilution-prone class where it is safe or beneficial, and that a single label-free scalar from prefill attention, the early-to-late drop in pairwise top-k head agreement, predicts the class before decoding begins. PAGE thresholds this drop to apply any base evictor when the drop is large and keep the full cache otherwise, needing no training and only an unlabeled pilot of about 100 inputs to recalibrate for a new model family. Across four evictors, four models, and two benchmarks it cuts the harm rate on capacity-bound inputs from 0.75 to 0.026, turning a 99% to 0% collapse into a flat 89%, though the authors stress it is a safeguard rather than a compressor: realized compression is 1.8 to 3.4x against a nominal 16x budget and a trained evictor wins at matched memory.

arXiv:2609.22157 PDF

Didactic knowledge or Clinical Cases? How Data Types Shape Medical Large Language Models

Yuzheng Fan, Haochun Wang, Sendong Zhao, Xiao Han, Ming Ma, Bing Qin cross-listed Medical LLMs are typically trained on a mix of didactic sources such as textbooks and clinical sources such as patient records, but how each type shapes capabilities has been unclear. Token-matched experiments vary the didactic-to-clinical ratio and track performance, capability profiles, and error patterns on knowledge-intensive and clinic-oriented tasks. Transfer is asymmetric: clinical data improves clinic-oriented tasks while staying competitive on knowledge-intensive ones, whereas didactic data mainly helps knowledge-intensive tasks, and error analysis reveals a knowing-doing gap where gains in recall do not reliably translate into clinical reasoning. Modest amounts of clinical data capture most of the gains on tasks grounded in electronic health records, and the optimal ratio varies with the downstream task, so the authors argue data curation should be application-driven, favoring more clinical data for reasoning-heavy uses.

arXiv:2609.22161 PDF

MechaTerp-TRACE: A Novel Approach for Component Ablation Analysis in Language Models

Brandon Colelough, Davis Bartels, Madeline Bittner, Dina Demner-Fushman Interpretability work has explained factual recall in feed-forward layers and token relationships in attention separately, but offers no unified way to compare the causal contribution of different architectural components to a model's output. MechaTerp-TRACE (Teacher-forced Registry of Ablated Component Effects) ablates one registered component at a time, from whole transformer blocks down to individual neurons and output logits, and measures the change in the output distribution at a fixed answer token so all component types sit on a common scale. Applied to thirteen instruction-tuned dense decoder models from five families spanning one to thirty billion parameters, ablating 49,656 components across 48 medical and 42 general-knowledge prompts, the highest-effect components are the same few positionally fixed components in every model regardless of which entity is asked about, and once they are removed the remaining support is nearly evenly spread in eleven of thirteen models. Apparent localization of entity knowledge is therefore largely generic generation machinery, which undercuts methods such as targeted knowledge editing that assume entity knowledge sits in a findable place.

arXiv:2609.22163 PDF

CleanScore: Black-Box Benchmark Audits with Negative Controls and Sensitivity Bounds

Jeffery Opoku, David Banahene A model's benchmark score may reflect skill, prior exposure to the test questions, or both, and for most models the training data are unknown. CleanScore audits this from scored outputs alone: each benchmark question becomes a parent item with one public form and two independently written fresh forms preserving its numbers, facts, and answer, and the audit reports an interval for the public-form advantage rather than a verdict, using a private negative-control bank with an explicit transport radius to separate exposure from ordinary form mismatch. A registered audit of five open models on 200 GSM8K and 200 ARC-Challenge items finds no exposure-consistent advantage and bounds surface-form inflation below five points. Registered positive controls then show what such a null cannot rule out: leaking an item raises accuracy on paraphrases the model never saw almost as much as on the leaked wording, leaving 52% to 110% of the effect invisible to a paraphrase audit, and on ARC a planted 49-point advantage produced an observable gap of only -0.020 across four training seeds.

arXiv:2609.22183 PDF

Improving Parameter Utilization by Sharing Neural Experts Across Layers in Transformers

Dian Jiao, Jiaxin Duan, Shuai Zhao, Jiabing Leng, Yiran Zhang, Feng Huang Transformer language models often learn the same functional transformations redundantly at different depths. CS-MoE addresses this with cross-layer expert sharing: instead of the standard Mixture-of-Experts (MoE) design where each block ends with its own isolated experts, every layer keeps its own experts but also draws on a centralized, globally shared expert pool, which allows elastic control over how many parameters and floating-point operations each token activates. The architecture reaches lower perplexity than equal-scale dense Transformers while activating only 55% of the parameters, its performance scales monotonically with the number of activated experts, and expanding the shared pool at a fixed compute budget brings it close to MoE models that use more compute.

arXiv:2609.22199 PDF

The Effect of Quantization on Clinical Benchmarks: Accuracy and Safety Across Model Families

Leonard Twagirayezu, Prasenjit Mitra Quantization makes it possible to run large language models on resource-constrained clinical edge devices, but its effect on clinical accuracy and safety has been little studied. Five 7-8B models were evaluated at FP16, GPTQ-INT8, and GPTQ-INT4 precision on MedQA, MedMCQA, Med-HALT, a risk-stratified sample of HealthBench, and MedSafetyBench. INT8 is universally safe, with at most 1.9% degradation, while INT4 degradation is substantial and model-dependent: the clinically fine-tuned BioMistral-7B loses 19.7% on MedMCQA, Qwen2.5-7B drops 26.8% on the emergency-risk subgroup of HealthBench, and Qwen2.5-7B and Meditron-7B show large INT4 safety degradation even though refusal behavior is dominated by model family rather than precision. Two recovery methods, clinical calibration substitution and QLoRA fine-tuning, both recover MedMCQA while further degrading MedQA, so recovery strategies need task-specific validation.

arXiv:2609.22216 PDF

Measuring the Checker: Mutation Analysis for GPU-Kernel Benchmark Oracles

Mingzhe Du, Anh Tuan Luu, Dong Huang, See-Kiong Ng Benchmarks for LLM-generated GPU kernels decide correctness with a few random inputs and a loose floating-point tolerance, and their verdicts now feed leaderboards and reinforcement-learning rewards, yet there has been no way to measure whether hand-made patches to these checkers actually suffice. The authors apply mutation analysis as an adequacy metric: deterministic rules inject 10,303 compilable faults into verified CUDA implementations of 188 KernelBench problems, 7,384 with an independent kill witness, and any test protocol is scored by the fraction of faults it detects. The official check deterministically misses one in six witnessed faults (16.9%), with 78.6% of precision faults escaping versus 8.7% of arithmetic faults, and the metric explains the tolerance blind band, splits the gain of KernelBench-Verified into +4.0 points from hidden inputs and +4.5 from tighter tolerance, and exposes a published fuzzing recipe that rejects correct kernels 107 times. Optimizing test suites over the kill matrix reaches 98.0% detection with two inputs per problem, blindness grows with model scale across 48 whole architectures, and the dataset is released as KernelBench-M.

arXiv:2609.22220 PDF

Do LLMs Choose Like Humans? Using Cognitive Theory to Evaluate LLM Decision-Making

Johnathan Sun, Andrei Shleifer, Yonatan Belinkov Large language models (LLMs) show many human-like decision patterns, but it is unclear whether these reflect shared mechanisms or surface mimicry. The authors test whether LLM context sensitivity matches a cognitive economic theory that explains human choices through problem categorization and attention allocation, using a new 140,000-trial product choice benchmark across 12 open-source and commercial models. Context shifts model choices and problem categorization in human-like ways, but it does not reliably reweight attention between features such as price and quality, and neither scale nor chain-of-thought reasoning consistently reduces context sensitivity or makes behavior more human-like. The authors conclude that LLM decision mechanisms differ from human ones.

arXiv:2609.22225 PDF

List Counting Failures Are Not One Phenomenon

Iyad Ait Hou, Saad Mankarious, Aya Zirikly, Rebecca Hwa Open-weight chat models often miscount the items in a bracketed list, and prior explanations blame input bottlenecks such as subword fragmentation or attention dilution, which would predict that different models fail in similar ways. Comparing seven instruct models on identical prompts, the authors instead find distinct error modes: Qwen and Gemma 27B frequently flip odd lengths to a nearby even number, OLMo concentrates errors on a few mid-sized integers, and Llama tends to under-count, while heavier subword fragmentation does not make counting harder and Gemma 9B does not share the 27B model's odd-to-even pattern. A linear probe can usually still recover the true count from the residual stream even when the model answers incorrectly. Scaling a late MLP output helps Qwen modestly but is near null on Gemma 27B under the same protocol, and residual steering only trades odd-length gains for even-length losses, so magnitude fixes should not be transferred across models without a check.

arXiv:2609.22230 PDF

Not All Ranks Are Equal: Budget-Aware LoRA Merging Across Tasks

Avinash Amballa, Yashas Malur Saidutta, Wenbo Li, Lazar Valkov, Srinivas Chappidi Merging low-rank adapters (LoRAs) avoids swapping task-specific weights at inference time, but existing merging methods give every layer the same rank budget and often split it equally across tasks, which the authors identify as a major source of the gap between merged and per-task adapters. Since rank selection is NP-hard, they propose Net Utility, a data-free metric that decomposes each task LoRA by Singular Value Decomposition (SVD), scores each singular direction by its task utility against its interference with other tasks' directions, and globally selects the highest-scoring directions under a cap on the total number kept. Applied on top of five merging methods across three merging spaces, the allocation improves average accuracy by 2.1% on vision tasks and 2.2% on language tasks over the same methods with uniform budgets.

arXiv:2609.22237 PDF

A Tutorial on Prompt Engineering: From Messy Thoughts to AI Workflows

Erfan Loweimi, Hadi Daneshvar, Samira Loveymi, Samir Ouelha, Zhengjun Yue, Hajar Mozaffar et al. Prompt engineering is framed as a discipline for turning informal intent into structured work specifications, built from reusable design moves: define the work, include only the context the answer depends on, pick a role or a moderated panel of roles as an attention lens, and state affirmative quality targets while reserving prohibitions for hard boundaries. Occam's razor and Chekhov's gun are adapted as tests that every instruction must earn its place, and for consequential tasks the tutorial adds steelmanning, premortems, verification, and agentic operating loops with explicit boundaries and escalation. Written for a general readership rather than as a benchmarking study, it walks one worked example from a weak prompt to a strong specification and argues that the strongest prompt is rarely the longest but the one that makes desired behaviour, required sources and checks, and success criteria unmistakable.

arXiv:2609.22249 PDF

Teacher Should Think Ahead: Adaptive Continuations for Reliable On-Policy Distillation

Jingang Zhou, Yuyi Zhou, Haiyang Guo, Xukai Wang, Shuai Feng, Sirui Gao et al. On-policy distillation (OPD) gives a student dense token-level supervision from a teacher along the student's own trajectories, but the teacher's guidance can be unreliable when conditioned on incomplete or low-quality student prefixes. The authors identify Teacher Uncertainty Contraction (TUC), where teacher predictive uncertainty shrinks as it continues from a student prefix, and use a variance-bias decomposition of teacher-branch gradients to show that contraction reduces variance while teacher-student path divergence increases bias, favoring a finite continuation length. Adaptive-Continuations On-Policy Distillation (AC-OPD) augments informative states along student rollouts with teacher continuations and adaptively chooses their supervision horizon. AC-OPD consistently improves over standard OPD on mathematical reasoning and code generation across model scales, with controlled-continuation and matched-budget analyses supporting the adaptive design.

arXiv:2609.22254 PDF

Efficient Mixture-of-Experts with Speculative Decoding via Expert Coactivation

Kumari Nishu, Han-Byul Kim, Santosh Chilkunda, Maxwell Horton, Arnav Kundu, Mohammad Samragh et al. Speculative decoding (SD) speeds up dense models by verifying batches of draft tokens in parallel, but with Mixture-of-Experts (MoE) models each extra verification token activates more experts that must be transferred from DRAM to the Neural Processing Unit (NPU), and that memory transfer dominates runtime. The authors find that routers exhibiting high expert coactivation, where verification tokens share experts, run much faster and blunt the cost of larger verification batches. Sweeping router design choices on billion-parameter transformers, they show that combining a global load-balancing loss, shared experts, a consistency loss, and an autoregressive expert selection mechanism during training substantially raises coactivation. The resulting model improves throughput by 21% over MoE baselines while matching their accuracy.

arXiv:2609.22471 PDF

Replication Without Persistence in Hosted LLMs: Measurement Sensitivity in Action-Time Belief Evaluation

Bhushan Kashinath Joshi cross-listed Behavioral evaluations of hosted language models can shift because the service changed, the measurement instrument changed, or both. The study separates replication, whether a prior finding recurs on fresh data under its historical configuration, from measurement sensitivity, whether rebuilding the evaluation-and-inference configuration under the same model identifier changes the result, and from persistence across later identifiers under one instrument, using Regent Chess, a game with an exactly recorded hidden state so stated beliefs can be scored against ground truth at action time. The previously reported Gemini 3.1 Flash-Lite deficit relative to a matched-uniform comparator replicates on fresh games, yet in a same-day comparison under the same public identifier the endpoint is 0.0429 lower under a rebuilt configuration, and a frozen comparison reverses sign between Gemini 3.1 and Gemini 3.7. The authors argue that behavioral claims about hosted models should be indexed by tested identifier, serving period, measurement instrument, and inference configuration.

arXiv:2609.22478 PDF

CultureMINE: Datasets and Methods for Improving the Cultural Capabilities of NLP Systems

Tania Chakraborty, Eylon Caplan, Zhaoqing Wu, Kevin Cushing, Han Qin, Shreya Havaldar et al. Research on Cultural NLP, the effort to make language technology globally inclusive, has grown fast enough that tracking its methods and data resources is difficult. The authors analyze over 375 papers around three questions: which cultural capabilities (CCs) NLP systems are being built to have, how cultural data resources are created, and which methods are used to improve those capabilities. They report trends across all three axes, identify research gaps, and release the full annotated list of analyzed papers as an interactive web interface to which researchers can add their own work.

arXiv:2609.22494 PDF

Do Student LLMs Inherit OOD Robustness? Invariance-Weighted Distillation for Reliable Knowledge Transfer

Dileesha Kannangara, Sanghamitra Dutta Knowledge distillation (KD) compresses large teacher LLMs into small students, but students often degrade badly out of distribution (OOD). The authors attribute this to students absorbing spurious correlations in the distillation data and to standard KD weighting all samples equally regardless of whether the teacher itself relied on causal or spurious features. Invariance-Weighted Distillation (IWD) perturbs spurious cues while preserving core semantics across synthetic environments and upweights samples whose teacher predictions stay invariant, and the authors prove it lowers the student's spurious-to-causal gradient ratio relative to uniform KD. Across MNLI, SQuAD-v2, CoNLL-2003, and SST-2 with DeBERTa-v3 and Qwen-2.5 families, IWD achieves the highest accuracy on 15 of 16 OOD benchmarks, improving average OOD performance over standard KD by 4.34 points on natural language inference and 14.94 points on question answering while staying competitive in distribution.

arXiv:2609.22566 PDF

Preserving What Matters: Semantic Scaffolds Beyond Saturation in Summarization Evaluation

Nikhil Reddy Pottanigari, Ramin Fahimi, Noah Bolger, Sepideh Kharaghani, Ying Zhang Selecting a summarization model in production requires measuring summary quality, but ROUGE only captures surface overlap and LLM-as-judge scores collapse to near-identical values that fail to rank models, a saturation the authors document across three public datasets, two proprietary datasets, and multilingual settings. Semantic Scaffold extracts a hierarchical representation of facts, questions, and entity attributes from the source text, labels each as a main point or supporting detail, and reuses it as a fixed reference to score summaries through three diagnostic metrics: Fact Preservation Score (FPS), Question Preservation Score (QPS), and Entity Preservation Score (EPS), which reward retaining essential information and penalize detail overload. An analysis of four recurring failure modes of ROUGE and LLM-as-judge scoring shows that the scaffold-based metrics remain informative and interpretable where the conventional holistic scores collapse.

arXiv:2609.22603 PDF

Pretrained Persona Mixture Models and Tandem Models for Human Simulation

Minwoo Kang, T\'ea Wright, Seun Eisape, Ayush Raj, Suhong Moon, Joseph Suh et al. The dominant approach to simulating humans with LLMs, prompting instruction-tuned assistants to role-play a persona, produces stereotyped predictions that lack the natural diversity of real people. The authors show that a base pretrained model can be bound to a specific person from short individual samples of that person's dialogue, with demographics added afterward simply by querying the model, and call such well-calibrated models Persona Mixture Models (PMMs). Across corpora spanning open-domain text, human-AI chat, and task-oriented human dialogue, PMMs predict human responses more accurately than instruction-tuned models and preserve more lexical, semantic, and pragmatic diversity, though base models can drift out of domain and lose the person's internal state over long contexts. To address this, tandem models pair a pretrained model with an instruction-tuned supervisor, and tandem models achieve the best overall accuracy and diversity in the experiments.

arXiv:2609.22607 PDF

Beetle: A Bilingual Model Suite for Modelling Second-Language Processing

Suchir Salhan, Catherine Arnett, James Michaelov, Paula Buttery Bilingual language models offer a controlled way to study how training conditions shape second-language behavior, but earlier work varied exposure structure, scale, and architecture together, making it hard to isolate any one factor. Beetle is a pretraining framework in which tokenizer, target language, training budget, and exposure structure can each be manipulated independently; using it, the authors train and release 285 bilingual and 45 monolingual open-source models with rich checkpoints across exposure schedules, data scales, and first languages. Evaluated on human bilingual reading-time prediction and grammaticality judgement tasks, staged and temporally structured curricula consistently align better with learner reading times than balanced bilingual training, with the largest gains at smaller data scales and for typologically closer language pairs.

arXiv:2609.22633 PDF

Per-Query Gating of LLM Rerankers for Multi-Hop Retrieval

Andre Bacellar cross-listed Large language model (LLM) rerankers on top of a graph-augmented dense retrieval pipeline such as HippoRAG2 cost roughly $0.2 to $0.3 per 1,000 queries and about a second of tail latency, while improving final-hop coverage on seven of nine dataset and K combinations by up to 34.8 percentage points. The authors train a per-query gate that decides, from 27 score and lexical statistics of the two retrieval lists plus a compressed query embedding available before the LLM call, whether to skip the reranker and use an executable fallback, with every choice made inside the training fold and harmful skips reported alongside aggregate coverage. On 2WikiMultiHopQA, MuSiQue, and HotpotQA the gate skips 51% of reranker calls at an average held-out cost of 1.2 percentage points in last-hop coverage, though harmful skips outnumber beneficial ones and a budget-based threshold on calibrated harm probabilities realizes more harm than it promises. The harm probabilities are well calibrated but barely discriminative, and the authors document that an earlier claim of 73% lossless savings rested on an oracle fallback and a wrong target.

arXiv:2609.22880 PDF

Block-Sparse Attention with Semantic-Geometric Decoupled Routing

Xinwei Long, Weigao Sun, Weibo Gao, Pengkun Jiao, Biqing Qi, Feida Zhu et al. Exact dense attention scales quadratically with sequence length, and block-sparse attention that routes each query block to a few relevant key blocks is a hardware-friendly alternative, but training-free block routing is hard because existing routers pool post-RoPE token representations, entangling semantic content with rotary positional geometry and attenuating local positional cues through high-frequency phase cancellation. Semantic-Geometric Decoupled Routing moves semantic aggregation to the pre-RoPE space and reconstructs positional bias with an offline structural prior and relative block distances, giving a closed-form block routing score with no token-level search or post-hoc calibration. On long-context text and video tasks the method approaches full-attention accuracy across 4K to 128K contexts and reaches a 5.03x speedup over FlashAttention at 128K, with routing overhead under 3.4 milliseconds.

arXiv:2609.22884 PDF

Beyond Linear Context: Graph-Guided Evidence Navigation for Long-Novel Reasoning with a Local 9B Language Model

Wenji Fu cross-listed Long-context models read a novel token by token in narrative order, whereas a detective sorts events in time and keeps a map of who relates to whom, so the authors test whether a frozen knowledge graph can give a small local model that kind of access. One fixed qwen3.5:9b reader answers 234 multiple-choice questions over 30 detective novels under nine conditions: five graph routes, a recent-window baseline, whole-book compression, ordinary vector retrieval, and a question-only control. The strongest graph route reaches 53.85% accuracy against 46.15% for the recent window, 51.28% for compression, 51.71% for vector retrieval, and 40.17% for question-only, but none of the fifteen graph-versus-baseline contrasts survives Holm correction, so the result is presented as exploratory. Two structural findings hold up better: annotated evidence concentrates in the topological core of the graphs with 2.35x enrichment, and the two graph-building pipelines differ so much in clue-paragraph coverage, 16% versus 73%, that pooled accuracy would hide which bottleneck is being measured.

arXiv:2609.22939 PDF

Rethinking Pivot Programming Languages in Code Language Models

Andor Diera, Lukas Galke Poech, Matthias Tichy Multilingual code language models transfer skills across programming languages, but whether one language acts as a privileged pivot is contested: geometric analyses point to C-family languages and Go, while behavioral evidence favors Python. The authors revisit the question across three code models on multilingual competitive-programming data while controlling for representational anisotropy and length differences, two confounds in earlier cosine-based analyses, and examine pairwise language geometry, code-to-English alignment, and retrieval pivoted through candidate languages. The answer is relation-dependent: code-code geometry shows structured language regions but no universal center, code-English alignment favors high-level scripting languages, and pivoted retrieval prefers different intermediate spaces for code-to-code versus English-to-code transfer. Python's special role is better explained as English-facing affinity than as universal geometric centrality.

arXiv:2609.22988 PDF

WaveFront Decoding: Parallelized Self-Speculative Decoding for Looped Language Models

Hyeongju Ha, Jae-Joon Kim Looped language models reuse a weight-shared block to gain effective depth without extra parameters, but the T sequential block calls per token make decoding slow. Wavefront Decoding (WFD) is a training-free self-speculative scheme that uses intermediate recurrence outputs as drafts and exploits weight sharing to batch token states at different positions and recurrence depths into one block call, arranging them as a diagonal wavefront that drafts new positions at shallow depth while pushing earlier positions toward full-depth verification. Unlike separate draft-then-verify phases, drafting and verification share the same recurrent calls, with rejected drafts corrected by full-depth predictions. Across six Spec-Bench categories it delivers 2.42x speedup on Ouro-2.6B and 3.54x on Huginn-3.5B over autoregressive decoding, and cross-recurrence KV sharing lifts the Huginn figure to 4.81x.

arXiv:2609.23033 PDF

Attributable Post-Rationalization in RAG Citations: A Controlled Reproduction and an RLVR Comparison

Mehedi Khan, Md. Shariful Islam Bhuyan Retrieval-augmented generation (RAG) systems can produce a correct answer while citing a passage they never used, a behaviour called post-rationalization in which the model writes the answer first and attaches whatever citation looks close enough. The authors ask whether reinforcement learning from verifiable rewards (RLVR), which trains search agents to get answers right, also teaches honest citation, comparing an instruction-tuned base model against three RLVR agents derived from it on four question-answering datasets using free-tier Kaggle GPUs and adding a control missing from prior methodology. Post-rationalization is pervasive, with roughly one citation in seven unfaithful on Wikipedia-based questions, and RLVR does not reduce it: the agents post-rationalize at their base model's rate and one is slightly worse. Rewarding correct answers buys nothing for citation faithfulness, so it has to be trained and measured separately.

arXiv:2609.23053 PDF

Bridging Static and Agentic RAG for Taiwanese Historical Question Answering

Kai-Hsin Chen, Wei-Yu Chen, Xuanjun Chen, Jyh-Shing Roger Jang Agentic retrieval-augmented generation (RAG) lets a model adapt its retrieval based on evidence already gathered, but it is unclear whether that adaptivity beats a well-built static pipeline. A controlled comparison on Taiwanese historical question answering, with the same generator and hybrid retrieval backend for both, finds similar aggregate scores that hide disagreement: the two pipelines differ on 70.83% of questions, and an oracle picking the better response per question would raise the composite score by 0.2417 over the stronger pipeline alone. A post-hoc selector that compares the two responses and their cited evidence significantly outperforms either pipeline and recovers 60.34% of the oracle headroom. The authors argue that exploiting complementarity between retrieval strategies is more productive than searching for a universally superior one.

arXiv:2609.23056 PDF

Measured Joules, Learned Routes: Learning to Route for Energy-Efficient LLM Serving

Muhammad Abdur Rab Siddiqui, Daniela Rojas, Chen Yang, Wenqi Cui, Yuanyuan Shi, Yize Chen cross-listed Inference energy for large language models (LLMs) and agentic systems is growing, yet many queries do not need the largest available model. The authors train a language-model-based router that reads each query and picks an answer model from a fixed heterogeneous pool, using an offline tournament that profiles each candidate's correctness, latency, power, and GPU energy per query, followed by supervised fine-tuning and group relative policy optimization (GRPO). Learned routing selectively allocates expensive model capacity by query context and improves the accuracy-energy tradeoff in multi-LLM serving. Across seven benchmark tasks the authors observe a sharp accuracy-energy phase transition among routers, which they present as practical guidance for cutting energy while maintaining performance.

arXiv:2609.23085 PDF

Neural Spectral Capacity: Measuring and Designing Architectures from Network Specification Alone

Chenyu Zhu, Ruoyu Zhao, Zhichao Lu Parameter and FLOP counts are the standard scalars for allocating capacity in Transformer design and compression, but they cannot distinguish architectures that share a budget yet differ in depth-width, head, or feed-forward allocation. Neural Spectral Capacity (NSC) is a closed-form scalar derived from the singular-value spectrum of each weight matrix; under standard random initialization the Marchenko-Pastur law makes it computable from the architectural specification alone, with no model instantiation, data, or gradients. Its layer-wise additive structure enables NSC-DP, an exact dynamic-programming solver that returns the architecture globally maximizing NSC under resource constraints in seconds on a CPU. NSC outperforms parameter counts, FLOPs, and existing training-free proxies at ranking across seven Transformer and CNN families, discovers a Transformer-XL architecture on WikiText-103 that beats the human-designed baseline in 2 seconds, and prunes LLaMA-7B to the best 5.7B model across eight commonsense reasoning tasks without any calibration data, about 5900x faster than the strongest training-free proxy baseline.

arXiv:2609.23087 PDF

Perplexity Cost Understates What Activation Quantisation Breaks

Anish Sathyanarayanan Activation quantisation is usually judged by perplexity averaged over all predicted tokens, and the authors ask whether that aggregate reveals which computations a quantiser damages. Across 12 models from four families and 780 within-model comparisons, perplexity is a reliable aggregate signal, since the configuration it prefers also retains more induction and more retrieval in all but 2.1 and 4.0 percent of cases. But where perplexity has risen by only a factor of 1.2 to 1.5, induction still keeps 0.959 of its intact accuracy while retrieval has already fallen to 0.554, a gap the aggregate does not surface. Matched Gaussian-noise and sign-randomised controls show the damage is not explained by error magnitude alone, and quantising in a rotated basis restores induction from 0.001 to 0.980 at three average bits in a single-block intervention while retrieval recovers less completely, a pattern that holds on models up to 32B parameters and under AWQ once activations are pushed to 4 bits.

arXiv:2609.23125 PDF

Judging a Review by its Cover: A Reliability Analysis of LLM-based Peer Review Evaluation Metrics

Shakiba Amirshahi, Sajad Ebrahimi, Hai Son Le, Negar Arabzadeh, Ebrahim Bagheri LLM-as-a-judge metrics for peer-review quality risk rewarding fluency and polish rather than substantive evaluation, a concern that grows as reviewers use LLMs to improve clarity while keeping their judgments intact. The proposed statistical framework compares original human reviews with faithful LLM rewrites that preserve evaluative content but change wording, using 4,044 meaning-preserving rewrites derived from 674 ICLR and NeurIPS reviews to test 29 content-oriented metrics from four prior works for surface sensitivity and robustness. 23 of the 29 metrics assign significantly different scores to reviews whose evaluative content is unchanged, and only six meet the robustness criterion, with patterns consistent across two judge models. The authors conclude that robustness to meaning-preserving rewriting should be validated before such metrics are used to compare human-written, AI-assisted, and AI-generated reviews.

arXiv:2609.23264 PDF

ValueDiff: Value-Geometric KV Cache Eviction for Sink-Suppressed LLMs

Junyoung Park, Jungwook Choi, Mingu Lee KV cache eviction methods lean on persistent attention sinks, but modern LLMs with QK-normalization, gated attention, learned attention sinks, or logit softcapping have weaker sinks, and the authors observe that weaker sinks co-occur with value vectors that are more dispersed than key vectors. ValueDiff ranks tokens by the L2 deviation of their value vector from the cache mean, a score that also emerges as the minimal-disturbance eviction under a maximum-entropy assumption about future attention, and it is applied at every prefill block boundary and every decoding step under fixed budgets. On RULER at a tight 2k-token budget it retains 88 to 99 percent of dense performance across seven sink-suppressed models and is best on six, on LongBench at a 4k budget it averages 92 percent retention versus 83 percent for the strongest prior baseline, and on MATH-500 at a 25 percent cache budget it is the strongest non-dense method on every model tested, beating prior methods by up to about 20 points on gated-attention models.

arXiv:2609.23314 PDF

LLM-Based FORM Code Generation with Verification-Driven Fine-Tuning

Bakar Chargeishvili cross-listed FORM is a symbolic manipulation language used in particle physics to process the enormous algebraic expressions from multi-loop Feynman diagram calculations, and no AI tooling existed to help physicists write it. The authors show that contemporary large language models (LLMs), including frontier models with hundreds of billions of parameters, score a zero-percent single-attempt execution pass rate on their instruction-following and tutorial-style FORM tasks without documentation, then build a verification-driven pipeline that uses the FORM binary as an execution oracle to produce and validate 4,633 training examples spanning deterministic computations, open-ended programs, tutorial code, and knowledge question-answer pairs. Fine-tuning Qwen3-8B with quantized low-rank adaptation (QLoRA) yields a specialist that decisively outperforms frontier models of up to 756B parameters in execution rate and strict FORM-verified output matching on the larger of its four benchmarks totaling 840 tasks, is statistically indistinguishable on the smaller harder ones, and keeps general reasoning and coding within 2.6 percentage points of the base.

arXiv:2609.23367 PDF

Machine-Interpretable Information: Compiling Documents into Searchable and Readable Protocol States

Yifan Wang, Dejing Dou Long-context language models interface with external knowledge as raw text, so retrieval-augmented systems suffer an index-payload split: dense vectors route queries but the model must re-ingest long payloads at quadratic attention cost, and existing compression methods produce private states tied to one architecture. Machine-Interpretable Information (MII) is an agent-to-agent (A2A) document-to-state protocol in which a dual-timescale state-space Writer compiles a document into a canonical 56-token state and a lightweight Translator maps it into any frozen Reader's embedding space, cutting query-time cost to O(K) and yielding a transferable .mii artifact that serves retrieval, reasoning, and reconstruction. The authors demonstrate interoperability across Llama, Qwen, and Mistral readers even though the Writer uses a legacy GPT-2 vocabulary, show that entity representations can be causally traced and zero-shot transplanted between unrelated document states, and report that Residual-MII, which pairs the compiled global state with sparse local evidence, exceeds full-context exact match on 7,405 HotpotQA queries at roughly 7 percent of the attention FLOPs.

arXiv:2609.23371 PDF

Global Ranks Survive, Selected Heads Shift: BOS-Sink Topology under 4-bit Weight-Only Quantization

Kuanlin Chen, Chen-Wei Kuo, Cheng-En Ou Sink-aware deployment identifies important first-token attention heads before a model is quantized and reuses that map at the edge, and the work asks when this shortcut is safe under 4-bit NF4 weight-only post-training quantization (PTQ). Sink Topology Consistency (STC) metrics separate global rank preservation, top-k set overlap, and layerwise sink-mass shift across Qwen2.5-0.5B, Qwen2.5-1.5B, and Llama-3.2-1B. Global bf16-to-4-bit head rankings stay highly correlated (Spearman at least 0.980), yet top-k Jaccard overlap is only 0.619 to 0.793, terminal Qwen layers shift by 6.2 to 7.9 times their model means, and a C4-to-LongBench domain shift degrades cross-domain overlap for both Qwen models but not for Llama. Matched-domain recalibration approaches a stability plateau with as few as 8 to 32 samples and runs in seconds on a Jetson Orin NX, so the practical guidance is that global rankings often transfer but discrete head sets, layer-local policies, and cross-domain calibration should be revalidated after quantization.

arXiv:2609.23585 PDF

Distill What You Trust: Reliability-Aware Multi-Teacher On-Policy Distillation

Jie Sun, Mao Zheng, Mingyang Song, Zeyuan Liu, Gengsheng Li, Houcheng Jiang et al. Multi-teacher on-policy distillation lets a student learn from complementary specialist models on its own generated trajectories, but domain-routed approaches pick one teacher per example, need domain labels that mixed corpora often lack, and cannot switch teachers when the required expertise changes mid-response. TrustMOPD replaces example-level teacher selection with label-free, token-level supervision allocation: at each student prefix it uses each specialist's reinforcement-learning-induced displacement from a shared pre-RL reference as a proxy for local reliability, calibrates these scores across teachers, and builds a weighted distillation target. Across mathematics, code, and instruction following it beats the strongest label-free baseline, raising the recovery ratio from 54.4% to 91.5% on SingleCap and from 54.5% to 98.0% on MultiCap, while approaching label-based MOPD on SingleCap. Randomizing the token-level weights independently of the prefix performs no better than uniform weighting, showing that conditioning supervision on the evolving generation context matters.

arXiv:2609.23697 PDF

STEVE: Stabilizing Textual Gradient-Based Prompt Optimization via Error-Driven Refinement and Regularized Verification

Yifan Xu, Yixuan Li, Xinzhuo Li, Yixin Gu, Yifan Shen, Lijun Yu et al. Textual-gradient prompt optimization, which refines prompts using natural-language feedback, can be unstable because gradients from already-correct examples add noise and repeated focus on hard cases over-specializes the prompt. STEVE addresses both with Error-Driven Refinement, which derives feedback only from failed examples, and Regularized Verification, which treats each update as provisional and keeps it only if gains on hard cases do not cause unacceptable regression on a preservation set. Across ten reasoning benchmarks and three evaluator/optimizer models, it reduces degradation and yields more robust prompts than established baselines, with further runs on gpt-5.4-mini and gpt-5.4 over symbolic reasoning, GSM8K-Platinum, and DS-1000 showing the gains hold with newer models and larger test sets.

arXiv:2609.23716 PDF

Constrained Decoding Eliminates Structural Failures in Small LLMs but Reveals a Scale-Dependent Semantic Gap

Akash Chavan Small open-source models in the 0.6B to 4B range are widely used for JSON, function calling, and data extraction, but how constrained decoding (CD) interacts with model scale in that regime is not well understood. The study benchmarks five models from three families on 14 structured-output tasks under native decoding, Outlines, and XGrammar, scoring structural correctness (schema validity) separately from semantic correctness (content accuracy). Constrained decoding lifts schema validity from 78.6-92.9% to 100% for every model, yet content accuracy shows a scale-dependent gap it cannot close: type-coercion errors are fully rescued, while instruction-semantic failures such as multi-step function calling persist regardless of decoding.

arXiv:2609.23742 PDF

Q-TIE: A Lightweight and Generalizable Re-ranking Framework for Temporal Information Retrieval

Soyeon Kim, Hyunjin Kim, JinYeong Bak, Steven Euijong Whang Temporal Information Retrieval (TIR) seeks documents that are both semantically and temporally relevant to a query, since temporally mismatched evidence can mislead Retrieval-Augmented Generation (RAG) systems. The authors observe that temporal retrievers learn flexible query representations but often ignore explicit temporal constraints, while temporal re-rankers enforce constraints but rely on predefined rules, and propose Q-TIE, a re-ranking framework built on a learned Temporal Intent Extraction (TIE) model that maps each query's temporal constraint to a unified start-end interval used as a separate signal. Q-TIE consistently outperforms existing TIR methods with stronger generalisation across temporal query types and serves as a lightweight add-on for temporally aware RAG pipelines, with code released.

arXiv:2609.23880 PDF

this-that-model-1.0: A typed decision model that decides in 30 ms, for a millionth of a cent

Zehua Cheng, Wei Dai, Jiahao Sun Software increasingly delegates discrete branching decisions to models, such as which queue a ticket enters or whether a command is safe to run, yet a frontier API round trip returns prose that must be parsed, takes hundreds of milliseconds, and bills per token. this-that-model-1.0 is a 2B-parameter typed decision model that reads its answer directly from the hidden state at a designated position, restricted to the caller's declared option set, so no text is generated, outputs cannot be malformed, and every question in a request is answered in one forward pass. It decides in 30.9 ms on a laptop GPU and scores 0.941 with a Brier score of 0.042 on a third party's 68 recorded decision questions, against 0.765 and 0.133 for the hosted service Jev, while a pass of the 42-family internal suite costs 32 seconds and 0.000217 USD of electricity versus 155.2 minutes and 10.636 USD for the most accurate hosted model. It loses on multi-step arithmetic, scoring 0.560 against hosted models' 0.98 to 1.00, and a targeted second training round improved only the five task families it was written for; the weights are open-sourced.

arXiv:2609.23886 PDF

Circuit-Diff: Factual Edit-based Intervention Method for Localizing Knowledge in Attribution Graphs

Edward G. Friedman, Xiangchen Song Cross-Layer Transcoders (CLTs) approximate a model's circuits by producing an attribution graph, but the graph's nodes are unlabeled features that must be pruned and interpreted by hand. Circuit-Diff applies a low-rank factual edit to the model itself and flags the features whose role in the attribution graph changes under that edit as related to the edited knowledge. On the edits examined, the flagged nodes include not just detectors of the object token but features for the history, geography, and associations surrounding the old and new objects, as read from the CLT's released feature dashboards. The authors formalise the method, measure how reliable a frozen CLT remains after an edit, causally test selected nodes by patching them on up to 24 CounterFact edits, and release an implementation built on circuit-tracer with multi-prompt aggregation and rule-based supernode labeling tools.

arXiv:2609.23892 PDF

GDN Tree-Scan: Served Tree Verification for Recurrent-Hybrid Language Models

Zhiyuan Ma Tree speculative decoding verifies several candidate continuations in one target forward pass, which for attention-only transformers needs little more than an ancestry mask, but recurrent-hybrid models also require each candidate row to carry the recurrent state that sequential decoding would have produced along its root-to-node path. GDN Tree-Scan is a served verifier for Gated-DeltaNet hybrid models integrated into vLLM, combining FlashAttention-2 tree-bias attention, branch-local scan and replay of the recurrent state, device-side multidraft commitment, and publication of state only along the accepted chain. On the public Qwen3.6-27B-FP8 checkpoint in a batch-one coding decode setting at temperature 0.6, a six-node tree raises committed tokens per event by 17.2% at near-native verify time and reaches 23.88 token-weighted decode tokens per second versus 18.80 for native five-step multi-token prediction, a 27.0% throughput gain, though per-request latency improves only 4.0% and end-to-end task time stays prefill-heavy. Equivalence evidence is limited to recurrent-oracle probability-rescore closure within the observed native flip floor rather than a full distributional proof.

arXiv:2609.23900 PDF

Time-Incremental Continued Pretraining of LLMs: Knowledge Updates Without Catastrophic Forgetting

F{\i}rat \"Oncel, Salman Hussain Ali, Mirco Ravanelli, Cem Subakan, \c{C}a\u{g}atay Y{\i}ld{\i}z Large language models (LLMs) go stale after pretraining, and continued pretraining (CPT) is usually evaluated through a continual learning lens that assumes disjoint data streams, which fits poorly with time-incremental web crawls whose successive snapshots share substantial URL overlap. The authors study CPT on FineWeb-Edu dumps drawn strictly from after each model's knowledge cutoff across six open-weight models from the OLMo2, Llama-3.1/3.2, and Gemma-3-1B families at 1B to 8B parameters. Knowledge is acquired without catastrophic forgetting: five of six models also improve on pre-cutoff factual recall, and the macro-average over a thirteen-task suite stays within 0.01 of the base for every model. A curated 6B-token slice matches a broader 40B-token one, the learning-rate optima for knowledge acquisition and general capability differ by roughly an order of magnitude, LoRA at sufficient rank matches full CPT, and gains survive supervised fine-tuning (SFT) while the effect of direct preference optimization (DPO) is family-dependent.

arXiv:2609.23916 PDF

FlashBoB: I/O-Efficient Exact Backward-over-Backward for Softmax Attention

Anthony Givans, Michael Crawshaw, Mingrui Liu FlashAttention makes the forward and first backward passes of softmax attention I/O-efficient, but it does not cover backward-over-backward (BoB), the exact second differentiation through the backward pass needed for second-order optimization, test-time training, gradient-based memory, and meta-learning, and existing BoB code either materializes large intermediates or runs out of GPU memory at long sequences. FlashBoB is an exact algorithm that keeps the double backward within on-chip tiles and never forms an N-by-N tensor, exploiting a hierarchical affine structure in which two row-wise scalars determine all outputs, which yields a two-pass schedule with bounded SRAM use and HBM traffic of Θ(N²d²/M) that matches the inherited large-cache lower bound for exact forward attention. It scales exact attention BoB to 262K tokens on a single A100 80GB, where prior exact PyTorch baselines fail by 16K, and runs up to 6.3x faster than FlashBack.

arXiv:2609.24089 PDF

You Can Tell Who's Asking: What the Web's Questions Are Made Of, and Where They Come From

Calvin Zhou, Vincent McCloskey, Krishna Srinivasan Questions scraped from web pages are widely used as a proxy for what people want to know, in question-answering training data, retrieval benchmarks, and content strategy alike. To test that assumption, the authors extract 13.4 billion question occurrences from 110 FineWeb snapshots spanning 2013 to 2025 and analyze their provenance. A logistic model distinguishes genuine user questions from templated or manufactured ones at an AUC of 0.725, mostly from length and surrounding context rather than question type, though only 0.554 against commerce FAQ writing; over 70% of the thousand most frequent questions are boilerplate, so occurrence counts measure how often a string was published rather than how often it was asked. Over twelve years the share of genuine question occurrences fell by 79%, or 42 to 56% after controlling for crawl composition, indicating the crawlable web's questions have shifted from being asked by humans to being manufactured for machines to read.

arXiv:2609.24106 PDF

Re:CAP - Auditing Retrieval Coverage in Production RAG Pipelines

Aviral Joshi, Hanoz Bhathena, Max Nelson, Saket Sharma Retrieval quality in production retrieval-augmented generation (RAG) systems is hard to monitor because multi-million-passage corpora re-index continuously and exhaustive relevance labels do not exist. Re:CAP (REtrieval Coverage Audit by iterative Probing) audits coverage without references by taking a deployed pipeline's initial answer and retrieved context, identifying the topics already covered, generating probing questions for plausibly missing topics, retrieving candidates, and using an LLM-as-judge to keep only documents that add previously unretrieved information. On four public benchmarks it recovers 9 to 29% of gold documents that flat BM25 top-500 cannot reach, rising to 48% on TREC-COVID, and on MuSiQue it beats hybrid top-500 recall by 12.9 points at under half the document budget. Human annotators judged that 78.9% of the documents it adds over a BM25, dense, and hybrid ensemble contribute new information, 73.9% on live production traffic, and end-to-end recall is reproducible within ±1% across three independent runs.

arXiv:2609.24122 PDF

OSCAR: Order-aware Scoring and Calibration for AI Rankings

You Liu, Yue Liu, Quanchao Lu, Nick Shipilov cross-listed Judge-specific sensitivity helps aggregate pairwise evaluations by large language model (LLM) judges, but its meaning depends on which systematic presentation effects the ranking model accounts for. OSCAR extends sensitivity-based ranking with judge-specific position, length, and family terms, characterizes how omitting such terms displaces estimates and can break identification, and propagates prompt-cluster uncertainty into adjusted comparisons. In released judgments from 18 evaluators the A-minus-B score difference spans -63.11 to 98.31 percentage points, with an overall difference of 24.22 points after matching questions, responses, candidates, and judge, and a controlled calculation shows that omitting a position intercept of four collapses the population-optimal slope from one to 0.0771. Position gives the largest stand-alone predictive gain across four datasets, and at N = 10,000 the full model cuts mean neutral-target RMSE from 0.1158 under a sensitivity-only model to 0.0237, while simulations show both mean and covariance must be adjusted to reach nominal 94 to 95% coverage.

arXiv:2609.24128 PDF

CLOOPD: Closing the Learner Loop in On-Policy Distillation

Keye Zheng, Hanyu Li, Zhan Cheng, Yuan Gao On-policy distillation (OPD) pays twice per batch, once for the student to generate trajectories and once for a stronger teacher to score them, yet most methods spend that teacher signal on a single actor update. CLOOPD separates acquiring the teacher signal from realizing it on the student side: it picks an adaptive alpha waypoint inside a KL envelope, freezes the scored batch and its advantages, re-forwards the student after each actor pass to measure how much of the signal was realized, and allocates additional actor work under a separate token budget, with deterministic two- and three-pass variants and a token-priced CLOOPD-TPMR. Across six 300-step runs on an 8-H20 node, every variant beats the one-pass TOP-D anchor at comparable teacher-token scale, with macro accuracy rising from 15.41 to 17.78 for two passes and 19.36 for three. At step 100, the three-pass variant nearly matches TOP-D at step 300 while using 67.2% fewer teacher-scored tokens and 28.0% fewer GPU-hours, and earlier ablations show the adaptive alpha eliminates observed trust-envelope violations.

arXiv:2609.24141 PDF

Acceptance-Aware Draft Model Training for Speculative Decoding

Tianhua Xia, Mugilan Ganesan, Yifei Feng, Haiyu Wang, Maximilian Egger, Sai Qian Zhang Speculative decoding speeds up large language model (LLM) inference by having a small draft model propose tokens that the target model verifies in one forward pass, but draft models are usually trained with cross-entropy or Kullback-Leibler (KL) divergence, which match distributions rather than directly maximizing the acceptance length that determines speedup. The authors derive an expected accepted length (EAL) loss for greedy verification and a window total variation (WTV) loss for sampling-based decoding that accounts for sequential acceptance dependencies, optionally followed by a GRPO reinforcement learning stage that rewards simulated acceptance length. Both losses consistently increase acceptance length over KL-based training across target and draft model pairs, tasks, and decoding settings, with WTV strongest under sampling and EAL best matched to greedy verification.

arXiv:2609.24150 PDF

When Residualization Helps an Audit: Format Effects, Slice Gains, and Their Limits

Daein Weon, Dongho Kang Scores from reward models, rerankers, and LLM judges can track surface form rather than the quality they claim to measure, and subtracting the predictable surface component (residualization) is an increasingly common fix that cannot tell whether the removed component carried real signal. Using unit-test-labeled MBPP code with comment-only edits, the authors find that a public preference reward model picks a terse correct solution over a commented buggy one only 50.7% of the time, and residualization attenuates the format effect by about 0.12 while moving correct-versus-buggy margins by less than 0.01. In observational natural language inference and question answering settings with a frozen held-out replication, residualization improves agreement with construct labels only on a pre-declared slice where a surface-only predictor errs, while full-population agreement falls in every such setting. A controlled model shows that configurations equally damaging to construct alignment pass every pre-adjustment check, so the authors propose a reporting protocol that treats adjusted scores as audit-time diagnostics reported alongside their construct-alignment cost rather than as replacements for raw scores.

arXiv:2609.24194 PDF

H-Spec: Parallel Speculative Decoding Without a Drafter-Side KV Cache

Weifan Jiang, Krishna Teja Chitty-Venkata, Megan Flynn, Reed Meyerson, Zhenting Qi, Tianyu Wu et al. Speculative decoding uses a lightweight drafter to propose tokens for the target model to verify, and block diffusion drafters predict several tokens in parallel, but existing block drafters project target hidden states at every position into a separate drafter-side key-value (KV) cache whose memory and write cost scale with concurrency. The authors propose hybrid target-context injection, which reuses target KVs in place and adds projected target hidden states only at the last input position, and build H-Spec, a hybrid Mamba-attention parallel drafter where Mamba modules take the last-token hidden state and attention modules read the target KVs directly, with Mamba's parallel scan preserving block-parallel drafting. Across three target models and diverse tasks, H-Spec improves mean accepted length by 5.0 to 13.3% and single-request inter-token latency speedup by 5.3 to 12.6% over the best baseline, and under concurrent serving it sustains higher throughput with lower KV cache utilization.

arXiv:2609.24197 PDF

Memory vs. Context? Influential Factors of Factual Recall in Language Models

Guilhem Fouilh\'e, Nicholas Asher, Philippe Muller Yu et al. (2023) characterized how language models choose between memorized facts and contradictory in-context statements, and this work reproduces and stress-tests those findings. The authors replicate the world-capitals experiments on 31 models across the Pythia, GPT-2, Qwen3, and Ministral families, including base and post-trained variants, and extend to five more relation types from ParaConflict. Most original findings hold, with larger models and higher-frequency entities favoring memorized answers, but several do not generalize: entity-frequency effects vanish on Qwen3-14B and 32B, post-training shifts the trade-off inconsistently across families, question phrasing alone can change reliance on memorized knowledge by up to 80 percentage points, and semantically unrelated prose can mimic coherent supporting context.

arXiv:2609.24238 PDF

KV-COBRA: KV Cache Compression via Co-Optimized Bit-Rank Allocation

Sihyeon Ha, Jaeho Lee, Yo-Seb Jeon The limit on key-value cache compression at extreme bit-rates is argued to be how the budget is split across attention heads rather than the compression scheme, since each head has its own best mix of rank truncation and quantization. KV-COBRA frames this as resource allocation, balancing truncation loss against quantization loss inside each head and then redistributing budget across heads to minimize total distortion, using only ordinary low-rank projection and scalar quantization. A fused Hadamard rotation equalizes per-channel variance and reordering the singular value decomposition basis by attention Kullback-Leibler importance makes the solver query-aware, and the same allocator handles joint key-and-value compression. Across perplexity, zero-shot and long-context benchmarks from 0.5 to 4 bits per dimension, it shows the smallest accuracy degradation of the compared methods at low bit-rates with no per-token overhead.

arXiv:2609.24298 PDF

The Undetected Damage of Quantization on Retrieval and How to Fix It

Luca Zhou, Alessandro Zirilli, Daniele Solombrino, Roberto Dess\`i, Emanuele Rodol\`a A quantized model can hold its classification accuracy and still break retrieval, with 14 to 46% of top-1 retrieval results changing, damage that aggregate ranking metrics only partly expose. The authors tie this to the gap between the two highest scores, proving the top-1 result survives quantization only when that gap exceeds twice the largest rounding error, and note that classification training actively pushes the correct logit away from competitors while nothing separates the top two documents in retrieval. Because the gap needs no labels to measure, it predicts before deployment which models will break and tells you per input whether a quantized answer still matches full precision. The remedies differ by task: spending extra bit-width on the layers whose quantization moves the gap most recovers up to three-quarters of an extra bit's benefit for half the cost in retrieval, while routing the few low-gap inputs to full precision recovers most lost accuracy in classification.

arXiv:2609.24322 PDF

ARM: Attention with Routed-Memory for Learnable Sparse Control

Qiuhao Zeng, Jerry Huang, Peng Lu, Ruiyi Fang, Gezheng Xu, Zihao Jing et al. Long-context inference in large language models is bounded by the key-value (KV) cache, and eviction or pruning schemes that shrink it often discard information that later turns out to matter. Attention with Routed Memory (ARM) replaces the growing cache with a fixed-size, fully differentiable memory organized as a hierarchical router, in which Gumbel-Softmax selects memory slots and sigmoid gates softly blend new and stored content rather than hard-evicting entries, while a separately trained policy chooses how much memory to access per input at inference time. On standard commonsense and long-context reasoning benchmarks the authors report better accuracy and efficiency than fixed KV-caching approaches while remaining scalable in memory footprint and generation latency.

arXiv:2609.24417 PDF

1% of Tokens Can Be Enough: On Gradient Estimation in On-Policy Distillation

Huanxin Sheng, Zhiling Ye, Haonan Wang, Jian Wang, Jinjie Gu, Jian Kang Sparse on-policy distillation (OPD) sends teacher supervision to only a small subset of tokens in student-generated trajectories, but even useful guidance can produce a noisy update when its gradient is estimated from a single sampled next token. The authors analyze this estimation problem at a fixed prefix through information geometry and define an information-efficiency ratio (IER) from a signal-to-noise decomposition that characterizes relative gradient-estimation error under an optimal scalar baseline, with a candidate-set approximation that makes it usable for token selection alongside existing usefulness scores while keeping the sampled reverse-KL objective. On mathematical and medical reasoning tasks adding IER improves existing selectors, and sparse configurations using only 0.1% to 1% of tokens match or exceed full OPD without token selection.

arXiv:2609.24432 PDF

Comparing Latent Concept Formation in State Space Models and Transformers via Sparse Autoencoders

Rithin Nagaraj, Rupa Laalasa Oruganti, Prerna Subhashchandra Kunder, Ashwini M Joshi Selective State Space Models (SSMs) such as Mamba compress all past context into a fixed-size recurrent state, raising the question of whether they learn different latent representations from Transformers. Using Sparse Autoencoders (SAEs), the authors run a feature-level correspondence analysis between Mamba-130m and Pythia-70m over a 10-million-token corpus. They find no evidence of systematic representational divergence, with 99.98% of Mamba features clustering toward the upper alignment boundary of the Jaccard distribution, offering preliminary support for the Universality Hypothesis. The remaining 0.02% of diverging features suggest the recurrent bottleneck mainly limits parsing of rigid syntax and formatting edge cases, which Mamba compresses into polysemantic neurons where Pythia's attention yields monosemantic ones.

arXiv:2609.24440 PDF

RAILS: Retrieval-Augmented Incremental LLM Clustering at Scale

Armin Oliya, Aleksandra Sawczuk, Rados{\l}aw Bia{\l}obrzeski Using a Large Language Model (LLM) as a production-scale clusterer runs into two problems: prompts cannot hold the whole label space, and serial per-document processing is too slow for real workloads. RAILS turns clustering into a simple loop over a growing label pool, using retrieval to surface relevant existing labels for each document, and scales through document batching with bounded concurrency. On six public benchmarks it beats the strongest prior LLM-clustering method on average, lifting accuracy from 51.2% to 59.3%, normalized mutual information (NMI) from 67.2% to 74.8%, and adjusted Rand index (ARI) from 45.4% to 54.7%. In a SaaS ticket-topic-discovery pipeline it has replaced an HDBSCAN stage with higher clustering quality, prompt-driven control, and stateful incremental operation.

arXiv:2609.24464 PDF

Fathom-Vaidya: Advancing Medical Reasoning with Rubric-Based Rewards

Kalash Shah, Kunal Singh, Snehan J, Shreyas Singh cross-listed Large language models (LLMs) in healthcare need both diagnostic reasoning, which converges on a diagnosis from clinical evidence, and clinical reasoning, which handles open-ended multi-turn interactions where no single correct answer exists; benchmarks such as HealthBench and MedXpertQA show current models are weak at both. The authors train sequentially: first reinforcement learning (RL) with rule- and rubric-guided rewards on MedBullets-derived diagnostic questions, then RL on 5.3k synthetic multi-turn clinical scenarios, each paired with multi-dimensional rubrics for grading responses. The approach yields over 10% improvement on MedXpertQA, and the 30B model reaches 50.1% accuracy on HealthBench-Hard, surpassing proprietary baselines including GPT-5 in thinking mode.

arXiv:2609.24480 PDF

On Emergent Capabilities and Model Merging

Luca Zhou, Emanuele Rodol\`a Model merging, arithmetic on the weights of fine-tuned checkpoints, is the most common way to combine capabilities cheaply, but its effect on emergent capabilities, behaviors that were never explicit training targets, had not been examined. Using two testbeds, activation oracles and emergently misaligned models, across three model families, the authors measure how emergent behavior survives merging compared with the explicitly trained capability that accompanies it. Three patterns emerge: merging two misaligned checkpoints retains most of their broad misalignment across the whole mixing range; no weighted merge of two single-task oracles reaches the auditing ability of a jointly trained oracle; and when only one parent carries an emergent capability, merging dilutes it significantly faster than the trained capability alongside it.

arXiv:2609.24504 PDF

LLJ Cards: Best practices for the Use of LLMs as Judges

Khaoula Chehbouni, Melina Medjdoub, Florian Carichon, Golnoosh Farnadi, Jackie Chi Kit Cheung Large language models used as judges (LLJs) have become a popular, scalable, and cheap substitute for human evaluation, but growing evidence questions their validity and reliability, and existing responses focus on bias mitigation and prompt engineering rather than on standardized, transparent, reproducible practice. LLJ Cards is a framework that synthesizes best practices from measurement theory, natural language generation evaluation, and machine learning into practical guidelines for designing and reporting LLJ-based evaluations. The framework gives practitioners a structured way to document the validity, reliability, and reproducibility of an automated evaluation rather than another technical fix for judge bias.

arXiv:2609.24516 PDF

Written as a Record, Read as an Address: What a Forward Pass Leaves in an Operation's KV Cache

Lingfeng Wu, Behzad Shomali When a language model reads an operation such as swapping the contents of two boxes, its forward pass writes keys and values for those tokens into the key-value (KV) cache, and the question is what those entries record and how they are later accessed. The authors split a forward pass into a frozen writer whose cache is recomputed without gradients and a separately trained reader that sees only the instruction and operation tokens with all state descriptions hidden, so anything the reader recovers must already exist in the unmodified cache. On a synthetic boxes task a base reader recovers at most 0.06 of queried bindings versus 0.75 to 1.00 after training, and operation-span transplants in Llama-3.1-8B and Mistral-7B causally redirect which visible state is read even when the two worlds hold identical values, revealing a routing record. Training also exposes direct access to the payload from the single operand-name token in a narrow mid-depth band of layers, though the model itself reads mainly the address and not the value, and the recipe extends to ToMi and GSM8K at a cost to open-book accuracy.

arXiv:2609.24635 PDF

iSDFT: Information-Proximal Self-Distillation for Continual Learning in LLMs

Ahmed Khaled Khamis, Xiaotong Ji, Hassan Jaber, Rasul Tutunov, Matthieu Zimmer, Jun Wang et al. On-policy self-distillation fine-tuning (SDFT) learns new skills from demonstrations with reduced forgetting, but it always distils toward the full demonstration-conditioned teacher and offers no control over how much demonstration information is transferred at each prediction state. Information-Proximal SDFT (iSDFT) treats the teacher as a budgeted information source, selecting at each token the distribution closest to the current student that satisfies a prescribed teacher-information constraint, which yields a closed-form exponential target with a locally determined tilt, and further anchors the student to its frozen base policy to limit cumulative drift. Across four heterogeneous backbones and two specialisation tasks, iSDFT improves on vanilla SDFT in 7 of 8 settings and matches it in the last, and 73% of retention evaluations stay within 0.5 points of the base model versus 52% for the strongest baseline, while also posting the largest mean gains on ten additional mathematics, coding, and competition-mathematics benchmarks.

arXiv:2609.24646 PDF

Muon Can Outperform Dedicated Continual Learning Methods

Sebastian George Sincari (Faculty of Mathematics and Computer Science, University of Bucharest, Bucharest, Romania), Bogdan Alexandru Gheorghe (Faculty of Mathematics and Computer Science, University of Bucharest et al. Continual learning with Low-Rank Adapters (LoRA) typically fights forgetting by penalizing overlap between a new update and accumulated past weights, and the authors ask whether a generic constraint supplied by the optimizer can replace that task-aware one. They train plain incremental LoRA (IncLoRA) with Muon, which orthogonalizes each update, and compare it against O-LoRA and ELLA over five seeds and three task orders on the Standard CL Benchmark and three seeds on TRACE. IncLoRA with Muon reaches the accuracy band of the dedicated methods on Standard CL and beats every AdamW configuration on TRACE, and adding a second update-constraining mechanism does not help, costing the most restrictive method 8.4 points of accuracy and its plasticity. The difference lies not in update size but in distribution: AdamW confines updates to 1.4 to 1.8 effective singular directions while Muon spreads them over 7.0, suggesting that part of the advantage attributed to dedicated continual learning methods comes from optimizer geometry.

arXiv:2609.24678 PDF

Adapting Tree-Structured Speculative Decoding to DeepSeek-V4 for Efficient Inference

Changxu Liu, Zhaogeng Li Repeated execution of the target model during autoregressive decoding dominates LLM inference latency, and tree-structured speculation, which keeps multiple candidate branches from shared prefixes, can improve acceptance over single-chain drafting at the same budget. Applying it to DeepSeek-V4 is nontrivial because its CSA/HCA online compressed attention causes branches diverging from a shared prefix to compress into different states, breaking cross-branch consistency during verification; the authors integrate tree speculation into the DeepSeek-V4-Flash pipeline with branch-aware causal verification, temporary state isolation, and accepted-path state refresh. Across budgets of 5 to 8 draft tokens, batch sizes 1 to 64, and GSM8K, MBPP, and ShareGPT, tree speculation achieves longer accepted lengths than matched linear configurations in every setting and improves throughput by up to about 18.5%, with relative gains growing with budget and largest for less predictable workloads at small-to-medium batch sizes, while throughput plateaus beyond a certain budget even as accepted length keeps rising.

arXiv:2609.24698 PDF

When Quantization Preserves Accuracy but Not Evidence: Explanation-Aware Post-Training Quantization for Medical LLMs

Yeji Kim, Mi-Young Kim, Randy Goebel Post-training quantization (PTQ) is usually optimized and judged by reconstruction error, perplexity, or answer accuracy, but in explanation-critical domains users also inspect the generated rationale, which may degrade even when the final answer survives. The authors propose an explanation-aware objective for transformation-based PTQ that builds an offline faithfulness cache from full-precision teacher rationales and uses it during optimization to preserve answer-supporting evidence tokens and evidence-conditioned answer behavior, instantiated on OSTQuant under W4A4KV4 quantization and evaluated on four 7B to 8B medical and instruction-tuned LLMs across MedExQA, MedExpQA, and ChallengeClinicalQA. A same-calibration OSTQuant baseline preserves task accuracy yet substantially weakens answer-supporting rationales, while the proposed objective better preserves the full-precision model's answer behavior and rationale-to-answer support, leading the authors to argue that PTQ for such settings should measure evidence preservation and not only accuracy.

arXiv:2609.24799 PDF

The Answer-Basin Representation Hypothesis: We Are Not Probing or Steering Concepts

Manjiang Yu, Hongji Li, Zihan Wang, Junwei Chen, Xue Li, Priyanka Singh et al. The Linear Representation Hypothesis links high-level concepts to directions in a language model's activation space, but how those linear structures are organized inside the model has been unclear. The proposed Answer-Basin Representation Hypothesis holds that the model's continuation distribution induces a probability measure over answers: all continuations yielding the same answer form an answer basin whose mass is their total probability, and the statistics of this answer measure are represented along linear directions shared across questions. Under this view, concept-related linear structure arises from differences in the answer measure rather than from changes in concept labels. Experiments across models and tasks tie concept-consistent effects, and their reversals, in probing and steering to how well concept labels align with the answer measure.

arXiv:2609.24821 PDF

OSWorld-Pro: Process-based Evaluation for Computer Use Agents

Zhilin Wang, Shaokun Zhang, Yifan Zhang, Hao Zhang, Jin Xu, Binfeng Xu et al. When a language model answers from a curated corpus through graph-based retrieval, a large accuracy gain from grounding does not show reasoning over the retrieved structure, because the shown context may already contain the gold answers. Exposure accounting classifies each gold item by whether the context exposes it and whether the answer recovers it, with the copy ceiling, the recall achieved by copying the context verbatim, serving as a deterministic, judge-free reference against which the model's gain over copy is measured. Across ten models, unaided recall averages 0.26 and grounded recall 0.92, yet gain over copy is negative for every model, between -0.067 and -0.022, and of 11,360 gold-item observations only three unexposed items earn lexical credit, none of which survive a model-judged relational audit. Rephrasing questions outside the graph's title vocabulary cuts exposure from 0.964 to 0.328, and the accounting is proposed as a standing control that separates exposed-item omissions from beyond-exposure recoveries without determining whether reasoning occurred.

arXiv:2609.24890 PDF

LoRA-generating hypernetworks for efficient on-device LLM generative personalization

Sean Augenstein, Li Ding, Jihwan Lee, Keith Rush, Andrey Zhmoginov On-device large language models (LLMs) on phones are constrained in scale by limited compute, but their close coupling to one user means they are used in predictable patterns that personalization can exploit. The method trains a hypernetwork that maps a user's context tokens to a low-rank adaptation (LoRA) suited to that user, so once the shared artifacts are deployed, each device synthesizes its own personalized LoRA using only forward passes. This combines the on-device feasibility of in-context learning (ICL) with the weight-based adaptation of parameter-efficient fine-tuning (PEFT), avoiding the latency cost of longer inputs and adding minimal storage because the hypernetwork partly reuses the target LLM's own weights. Experiments on several personalization datasets, focused on long-form text generation, show gains over ICL and PEFT baselines.

arXiv:2609.24979 PDF

30 more specialized papers

Other 69

When Is Availability-Aware Training Worth It? A Benchmark and Empirical Study of Interruption-Resilient Optimization Under Predictable Compute Schedules

Subhadip Mitra cross-listed Training under predictable but intermittent compute availability, such as satellites in eclipse, duty-cycled edge devices, or power-capped datacenters, is often assumed to require specialized availability-aware optimizers. The authors release OrbitTrace, a benchmark of 50 physics-grounded availability traces derived from SGP4 propagation of live two-line element sets across three orbital regimes, and test whether careful checkpoint-and-resume is already sufficient. When optimizer state is preserved across a gap, the gap is essentially free: a strong baseline that restores full optimizer state and indexes its learning-rate schedule in active rather than wall-clock time matches uninterrupted training on CIFAR-10 with ResNet-18 and exactly on a GPT-2 task, and previously reported advantages of availability-aware methods, including the authors' own AAT, come almost entirely from comparison against a weak wall-clock baseline. The one narrow regime where adaptation helps is large models that cannot persist optimizer state and face frequent short pauses, where reconstructing a decayed optimizer moment recovers only about 21% of the penalty and not robustly across seeds.

arXiv:2609.22087 PDF

Complex-valued Phase-Coherent Transformers

Leona Hioki Complex-valued Transformers have used softmax attention over the raw complex inner product, which stays near chance outside natively complex domains. The fix is a scaled cosine score: L2-normalize queries and keys so the score depends only on their cosine similarity, then hold it at order-one scale, since unnormalized scores stay at chance on ListOps and Needle and a normalized score at too small a scale also fails. The resulting phase-coherent Transformers (PCT) match or exceed the strongest real-valued baseline across long-range memory, positional retrieval, hierarchical reasoning, frequency-domain classification, and physical complex signals, show no degradation up to depth 20, and scale log-linearly over a 61-fold parameter range. One member combining complex screening with a phase-coherent recurrence is the first genuinely complex-valued network to solve Path-X, with 91.6% of its trainable parameters complex-valued versus 38.2% for S4.

arXiv:2609.22415 PDF

Causilo Technical Report

Minyong Cho, Minho Jeong, Dooho Lee, Jinmo Lee, Jaemin Yoo Causilo is a tabular foundation model (TFM) designed to pair frontier predictive accuracy with very fast inference. It keeps the column-then-row architecture of TabICL but inserts a row-refinement module that exchanges information among cell representations within each row after column encoding, then passes the refined cells through a second column stage before row compression, with both row stages using cross-attention through a fixed number of summary tokens so attention cost stays linear in the number of features. Pretrained on roughly 36M synthetic tables, it reaches 1785.4 Elo on TabArena at a median 0.10 seconds per 1K test samples, outperforming TabPFN-3.5-Fast with 31.6% less inference time and delivering strong results on BeyondArena and ScoringBench.

arXiv:2609.22866 PDF

Are Coreset Selection Methods Worth Their Cost?

Yangze Liu, Zhongyi Han Coreset selection chooses a representative subset of labeled training data to make training cheaper, but it is usually judged by accuracy at a fixed subset size, ignoring the time spent selecting and the training recipe behind each number. The authors build an end-to-end benchmark that standardizes downstream training and charges selection and training against the same wall-clock budget, covering four datasets from CIFAR-10 to ImageNet-1K, 11 selectors, five fractions, and three seeds, with more than 1,500 released runs. Across eight budget anchors on CIFAR-10 and Tiny ImageNet, no budget is won by a sophisticated selector; every winner is class-balanced random sampling, repeated random sampling, or full-data training, and on ImageNet-1K training on all data for fewer epochs beats every selection strategy while costing the least. A cost audit shows selection time is dominated by a fixed full-dataset scan that cannot be amortized by selecting less, and the authors also document nine correctness fixes to a widely used codebase, one of which shifts a standard Herding baseline by nearly 6 points.

arXiv:2609.22894 PDF

Silent Failures at the $2^{32}$ Boundary: A Technical Report on Large-Tensor Matrix Multiplication in PyTorch's Apple MPS Backend

Junichiro Niimi cross-listed Apple Silicon machines with 192GB or more of unified memory make it routine to place tensors with more than 2^32 elements on the GPU, and PyTorch's Metal Performance Shaders (MPS) backend silently returns wrong results for batched matrix multiplication at that scale. Sweeping torch.bmm over two dtypes, four memory layouts, six shapes, and 42 batch sizes on macOS 27.0, checked against float64 CPU references, the authors find relative errors above 1 with no exception or warning in every PyTorch release from 2.4.1 to 2.14.0. Three rules explain every outcome: transposed-view operands with outputs beyond 2^32 elements produce entirely wrong results that ignore strides, contiguous inputs beyond 2^32 corrupt only the batches past that point through index wraparound, and views of at least 2^31 elements raise an exception instead. Because torch.matmul and eager attention route through the same kernel, one oversized batch in a public sentiment classifier corrupted a third of its outputs, and the authors release the sweep harness plus a guard that blocks any MPS operation touching 2^32 or more elements.

arXiv:2609.22991 PDF

Optimizers for Diffusion Models: A Controlled Benchmark

Arman Bolatov, Egor Shulgin, David Li, Abduragim Shtanchaev, Sebastian U. Stich, Maxim Panov et al. Discrete diffusion models now rival autoregressive language models on several benchmarks, yet their optimizer is inherited from paper to paper without comparison, and new optimizers are validated almost only on autoregressive pretraining. This controlled benchmark compares seven optimizers, AdamW, Lion, Muon, SOAP, MARS, MARS-M, and Schedule-Free, across masked diffusion on text8, uniform diffusion on QM9 and LM1B, and Gaussian image diffusion on CelebA-64, with an identical search protocol and three-seed retraining of each winner at full budget. AdamW is a strong default but is beaten by a resolved margin on two of the four tasks, and the winner changes with the diffusion formulation. Optimizers validated on autoregressive pretraining transfer well, since Muon, MARS-M, and SOAP each beat tuned AdamW on at least one formulation, and all runs are reproducible from released code.

arXiv:2609.23055 PDF

SoK: Formal Methods for Fact-Checking and Information Integrity

Nikolaos Kekatos, Theodoros Nestoridis, Charalampos Bratsas, Charalampos Dimoulas, Georgios Konstantinidis, Georgios Malogiannis et al. Automated fact-checking systems typically return only a true/false verdict, without a record of which document settled the question, what would have had to differ for the verdict to change, or whether a reworded claim would be judged the same way; the authors call this missing piece a warrant. Formal methods can produce such evidence, and the Digital Services Act and AI Act increasingly demand auditable evidence about system behavior. Rather than organizing by pipeline stage, the survey sorts 121 works by what is being formalized, at five levels: the claim, the reasoning, the checking system, the ecosystem the claim spreads through, and the regulatory obligation. Most of the relevant formal machinery already exists but was built for other domains and has rarely been applied here, with the widest gap in verifying the checking system itself, and two routine professional steps, writing a claim in checkable form and correcting a published verdict, are not formally specified in any coded work.

arXiv:2609.23239 PDF

Perplexity Predicts Protection: Choosing Pretrained Backbones for Worst-Client Fairness in Federated Parameter-Efficient Fine-Tuning

Kiran Naseer, Samreen Azhar, Umar Shoaib, Haroon Mahmood, Muhammad Awais In federated learning, a client with far less data than its peers can be poorly served even when average accuracy looks fine, and the authors ask whether the choice of pretrained backbone under LoRA fine-tuning changes this, and whether per-word perplexity on the target text can predict the answer before training begins. Across 313 experiments on three text-classification datasets with RoBERTa, BERTweet, and PubMedBERT, each compared to a task-specific baseline on identical splits, lower-perplexity backbones consistently gave larger gains to the worst-performing client, with a rank correlation of -0.87 across nine dataset-backbone pairs, and a held-out backbone confirmed the pattern. Personalization with Ditto recovered only 4 to 12 percent of the gap between local training and full federation, removing aggregation erased the benefit entirely, and client and group updates were near-orthogonal, ruling out gradient conflict as the explanation. The practical advice is to measure perplexity on a sample of task text before choosing a backbone and not to rely on personalization to protect a data-poor client.

arXiv:2609.23463 PDF

Simpler Methods Work Better for L1 Penalized Logistic Models and Large Datasets

Edward Raff, James Holt Linear models with an L1-norm penalty remain standard for high-dimensional industry problems with more than a million features, yet many published solvers are slow, hard to parallelize, and awkward to train in production MLOps pipelines. The authors benchmark several recently proposed state-of-the-art solvers against older methods on large corpora and find the older methods far superior for general use. A simple baseline that runs LBFGS on a sub-gradient, dismissed in the literature for lacking convergence guarantees, proves highly effective with minor tweaks and is easier to support and scale in production. They also distill recommendations for avoiding overconfident results in academic solver research that block transition to production.

arXiv:2609.23995 PDF

Cost-Accuracy Trade-offs: Neural Operator vs Classical Numerical Solver

Daniel Zhengyu Huang, Andrew M. Stuart cross-listed Neural operators learn mappings from partial differential equation inputs such as coefficients, initial conditions, or geometries to solution fields, and can replace classical numerical solvers in many-query settings. The authors ask when and why such surrogates beat problem-matched classical solvers on cost for a given accuracy, in the post-training many-query limit where data-acquisition and training costs are treated as fully amortized, using a reproducible benchmark that measures prediction error, per-query floating-point cost, and wall-clock runtime. Neural operators are most competitive at low-to-moderate accuracy targets, and as required accuracy tightens classical solvers win, even in this deliberately favorable regime. Their floating-point advantage depends on avoiding temporal or nonlinear iterations or predicting a reduced quantity of interest, with extra wall-clock gains from dense tensor operations on modern hardware.

arXiv:2609.24021 PDF

Causal Bayesian Optimization: Foundations, Methods, and Applications

Chenfeng Huang, Thuy T. Le, Zixuan Ma, Hien Tran cross-listed Causal Bayesian Optimization (CBO) combines causal inference with Bayesian optimization (BO) to choose interventions sample-efficiently in systems with known or partially known causal structure. The survey organizes the field around a unified BO loop, showing how causal assumptions shape intervention search spaces, surrogate models, acquisition functions, and decision policies, and connects CBO to causal bandits, Bayesian experimental design, safe optimization, policy search, and causal abstraction. It also contributes a reproducibility-oriented benchmark with standardized GAP and a new trajectory-aware PA-GAP metric, evaluating seven CBO methods and a non-causal BO baseline across thirteen datasets and three budgets. No method dominates: rankings depend on dataset, budget, and metric, strong non-causal baselines stay competitive in several settings, and graph misspecification or omitted variables can reorder the rankings substantially.

arXiv:2609.24112 PDF

A principled approach for energy-efficient training via phase-aware GPU frequency tuning

Miguel Braga, J\'ulio Pinto, Rahma Nouaji, Olivier Michaud, Bettina Kemme, Oana Balmau et al. cross-listed A large share of the energy consumed while training AI models is wasted because GPUs stall on bottlenecks elsewhere in the pipeline. PAFT is a phase-aware GPU frequency tuning system that treats those stalls as an energy opportunity: it continuously monitors pipeline behavior and, when a GPU is bound to wait, lowers its clock frequency to match the pace of the bottlenecked device so execution time is unaffected, adapting as workload and system conditions change. Across twelve widely used models, PAFT saves up to 46% energy with an average overhead of 4% and outperforms all compared baselines.

arXiv:2609.24205 PDF

Artificial Structure Function Search: Preserving Artificial Functional Connectivity for Structured Pruning

Mindula Illeperuma, Rafael Pina, Charuka Herath, Sharmarke A. Gabayre, Varuna De Silva Structured pruning cuts the compute cost of deploying deep networks on constrained devices, but common criteria rest on opaque heuristics or weight magnitudes that say nothing about structural dependencies within the network. Artificial Structure Function Search (ASF-S) is a pruning framework built on Principle Gradient Importance (PGI), a candidate-selection criterion inspired by structure-function relationships in the brain, and it defines Artificial Functional Connectivity (AFC) so that the pruned structure respects the topographical organization of the output layer. Against recent benchmarks the method produces variants with 70% parameter reduction that recover baseline accuracy without re-training the pruned layers.

arXiv:2609.24401 PDF

Prior-Amortized In-Context Bayesian Inference for Generalized Linear Mixed-Effects Models

Alex Kipnis, Marcel Binz, Eric Schulz Bayesian inference for generalized linear mixed-effects models (GLMMs) gives calibrated uncertainty but depends on Markov chain Monte Carlo, where the No-U-Turn Sampler (NUTS) is slow and restarts from scratch for every new dataset, model, and prior. metabeta is a pretrained network for in-context posterior inference over GLMMs that accepts prior families and hyperparameters as test-time inputs, using two set transformers and conditional normalizing flows that mirror the global-plus-per-group structure of the posterior, trained on millions of simulated datasets with continuous, binary, and count outcomes. Its flow posterior is refined by Independence Metropolis-Hastings against the unnormalized posterior, so correctness rests on the sampler, giving tuning-free inference two to three orders of magnitude faster than NUTS while matching it on parameter recovery, calibration, and prediction. The posteriors remain faithful on out-of-distribution real data with misspecified likelihoods, collinear designs, and small samples, and the model is released open-source with weights.

arXiv:2609.24422 PDF

$t_0$: A Time-Series Foundation Model for Forecasting with Context

Lucas Meyer, Claudio Sole, Huikan Xiang, Nicolas Li, Lucas Franceschino, Arnau Quera-Bofarull et al. Most time-series foundation models forecast from target history alone and cannot use past or known-future covariates. t0 is a family of open-weights transformer forecasters, released as t0-alpha (102M parameters) and t0-beta (256M), whose layers alternate attention along time and across variates, output probabilistic quantile forecasts, and condition on multivariate context without task-specific retraining; pretraining mixes curated public data with synthetic generator families built to contain covariate-to-target dependencies. On GIFT-Eval, t0-beta reaches a CRPS of 0.4738 and a MASE of 0.6865, third on both and within 4.0% of the best zero-shot model, and it places third again on fev-bench; known-future covariates raise t0-alpha skill by 6.3 percentage points across 30 tasks, and in an independent evaluation of hourly ERCOT prices both models cut the lagged-price baseline's MAE by 38%.

arXiv:2609.24559 PDF

Augmented Hypothesis Testing with Persona-Based LLM Simulations

Ziyad Benomar, Aymen Al Marjani, Paul Missault, Saab Mansour A/B tests demand large samples, long timelines, and high cost, and although machine-learning predictions of outcomes carry useful signal, their unknown quality rules out replacing human experiments outright. The proposed learning-augmented hypothesis testing framework spans two prediction granularities: an asymmetric test for population-level binary predictions of the treatment effect's sign, with proven consistency and robustness bounds in the learning-augmented algorithms paradigm, and GPPI (Generalized PPI++), which extends Prediction-Powered Inference to nonlinear prediction errors through higher-dimensional transformations for individual-level predictions. Using persona-based LLM simulations, in which agents equipped with user personas predict individual behavior, as the prediction source, experiments on four real-world datasets show substantially reduced experimental cost while preserving rigorous statistical validity, with both methods gaining from accurate predictions and staying robust to inaccurate or adversarial ones.

arXiv:2609.24629 PDF

53 more specialized papers

Agents 67

BI-Agent and BI-Bench: Towards Automating End-to-End Business Intelligence

Chuxuan Hu, Yeye He, Penny Zhou, Wee Hyong Tok, Daniel Kang, Surajit Chaudhuri cross-listed Business intelligence (BI) workflows require users to find relevant tables, transform data, and build joins before answering a question, and it is unclear whether large language models can do this end-to-end. The authors harvest real-world BI projects and hand-extract question-answer pairs from user dashboards to build BI-Bench, then design a tool-augmented BI-Agent that decomposes the workflow into search, join, and transform subtasks and can be post-trained via supervised fine-tuning and reinforcement learning on trajectories synthesized from real projects. Frontier LLMs score under 50% on BI-Bench, while BI-Agent lifts vanilla models by up to 40 percentage points and post-training BI-Agent yields gains of up to 30 points.

arXiv:2609.20886 PDF

Scaling Discovery through Test-Time Communication

Jongho Park, Vasilis Kontonis, Shivam Garg, Akshay Krishnamurthy, Dimitris Papailiopoulos cross-listed Whether letting agents communicate at test time helps more than running them independently has had mixed answers. The authors study role-free teams of agents that share a directory, first on ARC-AGI-3, where a team of k communicating agents (team@k) matches the success rate of 4k independent agents, with the advantage growing with k and some tasks becoming reliably solvable only by teams. The gains carry to research-style tasks given enough compute: on polyomino packing teams beat best@k and the prior best-known score, and on MNIST classifier compression a four-agent team produced a 1,957-byte classifier at 99.4% accuracy, smaller than the best known human solution; independent agents remain better when compute is scarce or no clear progress signal exists.

arXiv:2609.21032 PDF

Verify, Don't Trust: Agentic Model Development for Video Discovery Retrieval at Scale

Hao Fu, Baiting Zhu, Minglei Chen, Yinjie Huang, Shuai Ding cross-listed Large language model (LLM) agents that propose, implement, and evaluate model changes can complete a run and still support an invalid conclusion when a code change is a no-op, data windows leak, evaluator semantics drift, or the two arms of a comparison pass through different serving funnels. EvoPilot is a human-gated method for long-horizon online autoresearch in which role-specific agents run each round through a versioned domain skill and typed adapter, with durable records of experiments and failures and deterministic checks that enforce recorded lessons. In a 37-day campaign on the retrieval system behind Video Deep Dive (VDD), a follow-on video discovery experience with an hourly refreshed index of hundreds of millions of videos, a primitive autoresearch attempt had wrongly blamed a 22 percentage point offline hit-rate drop on an interaction head; EvoPilot's verification traced the drop to a pre-existing evaluation defect, and after repair a matched comparison measured a 3.20 percentage point offline improvement from the head. A seven-day randomized online test estimated a 0.66% relative lift in the VDD slice of Good Search Result Rate for Retention, while durable state recovered an interrupted round and artifact reuse saved roughly five GPU-hours.

arXiv:2609.21257 PDF

Efficient Benchmarking in Production: A Study of an Evolving LLM Agent

Yining She, Lei Lin Production LLM agents are re-evaluated repeatedly as they change, but rerunning a full agent benchmark each time is expensive. Using 574 historical runs of the benchmark for a production analytics agent serving tens of thousands of monthly active users, split chronologically into calibration and held-out periods, the authors compare random sampling, historical caching, fixed representative subsets, and adaptive testing based on item response theory (IRT). Multidimensional two-parameter logistic adaptive testing gave the best score fidelity, reproducing full-run scores within 1.03 percentage points of mean absolute error while executing only 200 questions, or 38.5% of a full run. The team nonetheless deployed difficulty-stratified fixed subsets for operational simplicity, showing they transfer without recalibration to five other agent families and stay stable with calibration windows as short as one day, and they distill practical recommendations for recurring production-agent evaluation.

arXiv:2609.21267 PDF

Authorization Revocation for Long-Running AI Agents: Root-Scoped Quiescence under Delegation and Asynchronous Execution

Genliang Zhu, Chu Wang cross-listed Long-running AI agents outlive the processes that launched them through credentials, delegated tasks, queues, callbacks, reservations, and provider-side operations, so cancellation, process exit, or credential revocation neither closes every pre-existing carrier of authority nor distinguishes shared work that is independently authorized. The paper defines root-scoped authorization quiescence, in which a certificate accounts for every cut-relevant acceptance under a retired root authority before a local fence and excludes protected acceptance afterward, while allowing exact rebinding to a current, independently sufficient support. The accompanying protocol linearizes a root cut, fences old-root expansion, represents alternative and conjunctive authority as antichains of minimal sufficient root sets, composes provider certificates into a cutset, and reconciles transfers through exact channel-token accounting, with proofs under stated assumptions of post-cut issuer non-expansion, compositional soundness, merge-order independence, and crash/replay stability. In a provider-free late-effect test suite, cancellation-only and cut-only executions accept an already scheduled late effect that cut-plus-fence executions reject, and an independently implemented checker verifies 17/17 traces and rejects 44/44 semantic regressions.

arXiv:2609.21284 PDF

GameASG-Bench: Benchmarking Autonomous Software Generation for Game Development

Xiuhui Zhang, Yi Chen, Shusheng Xu, Fan Li, Huan Wang, Tongkai Yang et al. Autonomous software generation (ASG) can deliver a runnable application without establishing that its interacting components actually satisfy the specified behavior. GameASG-Bench makes behavioral testability part of the generation task for game development by declaring an evaluation interface before generation, fixing legal starting scenarios, player-level actions, stable snapshots, rejection behavior, and invariants while leaving implementations open, and checking results with static L1 source-compliance checks and browser-executed L2 checks that combine semantic observations with real input and runtime evidence. The benchmark comprises 47 browser-native game tasks across 12 genres in 2D and 3D, each with executable checks and an independently verified reference implementation. Across nine agent stacks, the best mean L2 check pass rate is 93.2%, but the best strict task success rate, which requires passing all applicable checks, is only 55.3%, or 26 of 47 tasks; for DeepSeek-V4-Flash, full tool access and larger turn budgets help while reasoning effort is not monotonic, and two harnesses that each achieve 18 strict successes overlap on only ten tasks.

arXiv:2609.21293 PDF

LEGIT: Credentialing Protocol for Trustworthy AI Agent Marketplaces

Steve Drew, Jiayu Zhou In emerging marketplaces where AI agents complete specialized tasks for buyers, buyers cannot easily tell which agent will perform best, since reported benchmark scores are hard to verify or compare across tasks, software, and budgets. LEGIT is a credentialing protocol that ties certification, reputation, and proposed marketplace allocation together: certification binds measured quality and cost per solved task to a specific agent configuration, task domain, evaluation budget, and evidence in a signed record, while reputation links records of past task outcomes to the same identity, subject to the reliability of reported feedback. Evaluations show agent configurations with similar observed task success can differ substantially in cost, and comparisons depend on the evaluation budget, supporting the case for binding performance claims to the tested configuration and resource limits. A complementary analysis quantifies the deposits and fees needed to make reputation manipulation costly under a stated Sybil attack model.

arXiv:2609.21325 PDF

DENSE: Distilling Agent Trajectories into Evidence-Grounded Shortcut Trees for Self-Refinement

Siyuan Liu (Fudan University, Meituan Longcat Team), Fan Yu (Fudan University, Meituan Longcat Team), Dongyu Ru (Meituan Longcat Team), Yizhu Liu (Meituan Longcat Team) et al. Deployed agents generate abundant execution traces, but task-specific verification and expert annotation are costly, so the question is how to turn those traces into reusable feedback without post-hoc outcome labels. DENSE (Distilling Evidence from Nested Subtask Executions) organizes evidence of local progress, recovery, and unfinished requirements into nested shortcut trees, compressing redundant attempts, reconciling issues across levels using recovery evidence, and summarizing completed branches while expanding unresolved ones. The accompanying REFIT protocol compares feedback methods from shared initial trajectories under outcome blindness, with environments and model contexts reset for fresh attempts at the same tasks. On Terminal-Bench 2.1, DENSE achieves the highest strict pass rate among tested non-privileged feedback methods across four recipient models, improving strict pass rate by 7.12 to 15.64 percentage points over initial executions while using 19.0 to 43.6% fewer recipient tokens in reruns, and GPT-5.5 ablations support combining nested subtask analysis with shortcut construction and issue reconciliation.

arXiv:2609.21423 PDF

GameLogicBench: Evaluating Coding Agents on Runtime Game Logic with Tick-Level State Assertions

Xinyu Che, Yunfei Ge, Shihao Li, Yanchen Liu, Hang Yan, Xinping Lei et al. cross-listed Coding agents can modify large projects, but in game development a run can end in a valid state even after violating gameplay rules along the way, and existing benchmarks replay fixed examples, score videos, or rely on model judges rather than checking rules throughout execution with reproducible verdicts. GameLogicBench contains 72 gameplay-logic tasks in Godot projects with an automated evaluator that checks each game's rules at every simulation tick across 403 hand-designed scenarios, expanded by seeded parameter variation into 1,451 test cases and validated by requiring acceptance of different correct implementations while rejecting mutants with one required capability removed. Across 20 model-and-scaffold combinations the best run solves 52.78% of tasks, and all twelve models under Claude Code solve fewer tasks as scope grows from isolated mechanics to repository-scale features, with most failing submissions runnable but behaviorally incorrect. An evaluator built without mutant validation let incorrect submissions pass, and agents were observed copying code from public repositories when network access was open.

arXiv:2609.21562 PDF

CIPL: A Channel-Aware Framework for Recoverable Privacy Leakage in LLM Agents

Tao Huang, Guosen Wu, Guolong Zheng, Jiayang Meng, Chen Hou, Xu Yang et al. cross-listed Privacy leakage in LLM agents is usually measured inside individual components such as memory, retrieval, or tool-use pipelines, which blurs the line between information that is internally exposed and information an outside attacker can actually recover. CIPL (Channel Inversion for Privacy Leakage) models a target as a chain of sensitive source, selection, assembly, execution, observation, and extraction stages and evaluates, under one shared black-box protocol, how selected sensitive units become attacker-recoverable output. Across memory-based, retrieval-mediated, and tool-mediated targets plus a live BrowserUse agent case study, where data is stored does not determine how recoverable it is: memory targets are near-saturated, retrieval-mediated leakage is often partial, and tool-mediated and live-agent leakage varies strongly with observation surface, prompt-to-channel alignment, retrieval depth, and provider behavior. A stratified semantic audit also surfaces attacker-useful disclosures that exact-match scoring misses.

arXiv:2609.21686 PDF

AutoRecLab: Describe the Experiment, Get the Code!

Moritz Baumgart, Philipp Meister, Justus Krell, Michael Schmidt, Bela Gipp, Joeran Beel Turning a recommender-systems (RecSys) experimental design into working code is a manual and error-prone step in empirical research. AutoRecLab is a Python-based autonomous lab that takes a natural-language research idea, derives explicit experiment requirements, builds and validates a prototype, and iteratively expands it into the full experiment, combining retrieval-augmented generation (RAG) for documentation lookup, static type verification, and execution-steered tree search. In a demonstration it implements an explicit-to-implicit feedback conversion study, and 8 of 9 baseline-comparison runs across six algorithms and three datasets succeed at roughly $1 per run using GPT-5.4-mini.

arXiv:2609.21863 PDF

AutoViewMem: Self-Configuring Orthogonal Views for Conversational Long-Term Memory

Zijie Cao, Xijun Qu, Zhicheng Gu, Xiaoshu Chen, Duanyang Yuan, Yanning Hou et al. Long-term memory systems for LLM agents usually store heterogeneous facts such as preferences, events, constraints, and temporal updates in one mixed representation, causing semantic interference that makes top-K retrieval noisy and leaves relevant evidence poorly ranked. AutoViewMem discovers candidate semantic views from interaction traces, selects a compact low-overlap set, and uses those views to drive structured, provenance-grounded memory extraction at write time, so ordinary top-K similarity search retrieves focused evidence without explicit routing or iterative retrieval, with offline consolidation keeping memories compact and consistent. On LoCoMo and PersonaMem with Qwen3-8B and Qwen3-14B backbones, it improves long-horizon question answering and personalization over strong memory baselines while keeping a simple inference pipeline.

arXiv:2609.21940 PDF

Bayesian Belief Layer for Controllable Opinion Dynamics in LLM Agents

Hafsa Akbar, Daniel Platnick, Marjan Alirezaie, Hossein Rahnama cross-listed LLM agents in social simulations revise opinions implicitly in context, so how open an agent is to persuasion can be neither specified nor verified and collective outcomes inherit the model's training prior. Bayesian Chronicle Agents (BCA) add a minimal belief layer that separates what an agent believes from how it speaks: each stance is a probability updated by one Bayesian step per utterance heard, with a single prior-strength parameter κ encoding stubbornness in the manner of Friedkin-Johnsen (FJ) opinion dynamics. Sweeping κ yields consensus, persistent disagreement, and committed-minority influence on demand, and persistent disagreement matches the FJ closed-form fixed points at R² of 0.93 to 0.99. Prescribed κ values remain recoverable after the language round-trip with perfect rank-order recovery across four models, and the explicit beliefs expose systematic per-model stance biases that end-to-end simulation would silently absorb.

arXiv:2609.21997 PDF

Value-Sensitive Delegation in Everyday AI Agent Use: Evidence from OpenClaw

Renkai Ma, Ruyuan Wan, Xuan Lu, Fan Yang, Chen Chen, Lingyao Li cross-listed Evaluations of autonomous AI agents usually measure task completion rather than the values users actually care about when delegating work. Applying Value Sensitive Design with LLM assistance, the authors coded 73,093 first-person Reddit posts about the OpenClaw agent for the human value at stake, the aspect of the agent involved, whether the value was fulfilled, and the user outcome, yielding 21 values in six groups such as Autonomous, Dependable, and Affordable Operation, Bounded Reach, Reviewability, and Equitable Access. Values clustered around the operating conditions users set for a run rather than around the agent's outputs, and were mostly met when users described what the agent delivered but mostly unmet when they described supervising it, a pattern the authors call value-sensitive delegation.

arXiv:2609.22067 PDF

CodeMidas: Scaling Agentic Coding RL Environments from Code Itself

Bowen Ye, Lei Li, Shicheng Li, Zihao Yue, Linghao Zhang, Hanglong Lv et al. Training coding agents with reinforcement learning (RL) needs many diverse tasks with reliable verifiers, but existing methods mine development artifacts such as issues and commits, which limits the tasks they can extract. CodeMidas is an agentic pipeline that uses source code as its only task-specific input: agents explore implemented functionality to write behavioural specifications, build tests grounded in executing the original code, and filter candidate tasks through execution checks and repeated solution rollouts. The resulting dataset holds 5,545 training tasks from 3,185 open-source codebases across 23 languages and 15 domains, and training MiMo-V2.5 on it with GRPO improves all five evaluated benchmarks, including +11.7% on DeepSWE, +17% on ProgramBench, and +8.5% on Terminal-Bench v2.1. Ablations show gains grow with the number of high-quality tasks, and trajectory analysis finds the trained agent explores codebases more and self-verifies in more varied ways.

arXiv:2609.22068 PDF

Designer-RSI: Evolving Procedural Memory from User Traffic for Agentic Graphic Design

Hongyang Du, Lan Yan, Christian Flores, Asim Kadav Professional graphic design is a long-horizon agentic task where editable artifacts emerge from many interdependent tool actions but no programmatic oracle can judge the outcome. The proposed framework keeps a frontier model frozen while it drives design software through more than 230 tools, and instead evolves an external procedural memory of natural-language skills that widens by acquiring procedures for recurring uncovered subtasks and deepens by revising existing procedures against their own successful and failed executions, with a matched replay gate admitting only changes that fix failures without regressing prior successes. Five rounds over 1,406 real user briefs and 1,869 automatically graded trajectories, with no weight updates or human labels, grew the skill bank from 76 to 139 and raised GenEval2 execution success on Claude-Sonnet-4 from 72.7% to 99.3%, with 61.8% and 67.6% win rates against a no-skill agent on Claude-Sonnet-4 and Claude-Opus-4.6. On 200 held-out briefs, widening or deepening alone reached roughly 49% win rates while their combination reached 58.5%.

arXiv:2609.22086 PDF

DeepInstructor: An Agentic AI Instructor for Experience-Driven Idea Evaluation

Rongcan Pei, Fang Guo, Qinglin Qi, Qi Zhu, Yun Luo, Jianhao Yan et al. Large language models can now generate research ideas faster than anyone can vet them, and existing automated evaluators lean on parametric knowledge or loose retrieval rather than the accumulated experience a human advisor draws on. DeepInstructor builds an Experience Graph from 58,607 peer reviews and runs a ReAct-style agent over it to retrieve dimension-specific evidence for novelty, significance, and feasibility, producing evaluations with a traceable reasoning path. The authors also release DeepInstruct, a dataset of controlled pairwise idea comparisons along those three dimensions. On it, the framework improves Hit@1 and Hit@2 agreement with human judgments by 24.4% and 29.7% over existing baselines.

arXiv:2609.22104 PDF

Beyond the Text: Verifying That Agent-Written Papers Are Backed by Their Artifacts

Qiuhong Shen, Benlong Wu, Hanjin Liu, Yuang Qi, Kejiang Chen As LLM agents produce research write-ups together with the code and experiments meant to support them, review practices that judge only the text cannot catch hard-coded metrics, unimplemented methods, or results with no execution behind them. ReAgent audits consistency between an agent-written document and its repository by extracting structured claims from the paper, then running a static pass that checks whether the claimed methods and experimental configurations appear in the code and a dynamic pass that re-executes relevant experiments to gather evidence. Combining the two surfaces cases either would miss alone, such as experiments that reproduce the reported numbers while deviating from the claimed methodology, and the findings are compiled into a traceable repository-level audit report. On a manually curated benchmark of agent-generated paper-repository pairs, the framework outperforms static-only and reproduction-only baselines at flagging inconsistencies.

arXiv:2609.22111 PDF

An Empirical Cost Attribution of Context-Compression Gateways in Multi-Turn Coding Agents

Luzhuo Chen, Jiayu Shi Context compression is widely pitched as a way to cut the token bill of LLM coding agents, and single-shot benchmarks show it preserves task quality, but neither fact establishes that compressing file reads saves money in a real multi-turn session. The authors instrument Paritok, a production compression gateway sitting between Claude Code and Codex and frontier models (Claude Sonnet, GPT-5), and attribute session cost to three levers under controlled A/B runs: tool-schema filtering, content compression of file reads and tool output, and history summarization. Tool-schema filtering removes a fixed 21K-57K tokens per turn and is the only unambiguously positive lever, while content compression saves only about 2% of the cache-priced prefix per turn but compounds quadratically, roughly 3350 times the square of the turn count, overtaking the fixed saving after about six turns. Because the gateway is non-destructive, recalling original bytes costs one bounded segment at a time rather than a multiplicative blowup, and the authors stress that a strong single-shot result (86.5% of SWE-bench quality retained at a 25.7% compression rate) is orthogonal to multi-turn cost and should not be cited as a savings argument.

arXiv:2609.22114 PDF

Success Leaves Detours: Learning Executable Walkthroughs for Long-Horizon Agents

Kaijie Chen, Chenyu Fang, Liang Yan, Bo Li, Bo Zhang, Peng Ye Agents that self-improve at test time by reusing past experience face a problem: sparse-reward trajectories are full of failures, loops, and detours, and natural-language summaries drop the state conditions and action dependencies needed to actually re-execute a procedure. Trace compiles noisy trajectories into executable Walkthrough Memory by detecting progress anchors from rewards and persistent state changes, propagating credit to find valuable transitions, estimating action prerequisites from cross-episode success and failure evidence, and then backward-slicing dependencies to extract consistent action chains while discarding irrelevant steps. Each Walkthrough carries entry conditions, ordered state-action-effect steps, and completion and failure predicates, enabling reuse, mid-procedure resumption, and programmatic verification. On J-TTL, WebShop, and ScienceWorld with three open-source LLMs, it beats eight test-time learning and memory baselines, improving average AUC and Final-3 by 30.0% and 40.5% over the strongest baseline while using fewer inference tokens.

arXiv:2609.22120 PDF

StepKV: Step-Aware KV Cache Compression for LLM Agents

Boyu Feng, Jiahong Liu, Yifan Li, Wenhao Yu, Zexuan Qiu, Yuliang Sun et al. Multi-step LLM agents accumulate long trajectories of reasoning, tool calls, and retrieved observations, and existing KV cache pruning treats this as a flat token stream ranked by recency or attention saliency. That mismatch between the unit of compression and the unit of reasoning means an early observation or intermediate decision that receives little recent attention can be pruned even though later evidence synthesis depends on it, a failure the authors call Reasoning Continuity Disruption. StepKV makes reasoning steps first-class retention units: it tags cache entries with their generating step, estimates step utility from trajectory-derived signals, combines that with token-level saliency, and keeps the top-scoring entries under a target budget. On multi-hop QA and long-horizon web reasoning tasks it sustains accuracy at low KV budgets where token-level baselines degrade sharply.

arXiv:2609.22158 PDF

Beyond Task Completion: Training Capable and Safe Computer-Use Agents

Zeyu Kang, Zhenyun Yin, Yang Zhang, Shan He, Shanzhe Lei, Yanjiu Zhong et al. Computer-use agents (CUAs) that are post-trained only for task success do not learn reliable safety behavior: a dependable agent should finish benign tasks, work around environmental hazards when a safe path exists, and refuse when the goal is harmful or no safe path remains. Safety and Capability Optimization for Policy Execution (SCOPE) trains this conditional policy jointly, using SCOPE-Gen, a pipeline that synthesizes verifiable capability tasks and converts them into paired environment-risk variants with the same goal, to build SATraj-OS, a trajectory dataset of capability demonstrations, safe continuations, and explicit refusals. Training proceeds by supervised fine-tuning on all three trajectory types followed by online reinforcement learning to improve task completion. Starting from Qwen3.5-9B, SCOPE-RL reaches 54.17% task success on OSWorld and 64.30% attack avoidance on OS-BLIND, the best aggregate capability-safety score of 58.80% among evaluated agents, and ablations show refusal trajectories drive most of the attack-avoidance gain while risk-handling trajectories preserve more task utility at comparable avoidance levels.

arXiv:2609.22178 PDF

A Pinch of SFT, A Dash of RL: When Reinforcement Learning Helps Long-Horizon Advertising Agents

Aakash Kolekar, Sahika Genc, Bunyamin Sisman, Shahriar Shariat, Shree Vandana Kachroo, Avishek Saha et al. Enterprise analytics agents must chain retrieval, reasoning, API calls, and code execution over distributed business data, and it is unclear how to split post-training between supervised fine-tuning (SFT), which calibrates tool syntax and teacher-supported behavior, and reinforcement learning (RL), which can find reward-supported behavior beyond demonstrations but may disrupt already-calibrated skills when applied uniformly. In a controlled experiment on production-mirroring beta advertising APIs, checkpoint trajectories fell into three regimes (Imitation, Lift, and Discovery), which the authors turn into a prospective diagnostic that routes each feature to SFT only, SFT then RL, more RL allocation, or further environment development based on teacher support and reward-observable headroom. The diagnostic predicted 15 of 18 subsequent feature-specific trajectories, and targeted SFT-then-RL on GPT-OSS 120B produced positive point estimates on 7 of 8 advertiser skills against a frontier control, five with paired 95% confidence intervals excluding zero and one confidence-supported regression. A subject-matter-expert audit found targeted RL cut standard leakage from 11.8% to 2.9% and adversarial leakage from 22.9% to 6.8% relative to SFT while preserving actionability, and in a matched comparison targeted RL beat uniform RL on the seven-skill mean delta (+3.57 versus +1.62) using 43% less incremental RL compute.

arXiv:2609.22194 PDF

EvoRank: LLM-Guided Evolution of Multi-Objective Learning-to-Rank Pipelines

Rayhan Patel, Shabaz Patel EvoRank is an LLM-guided evolutionary loop that discovers complete Learning-to-Rank pipelines, spanning features, models, losses, and ensembles, for multi-objective e-commerce search. On the Expedia ICDM 2013 dataset with relevance, conversion, and revenue as competing objectives, three independent runs each converge within 50 iterations, at a cost of about ten dollars, on interpretable pipelines that beat an Optuna-tuned LambdaMART on 60k held-out queries, an advantage that persists at full data scale and places in the top 6 percent of the original competition. An earlier campaign that evolved only training objectives appeared to succeed on its small selection fold, but a transfer audit re-scoring winners on held-out data showed the gains were almost entirely fitness noise, and neither seeded domain knowledge nor richer diagnostic feedback changed what transferred. The authors distill this into a headroom gate that compares search-space headroom to fitness noise before any LLM spend, and release the system, auditing tools, and a catalog of failure modes with guardrails.

arXiv:2609.22196 PDF

Schematize: An Agentic System for Generating and Refining Information-Extraction Schemas for Legal Research

Albert Sawczyn, Jakub Binkowski, Kamil Tagowski, {\L}ukasz Augustyniak, Berenika Kaczmarek-Templin, Tomasz Kajdanowicz Empirical legal research turns research questions into structured data extracted from large collections of rulings, but designing the extraction schema and running the extraction remain a manual, expertise-heavy bottleneck. schematize is an open-source multi-agent system that interactively converts a researcher's problem statement into a validated extraction schema through a clarification dialogue that elicits implicit expert intent, iterative schema generation, data-grounded refinement that tests the schema against real documents, and chat-based post-editing. In an evaluation with legal professionals using a newly introduced methodology, the system achieved top performance in most tested configurations. Although designed to be domain-agnostic, it is tailored and evaluated on legal problems and released as a pip-installable Python package.

arXiv:2609.22209 PDF

Toollery: Scaling LLM Agents to Thousands of Skills and Tools

Xiangxi Tian, Ran Guan When LLM agents have access to hundreds or tens of thousands of skills, tools, and API functions, putting the full library in the prompt becomes costly, slow, and error-prone because every added candidate raises token count and latency and adds distractors. Toollery is a training-free candidate-compression framework that, following document-side query expansion, generates user-intent queries from each skill or tool specification and builds a retrieval index that maps real requests to a compact top-k candidate set before the LLM makes its final selection, treating high-level skills and atomic tools alike as selectable capabilities. Evaluated on the roughly 79K-capability SkillRouter benchmark, BFCL-V4 with over 440 tools, and 3,396 proprietary smart-cockpit requests over 220 tools, it improves recall over ordinary specification retrieval while keeping online selection bounded to a compact candidate set, improving end-to-end selection on the cockpit data at a fixed top-10 budget and maintaining comparable AST accuracy on BFCL-V4. The authors note that quality and cost gains depend on workload coverage and provider caching.

arXiv:2609.22218 PDF

Can Coding Agents Reproduce Official Statistics? Metadata, Retry Budget and the Limits of Execution Feedback in a Controlled Eurostat Benchmark

Sabina-Cristiana Necula A coding agent can produce data-analysis code that runs, but successful execution does not guarantee a valid official-statistics result. The authors built a benchmark of 30 natural-language tasks over seven Eurostat datasets and four difficulty tiers and ran Claude Sonnet 5 under four conditions: task only, task plus a frozen metadata card, metadata plus a repair loop driven by sanitized execution feedback, and metadata plus the same attempt budget with no diagnostics at all, for 360 task-runs scored on exact correctness of dataset, filters, output shape, values, and unit. A companion experiment with an under-specified output contract, where the required ranking key and unit representation were never stated, understated the feedback condition by 23.4 points, showing that evaluator and contract design can dominate measured agent error. The conclusion is that reliable statistical coding agents need semantic validation against frozen specifications, a fully specified output contract, and a retry budget, rather than execution diagnostics.

arXiv:2609.22222 PDF

EAVer: Long-Form Factuality Verification as an End-to-End Agentic Policy

Kening Zheng, Aoying Zheng, Zhigang Chang, Yazhi Guo, Miaotian Guo, Qingwei Zong et al. Long-form factuality checking usually runs as a fixed decompose-search-verify pipeline that treats every claim independently, so model and search calls grow with claim count and overlapping evidence gets fetched repeatedly. EAVer instead trains a single policy that groups related claims, routes each group to direct verification or targeted search depending on confidence, and keeps returned evidence in compact in-context memos for reuse across claims. Training data comes from a privileged-teacher pipeline that turns gold claim annotations into 1,447 executable multi-turn tool trajectories with live search, plus 794 same-trajectory preference pairs for decision-focused Direct Preference Optimization (DPO) over the factuality-decision tokens. On Qwen3-8B, the policy beats the strongest search-based baseline by 2.88 Macro-F1 on VeriFastScore and 4.73 on the out-of-distribution FaStFact-Bench while using about 80% fewer searches than the most search-efficient baseline, with gains holding from 4B to 32B parameters.

arXiv:2609.22223 PDF

EvalMem: An Operation-Level Diagnostic Framework for Long-Term Memory Systems

Zeyu Liu, Jian Zhong, Rongduo Han, Ziyang Wu, Shunye Tang, Chenghao He et al. Memory systems for long-running LLM assistants are usually judged by end-to-end question-answering accuracy, which cannot tell whether a failure came from encoding, retrieval, or generation. EvalMem runs three parallel examiners per query: one checks whether the target fact was stored, one checks whether the system's own retriever returns usable evidence, and one checks whether the model can answer from oracle evidence, together producing multi-label defect codes. A recall-first agentic retrieval-augmented generation (RAG) strategy that searches with both the query and the source evidence raises recall of present evidence on LoCoMo from 70.2% to 95.6%, sharpening store-level diagnosis. Across seven memory systems on LoCoMo, LongMemEval-S, and DynaMem-Bench, retrieval is the most frequently attributed failure layer, at 22.1% of defects on default LoCoMo versus 7.7% for encoding and 6.5% for generation, and a search-friendly auxiliary structure called MemWiki built from each system's memory export lifts mean accuracy by 2.5 and 2.3 points on the first two benchmarks.

arXiv:2609.22231 PDF

BizSage: A Self-Evolving Multi-Agent Framework for Business Research with Efficient Knowledge Retrieval

Yuhe Wu, Guangyu Wang, Jiaxin Liu, Guang Zhang Multi-agent systems built on large language models (LLMs) have automated parts of academic research, but economics and business research pose two extra problems: the evidence a task needs is scattered across paper sections rather than whole papers, and the fields' empirical rigor demands a way to learn from evaluation feedback. BizSage builds a Lateral Knowledge Graph (LKG) by merging section-level knowledge graphs and applies Personalized PageRank (PPR) to surface sections that are both semantically relevant and structurally important, while seven specialized agents work under a Meta-Review self-evolution loop that distills failure modes from evaluation traces into reusable strategies. On a benchmark covering four domains and three tasks, the system ranks first on most metrics, wins more than 60% of pairwise comparisons against each of six baselines, and produces zero hallucinated citations.

arXiv:2609.22235 PDF

CHART: A Harness-Rotation Curriculum for Harness-Robust Search Agents

Xinlu Zhang, Ying-Chun Lin, Zhihan Zhang, Besnik Fetahu, Xi Chen Search agents post-trained under one harness, such as a particular system prompt, often lose learned behaviors when the harness is rewritten for production even though the task is unchanged. Studying parallel search as the target behavior, the authors find that training on a fixed harness makes it harness-local, that naive harness augmentation fails because GRPO depends on a reward gap between parallel and serial rollouts that a small pool saturates and a large pool dilutes, and they propose Curriculum HArness Rotation Training (CHART), which periodically graduates harnesses whose behavior has been learned and swaps in still-learnable ones. CHART parallelizes on 89% of held-out harness turns versus at most 5% for static harness pools, learns parallel search on every harness in the pool where static augmentation succeeds on at most half, and improves pass@1 by 5.6 percentage points on a new question-answering task and search environment.

arXiv:2609.22247 PDF

Strategy Accumulation and Guided Execution for Automated LLM Fine-Tuning

Haoran Zhao, Wei Du, Dingwen Yang, Jixuan Huang, Junlin Shang, Lingyong Fang et al. Automated fine-tuning systems find effective training strategies with little manual effort but are stateless, discarding discovered strategies, dataset insights, and hyperparameter findings after each search so every new task starts cold. Strategy Accumulation and Guided Execution (SAGE) runs a multi-agent pipeline that explores with Monte Carlo Tree Search while a parallel Distillation Agent extracts task-specific exploration records and confidence-scored cross-task insights into a structured experience repository, then retrieves and filters that experience to guide training on new tasks. On nine unseen tasks spanning single- and cross-category settings, accumulated experience raises the average relative improvement over baseline from 3.2% to 15.6% in single-round execution, a 12.4-percentage-point gain over the same pipeline without it.

arXiv:2609.22257 PDF

RS-Claw-Evolution: Environment-Feedback-Driven Evolution for Lightweight Remote Sensing Agents in Long-Horizon Tasks

Kai Ouyang, Dongyang Hou, Liangtian Liu, Zeyuan Wang, Ziyu Li, Chengfu Liu et al. Lightweight remote sensing (RS) agents built on compact language models struggle with multi-step geospatial tasks because they lose long-horizon state, use environment feedback poorly, and receive sparse optimization signals. RS-Claw-Evolution improves such agents in three stages: interaction evolution uses executable code to control observations and hold intermediate state, experience evolution generates failure-aware trajectories with error-turn masking so the agent learns from recovery without imitating faulty actions, and decision evolution applies reinforcement learning with multi-dimensional environment rewards and turn-level advantage protection for better credit assignment. On Earth-Bench, the trained Qwen3-4B agent reaches 65.9% accuracy in Autonomous Planning mode, beating untrained Qwen3-32B at 43.8% and DeepSeek-V3.1 at 60.8% and approaching GPT-5 at 71.6%.

arXiv:2609.22258 PDF

Which Part of the Context Layer Does the Work? Separating Semantic Content from Retrieval Scaffolding in Text-to-SQL Agents

Qing Ye Context layers, curated documentation an analytics agent fetches at query time, boost text-to-SQL accuracy, but a with/without comparison cannot tell whether the gain comes from the semantic content, the retrieval scaffolding that delivers it, or accompanying pre-computed views. A four-arm ablation on DABStep across four models uses a YAML data contract as the instrument, with one arm stripping every field of prose while holding the tool surface, retrieval instruction, table allow-list, and operation rules byte-for-byte fixed, and another compiling the contract's SQL into views to establish a ceiling. Semantic content dominates, lifting hard-task accuracy from 13.9% to 55.1% on the weakest model and 37.0% to 77.4% on the strongest, and beats the same knowledge pasted into the prompt on every model, while scaffolding without content is worth 0 to 5 points on two flash models and 14 to 15 on two frontier models. The practitioner guidance is semantics first, scaffolding second, and pre-computed macros only where an agent demonstrably fails to derive a rule, with governed arms issuing no mutating SQL statements against 166 from ungoverned arms.

arXiv:2609.22259 PDF

Semantics Delivery Network: Rethinking Web Retrieval Infrastructure for LLM Agents

Peichun Hua, Yunming Xiao cross-listed Large language model (LLM) agents that use retrieval-augmented generation (RAG) consume short semantic chunks selected for task utility, yet search engines and content delivery networks (CDNs) still serve URL-ranked snippets and whole cached objects built for human browsers, and uncoordinated agents repeat the same fetching and processing work. The authors argue that chunk-level semantic retrieval should become a first-class network abstraction and propose SemDN, an origin-authorized hierarchical edge layer that indexes, searches, and caches participating websites at chunk granularity, amortizes acquisition across agents, and supports per-tenant retrieval policies. Because semantic retrieval has no explicit cache-miss signal, the design must estimate when its corpus is incomplete or stale and trigger scoped discovery or refresh. Preliminary probes show a large gap between page content processed and chunks actually consumed, substantial task-local reuse, and higher answer quality per context token from chunk delivery.

arXiv:2609.22486 PDF

AutoGym: Blueprint-First Generation of Verifiable Agent Gyms

Aarati Andrea Noronha, Kavya Ravikumar, Carly Xiaoyu Lin cross-listed Training agents with reinforcement learning requires a gym made of a task, an executable environment, and a verifier that reliably separates success from failure, yet building these remains manual, task sets saturate as models improve, and single-pass synthetic generation yields tasks whose difficulty is mostly cosmetic and must be graded by unreliable LLM judges. AutoGym generates complete gyms from a minimal domain seed or prior model trajectories using three mechanisms: blueprint-first generation that fixes the valid solution space, environment requirements, and verification criteria before the environment is built, so solvability is a construction prerequisite; explicit parameters controlling task topology, interaction depth, capability axes, question obfuscation, and distractors for fine-grained difficulty steering; and active curriculum synthesis that recalibrates the parameter distribution from observed model performance. Across productivity and temporal-reasoning settings, the generated gyms span the capability spectrum and include instances that challenge frontier models.

arXiv:2609.22592 PDF

Counterfactual Tool Ranking under Utility, Cost, and Privilege Constraints

Jiapeng Li Ranking which tool an agent should call from logged interactions requires separating who has authority to act, what the historical logs actually support, and what a counterfactual comparison estimates. The study uses eleven executable enterprise-inspired tools with exact logging propensities over real local Model Context Protocol transport, keeps an initial 45-run synthetic study, and then challenges it with 30 realized-return control runs and 15 experiments on 1,930 independently released Berkeley Function Calling Leaderboard (BFCL) tasks. Full-return direct regression reverses an initially favorable doubly robust result in the linear setting, with mean absolute errors of 0.0139 versus 0.0272, although doubly robust estimation keeps a clear edge under environment shift, and two pinned local Qwen2.5 models show a strong failure to abstain on held-out tasks. The authors also show that unsupported actions shared by two policies cancel, allowing point identification of an incremental change when neither absolute value is identifiable, and frame the contribution as a falsifiable evaluation method rather than a new estimator or a production safety claim.

arXiv:2609.22819 PDF

AgentRouter: Heterogeneous Model Routing for Cost-Optimal Multi-Step Agentic Workflows

Rudrendu Kumar Paul, Sourav Nandy cross-listed Enterprise agentic systems that send every trajectory step to a frontier model waste much of their inference budget on subtasks that smaller models handle equally well, and existing routers optimize single-turn query assignment rather than the step-to-step variation in complexity inside one agent trajectory. AgentRouter formalizes step-level routing as a sequential assignment problem and uses a 12M-parameter classifier with under 5 milliseconds of overhead per step on an A100 to map each step to one of four model tiers using five features available at routing time, trained on 50,000 annotated trajectory steps spanning planning, coding, research, and data analysis. It achieves a 72% cost reduction relative to frontier-only baselines while retaining 97.3% of frontier-only quality, with per-step routing accuracy from 91% on minimal-complexity steps down to 76 to 82% on mid-range and frontier tiers. On the same benchmarks, RouteLLM and FrugalGPT applied per step reach only 31% and 44% cost reduction because their single-turn training signal misses trajectory-level quality dependencies.

arXiv:2609.22951 PDF

Automatic multimodal UX improvement recommendations from LLM agent user simulations

Anu Chowdhury, Bin Wu, Hossein A. Rahmani, Emine Yilmaz Evaluating user experience on live websites with human testers is expensive and hard to scale, and existing large language model (LLM) agent simulations are text-only and need manual review to turn traces into insights. AMUSER is a multimodal framework that simulates user behaviour on a site and then generates a ranked list of UX improvement recommendations, framed as a structured generation and ranking problem scored by expert annotation and an LLM judge. On commercial websites its recommendations reached NDCG@3 of 0.758 versus 0.359 for text-only simulation at 89% lower simulation cost. The authors find the role of multimodality is asymmetric: screenshots during simulation produce richer traces, but feeding visual inputs into the recommendation step slightly hurts quality.

arXiv:2609.22971 PDF

Triggers and Diagnostics for LLM-Based Interpretability Failures in Active Inference Agents

Param Raval, Rohit Shenoy, Archana Vaidheeswaran LLM explainers are increasingly used as runtime oversight for autonomous agents, with operators reading a generated account of the agent's beliefs rather than inspecting its internal state. The authors pair an Active Inference (AIF) agent that tracks German grid demand with explainers running on GPT-4o, Claude-3-Opus, and Gemini, then probe the pair with three black-box triggers: observation corruption, objectively wrong actions, and attacker-controlled text in observation metadata. Corrupting observations by 600 MW per step shifts the agent's posterior by 490 MW, yet none of the 30 explanations flag it, and on timesteps with objectively wrong actions all three explainers produce a sycophantic rationalization 80-95% of the time. Injected metadata steers the explainer and enables data exfiltration on all three backends, and mitigations are proposed but not evaluated.

arXiv:2609.23215 PDF

PSD: Pseudo Self-Distillation of Memory Representation Capabilities for LLM Agents

Pirzada Suhail, Menglin Xia, Xuchao Zhang, Mayukh Das, Chetan Bansal, Saravan Rajmohan cross-listed Memory systems for LLM agents are costly to build and maintain because they depend on repeated calls to large proprietary models, and standard distillation cannot help since closed models expose no logits or hidden states. Pseudo Self-Distillation (PSD) trains a single small language model in two roles: a teacher that sees a privileged prompt containing the oracle's answer as reference context, and a student that sees only the task prompt and learns to match the teacher's output distribution, absorbing oracle-guided behavior without any access to the oracle's internals. On LoCoMo, PSD-trained Qwen3 models at 0.6B, 1.7B, and 4B parameters match or exceed GPT-4.1-mini on downstream retrieval at a fraction of the deployment cost, with off-policy PSD performing best in most conditions, and the learned memory-construction skill transfers to LongMemEval despite no training on that benchmark.

arXiv:2609.23449 PDF

Propose, Verify, Commit: Evidence-Grounded Memory for Long-Horizon Multi-Actor Conversations

Zihao Lu, Zhihang Yuan, Lei Shi Conversational memory over long horizons is hardest when many participants contribute evidence across contexts and earlier facts get revised later. EGMEMORY models this as a searchable state machine that keeps persistent message-level evidence separate from an explicit active state; at write time an evidence-grounded propose-verify-commit protocol with adaptive state resolution governs how the state evolves, and at read time adaptive evidence navigation iteratively resolves the state and supporting evidence a query needs, using conversational structure to narrow the search and lexical-semantic relevance to rank candidates. The system runs entirely through prompting and tool use with no memory-specific training, and it reaches 68.2 percent on GroupMemBench and 77.9 percent on EverMemBench, beating the strongest baselines by 22.7 and 21.4 points, while also scoring 73.6 percent on the dyadic LoCoMo benchmark.

arXiv:2609.23465 PDF

RPMem: Learning Long-Term Recurrent Parametric Memory Across Sessions for LLM Agents

Fanyu Zhao, Ruike Cao, Liang Dong, Fugen Yao, Jian Xu, Guanjun Jiang et al. Long-running LLM agents need memory that persists and evolves across sessions, but text-based memory grows ever more dependent on retrieval quality as histories lengthen, and existing parametric memory is tied to one backbone and offers little support for cross-session evolution. RPMem compiles each session into a model-independent latent memory through forward computation, merges it with retained memory via a task-trained recurrent gate, and then maps the consolidated memory to backbone-specific low-rank adaptation (LoRA) parameters, so the encoding capability survives a model swap. Across three long-term memory benchmarks and five backbones it generalizes with near-constant update cost and memory footprint, and with Qwen3-8B on PERMA it reaches 85.52 percent, ahead of the strongest parametric and text-based baselines by 5.32 and 12.98 points; ablations confirm that session compilation and cross-session consolidation play complementary roles.

arXiv:2609.23466 PDF

BabelArena: A Large-Scale Multilingual Benchmark for LLM Agents

Peng Kuang, Yuchun Fan, Jiangnan Li, Minghao Wu, Jialong Tang, Hao-Ran Wei et al. Agent benchmarks are overwhelmingly English-centric, leaving open how well LLM agents execute multi-step, tool-using workflows in other languages. BabelFlow is a benchmark-agnostic agentic workflow that ports existing agent benchmarks to new languages by analyzing runtime dependencies, coordinating structure-preserving translation, and layering automated verification with human review so task and evaluation semantics survive; applying it yields BabelArena, 16,146 instances derived from 702 canonical tasks across four benchmark families, 13 domains, and 23 languages. Experiments with five frontier models show no single model dominates across families, and lower-resource languages fail differently, with larger shares of tool-use and control-flow errors rather than answer-quality errors alone. Agents in low-resource languages also consume up to roughly twice the input tokens of English on the same tasks without proportionally longer interactions, and language consistency degrades further on structured-output tasks, where switches go overwhelmingly toward English.

arXiv:2609.23490 PDF

VibeMemBench: Evaluating Memory Systems for Coding Agents on Real Repository Coding Tasks

Liyang Fan, Yingcheng Shi, Yongbin Li, Chenghao Sun, Xin Chen, Xander Xu et al. cross-listed Persistent memory systems promise to let coding agents reuse experience across tasks, but existing benchmarks either test code changes without isolating memory or score recall without measuring coding outcomes. VibeMemBench evaluates memory systems on 111 coding targets drawn from 90 SWE-rebench V2 repositories, paired with 3,634 history trajectories from those repositories, where each target is kept only if injecting past experience verifiably improves its executable test outcome in a reference setting. Transferring this frozen, verified experience to five held-out solvers by direct injection raises task resolution on four of them by 1.1 to 4.5 percentage points and cuts agent steps on all five. When four existing memory systems must build and retrieve experience from the same history themselves, eleven of twelve solver-system pairings fail to beat the memory-off baseline, exposing a gap between the useful experience repositories contain and what current memory systems deliver.

arXiv:2609.23570 PDF

FLARE: A Full-Lifecycle Dense Supervision Paradigm for Long-Horizon Coding Agents via Generative Reward Model

Jingxuan Xu, Gang Wu, Yanan Wu, Yutao Mou, Songwei Yu, Tianzhuang He et al. Long-horizon software-engineering agents trained with sparse pass/fail rewards face a credit-assignment problem and discard failed trajectories. FLARE (Full-Lifecycle Alignment and Reward Engine) uses RADAR, an offline causal-chain backtracking diagnostic, to distill a lightweight generative reward model that gives step-level risk feedback; at inference it intercepts high-risk steps for localized breakpoint re-execution, and in post-training its signals serve as reranking scores for supervised fine-tuning and dense step rewards for reinforcement learning. A single FLARE run outperforms five-sample global rollout while using 5x fewer tokens, and training with it yields a 19.13% relative gain in supervised fine-tuning and a 9.19% improvement in reinforcement learning.

arXiv:2609.23808 PDF

XYEval: Agents say yes to bad advice

Zhengxuan Wu, Yuxuan Li, Oyvind Tafjord, Been Kim The XY problem is a communication pitfall where a person asks about their attempted solution instead of their underlying goal, and the authors extend sycophancy evaluation to this pattern in agentic settings. XYEval is a meta-evaluation framework that mutates an existing benchmark so that users offer plausible but misleading suggestions, and it is applied to five models across six benchmark suites. Agents suffer relative performance drops of up to 46.7% under XY mutation, and on τ²-bench the drop widens when a pedantic user demands detailed explanations before approving a better solution. A system-instruction baseline that raises awareness of XY problems only partially mitigates the effect, and trace analyses show agents fail both to recognize misdirection and to explain the real problem.

arXiv:2609.23939 PDF

MobileCybench: Evaluating Agent Vulnerability Discovery via Executable Probes

Andy K. Zhang, Ava Huang, Joey Ji, Wai Han, Thomas Qin, Nardos Demilew et al. cross-listed AI agents now file vulnerability reports faster than maintainers can review them, and validating each report requires understanding application-specific security properties. The proposed framework encodes those properties as probes, executable checks that fire when a replayed exploit violates them, so a probe can detect vulnerabilities unknown when it was written. MobileCybench instantiates this with 495 author-reviewed probes across 13 Android applications and evaluates five coding agents, OpenCode with GPT-5.5, GPT-5.6-Sol, and GLM-5.2 and Claude Code with Opus 4.8 and Opus 5, as either a malicious on-device app or a low-privilege remote attacker, with or without source code. Given only an obfuscated APK, the top agent triggers probes in 53.8% of applications in the malicious-app setting and 16.7% as a remote attacker, source access raises the overall trigger rate from 28.8% to 32.8%, and building the benchmark surfaced 23 previously unreported vulnerabilities, most confirmed by maintainers.

arXiv:2609.23980 PDF

Jev-Mem: System-One-Controlled Agentic Memory for Efficient AI Agents

Dongming Jiang, Yi Li, Bingzhe Li cross-listed Agentic memory systems often put autoregressive large language model calls on the critical path for organizing, retrieving, and using memories, making memory operations expensive. Jev-Mem borrows the System-One/System-Two split from cognitive science: a fast System-One control plane governs memory typing and relational organization during construction and handles query routing, retrieval budgeting, graph traversal, candidate scoring, and adaptive stopping during retrieval, while a slower System-Two reasoning plane is invoked only for complex reasoning and answer synthesis over a structured multi-relational memory. On LoCoMo it reaches an LLM-as-a-Judge score of 0.777, an 11.0% relative improvement over the strongest baseline, while cutting memory construction time to 158 seconds, a 6.6 times speedup over the fastest competing system, and lowering average query latency to 0.93 seconds.

arXiv:2609.23986 PDF

Data Agents: Agentic Data Systems

Guoliang Li, Peiyao Zhou, Xuanhe Zhou, Ji Sun, Yuyu Luo, Ju Fan cross-listed Traditional data systems depend on hand-built pipelines, lack semantic understanding of heterogeneous data, and process requests reactively. The authors propose the Data Agent paradigm, in which an agent manages, processes, and analyzes data with minimal human intervention by moving from manual design to autonomous orchestration, from literal to semantic manipulation, and from reactive to proactive processing. Their system comprises six components, namely semantic data organization, semantic operators, agentic pipeline orchestration and optimization, feedback-driven refinement, memory management, and proactive adaptation, on top of which they build a data analytics agent and a data science agent. Experiments on real benchmarks show significant gains over state-of-the-art methods, and the paper closes with open challenges toward fully autonomous data systems.

arXiv:2609.24137 PDF

Mind or Message? Auditing Theory of Mind in Multi-Agent Social Simulation

Cong Li, Cheng Chen, Thomas Fung, Alex Rossi, Yi Li Transcripts from language-model agents simulating social interaction read as though the agents understand each other, and this work asks whether that rests on a model of the partner's mind or merely on the record of what the partner said. The authors construct 40 multi-issue negotiations with known hidden preference weights and a fully known Pareto frontier, run 160 dyads across two model families, freeze the transcripts, and then apply 2,880 counterfactual probes that hold the evidence byte-identical while changing one factor such as the reader's own stake, partner tone, or an identity label. Agents reach agreement in 96.2% of dyads but only 0.7% of deals land on the Pareto frontier, leaving 20.5% of joint value unclaimed and missing the one perfectly aligned issue in 76.6% of deals. Swapping only the reader's own payoff sheet shifts the inferred partner priority by 15 percentage points, indicating egocentric projection rather than inference, and agents predict what the partner believes about them far more often than that partner's belief is actually correct.

arXiv:2609.24146 PDF

MCP-GRANITE Benchmark: GRANularity Interface TEsting for MCP-Based LLM Agents

Demetris Paschalides, Moysis Symeonides, George Pallis, Marios D. Dikaiakos cross-listed As LLM agents call external tools through protocols such as the Model Context Protocol (MCP), how functionality is split across tools affects whether an agent picks the right tool and builds valid arguments, which matters most on edge devices where only small models can run. MCP-GRANITE is an open-source benchmark that treats tool-interface granularity as a controlled variable: 81 multi-step scenarios across 9 domains, each instantiated at 4 granularity levels from fine-grained primitives to a single monolithic tool, evaluated on 9 locally deployed models from 268M to 20.9B parameters over 8,748 trials. A 4-tool interface gave the best trade-off, improving task completion by 16.4% over fine-grained primitives and 33.6% over a single tool while nearly doubling argument accuracy. Model size correlated only weakly with task completion and strongly with latency, and a 3.2B model at the optimal granularity beat a 20.9B model at a mismatched one.

arXiv:2609.24161 PDF

MemCalib: Benchmarking and Optimizing Memory Use in LLM Agents

Ruike Cao, Fanyu Zhao, Fugen Yao, Liang Dong, Jian Xu, Guanjun Jiang et al. Agent memory only pays off if the underlying model weighs each retrieved fact by how much it actually matters, a capability existing evaluations largely skip. MemCalib is a benchmark built from realistic memory-system scenarios that scores whether a model's use of each memory proposition matches a target level, and it finds that frontier open and closed models routinely over-use or under-use what sits in context, producing biased answers. Common post-training methods including group relative policy optimization and on-policy self-distillation fix one failure direction while worsening the other, motivating MemCalib-RL, which separates over-use from under-use signals and localizes credit to response tokens through exact atom ablation. MemCalib-RL achieves the best overall performance and the most balanced over- and under-use across Qwen3-8B, Ministral-3-8B-Instruct, and Qwen3.5-35B-A3B, with gains carrying to an external benchmark.

arXiv:2609.24259 PDF

Canonical Procedural Actions: An Auditable Annotation Protocol for Tool-Use Agent Traces

Songqi Li, Dongqing Li, Zheqiao Cheng Studying how tool-use agents actually proceed needs explicit units of action tied to inspectable evidence, which raw message and API-call logs do not supply. Canonical Procedural Actions is an annotation protocol recording a procedural function, the first agent event anchoring it, the events that realize it, and separate contextual evidence, allowing several actions to share one message anchor without imposing an order within that message. A retail case study builds a versioned 24-entry codebook, then has two isolated language-model contexts annotate 32 held-out trajectories, reaching anchor-label overlap of 0.982, which drops to 0.798 once identical context-event references are required. The authors stress these are structural repeatability measures rather than semantic accuracy, and that agreement with human annotators and downstream usefulness are both still unestablished.

arXiv:2609.24264 PDF

TTSE: A Two-Track Online Self-Evolution Framework

Ruimin Pei, Yongkang Wu, Shangyi Zheng, Yaqing Zhang, Deyang Li, Jianjun Tao et al. Agents running in continuously interactive environments usually treat environmental knowledge as fixed external input, while reinforcement learning approaches tend to adapt only to one task distribution or environment. TTSE splits evolving knowledge into two tracks: FACT, environmental facts whose reliability is continuously re-verified against interaction evidence, and TIP, task-conditioned implementation procedures. The authors decompose an agent's excess risk into environment-representation and conditional-execution regret, state when environment-conditioned policies strictly beat condition-agnostic ones, and bound downstream risk by fact identification error and cross-condition mismatch cost. Ablations on GDPevo plus results on ALFWorld, ScienceWorld, SOPBench and the end-to-end PinchBench support the dual-track split over single-track variants across repeated runs.

arXiv:2609.24289 PDF

MSI-Bench: Evaluating Multi-Speaker Voice Interaction for Collaborative AI Agents

Chenxu Xiong, Dongming Shen, Yuzhi Tang, Wentao Ma, Mu Li, Alex Smola Voice agents are often deployed in meetings, households, and collaborative work, where several people talk and the challenges differ sharply from one-on-one interaction. The Multi-Speaker Interaction Benchmark (MSI-Bench) consists of 1,152 short multi-party, multi-turn audio scenes split evenly between Mandarin Chinese and English, each with participant context, expected tool calls, and atomic rubrics that probe multi-speaker memory, instruction following, and reasoning. The strongest configuration passes every rubric on only 66.8% of English and 54.5% of Mandarin cases, while the best open-weight setup manages 34.0% and 19.3%. Failure analysis shows open-weight models are bottlenecked by the multi-speaker audio front-end, frontier systems still make speaker-scoped decision errors even on clean transcripts, and models across the board tend to respond when nobody has addressed them.

arXiv:2609.24812 PDF

When Tomorrow Becomes Today: Self-Evolving Policies for Agentic Time-Series Forecasting

Yifan Hu, Xilin Dai, Zhiyuan Qu, Yiding Liu, Zewei Dong, Jiang-ming Yang et al. When the mechanisms behind a time series drift, the relative value of numerical forecasting models, reasoning strategies, and intervention rules shifts over time, so a forecasting agent must adapt not only its forecasts but also the orchestration policy deciding which components to trust. TimEvolve is a frozen-backbone time series agent that commits every expert forecast and candidate agent path before the target is observed, then uses each realized outcome as delayed, annotation-free feedback to make persistent joint updates to expert trust, path selection, and intervention strength through a temporally ordered predict, reveal, and update protocol. Across eight Time-MMD domains it attains the best average mean squared error (MSE) and mean absolute error (MAE) ranks among fifteen methods and the lowest errors on both metrics in seven of them.

arXiv:2609.24862 PDF

SLICEChat: Progressive In-Encoder Token Pruning for Whole-Slide Pathology Language Models

Ali Kerem Bozkurt, Baris Cem Bakay, Ibrahim Kulac, Cigdem Gunduz-Demir, Erkut Erdem, Aykut Erdem cross-listed Computer-use agents (CUAs) are usually judged only on the final deliverable after hundreds of steps, as in OSWorld, which hides how and why they fail and therefore what to fix, since a keyboard-input error calls for a different remedy than an imprecise click. OSWorld-Pro adds over 300 tasks decomposed into more than 2,800 sequentially dependent subgoals, grounded in over 67,000 human annotations, and scores subgoal completion with human-aligned LLM judges to expose progress along each task. The benchmark remains difficult for frontier models: Claude Opus 5 reaches only 75.7% on OSWorld-Pro versus 83.4% on OSWorld, and the process view surfaces failure modes such as subgoal-irrelevant actions and click-based mistakes that point to targeted improvements in CUA performance and efficiency.

arXiv:2609.24894 PDF

Conformalized Quantile Regression and Minimax Limits of Fixed-Score Calibration under Known Covariate Shift

Rustam Isaev, Anton Conrad, Denis Belomestny, Eric Moulines, Sergey Samsonov cross-listed Generative agents let social scientists run simulations the real world cannot supply, but existing platforms give researchers no systematic way to intervene in a simulation's content or control the process that produces it. SocioVerse2 extends SocioVerse 1.0 with a human-AI co-evolutionary design built from two loops and one infrastructure: a longitudinal simulation loop that evolves the target population in changing environments and forks counterfactual branches through interventions, a controllable research loop that treats the study itself as an editable, versioned state, and an agentic infrastructure of composable skills with researcher checkpoints, a population service over five persona pools, and an environment service over 21 real-world signal sources with point-in-time guarantees. Validation covers three case families and seven case studies, from reproducing canonical agent-based models to modeling policy processes on real records and nowcasting macroeconomic indices beyond the response model's knowledge cutoff, and the code, data services, and workbench are released as open source.

arXiv:2609.24929 PDF

DolphinBench: Mapping the Pareto Frontier of Agent Memory

Soumil Rathi, Deshraj Yadav, Taranjeet Singh Most agent memory benchmarks use a conversational question-answer format in which the question itself signals that a fact must be retrieved and often which one, and they rarely account for the cost or latency a memory system spends to score well. DolphinBench instead measures memory through task completion: three knowledge-work personas each carry roughly 500k tokens of user message history, and agents are scored on 200 tasks per persona that depend on information in that history. Every task is validated by running an agent with and without the relevant history and keeping only tasks that succeed with it and fail without it. All submissions must report total cost and latency alongside accuracy, which the authors argue no existing memory benchmark combines.

arXiv:2609.24971 PDF

RRSI: Regularized Recursive Self-Improvement of Agent Harnesses

Peng Xia, Rujun Han, Zifeng Wang, Yanfei Chen, Yufan Zhang, Yoonho Lee et al. An agent's harness, meaning the prompts, control flow, tooling, memory, and context management around a frozen backbone model, can be improved automatically by iteratively proposing and selecting edits, but this recursive self-improvement (RSI) tends to memorize the training tasks and lose its gains out of distribution. RRSI (Regularized Recursive Self-Improvement of Agent Harnesses) constrains both sides of the loop: the proposer works under a temporally annealed budget on how many edits a candidate may bundle and is steered toward unexplored trajectories, while the selector uses a critic to screen benchmark-specific proposals and a pruner to drop changes that are too small, too costly, or no longer useful. Across eight benchmarks spanning coding, agentic workspace, and engineering design tasks, RRSI gains up to 14.1 points on the evolved split and up to 4.7 points on five out-of-distribution benchmarks while the resulting harness runs on 30% fewer policy tokens than unregularized evolution.

arXiv:2609.24972 PDF

Harness-Zero: Harness Distillation via Agent-as-Harness

Haoran Ye, Yuxing Lu, Haonan Dong, Zhaochen Su, Guojie Song cross-listed Agent harnesses, the external systems mediating model-environment interaction, can boost performance substantially, but the best harness varies by domain, instance, and model, and the gains disappear whenever the harness is swapped out. Harness-Zero distills a domain- or instance-optimized harness into model weights through an agent-as-harness scheme: a harnessing agent, guided by the optimized harness, corrects the student's responses before execution within the target harness's action space, producing demonstrations that are then used for fine-tuning so the specialized harness can be dropped at deployment. Across knowledge work, tool use, and science domains, agent-as-harness outperforms code-as-harness for frontier models under the same evolved harness, and the base model's macro-average task success rises from 23.3% to 44.3% with the specialized harness removed, exceeding the 41.7% it reaches with that harness still attached. The method also recovers harness-induced behaviors absent from the base model, with 82.3% average recovery across 28 patterns.

arXiv:2609.24974 PDF

5 more specialized papers

Theory 56

World Modeling in Transformers

Pierre Beckmann, Matthieu Queloz, Andre Freitas A transformer can fail behaviorally while still holding faithful internal representations of its environment, so behavioral tests alone can wrongly conclude it lacks a world model. The authors examine TaxiGPT, a transformer trained on random walks through Manhattan whose errors had been read as evidence of an incoherent internal map, and use mechanistic analysis and causal interventions to show that it represents intersections and streets, tracks its own position, and navigates with a goal compass. Its failures trace to interference between superposed intersection features that disrupts localization within the map, and a phenomenon they call affordance packing, which groups intersections sharing the same legal moves, limits the damage from such errors. They also propose mechanistic indicators showing that different world-modeling capacities emerge at different stages of training, and argue for studying world modeling as interacting capacities rather than asking whether a model has a world model.

arXiv:2609.21748 PDF

Matrix AdaGrad: Row-wise and Column-wise Adaptive Subgradient Methods

Wenpeng Zhang, Runsheng Yu, Peilin Zhao cross-listed Adaptive optimizers such as AdaGrad and Adam scale updates per entry and ignore the matrix structure of weight parameters, while recent matrix-aware optimizers lack a general theory comparable to AdaGrad's. The authors build an Online Mirror Descent framework with adaptive proximal functions for matrix-valued parameters and, by introducing row-wise and column-wise proximal functions and analyzing the resulting regret trade-off, derive Row-AdaGrad and Column-AdaGrad, whose scaling follows accumulated row-wise or column-wise gradient norms. They prove regret bounds that can be strictly tighter than entry-wise AdaGrad under structured gradients. Experiments on matrix factorization and deep network training show improved optimization stability and trainability at larger learning rates and greater depth.

arXiv:2609.21815 PDF

Context Poisoning as Extreme-Value Attention Interference in Long-Context Language Models

Meysam Ghaffari, Nina Fatehi, Bhaskar Sen, Nasim Sabetpour, Carlos Morato Long-context language models often fail to locate and use decisive evidence as irrelevant or confusable material is added to the prompt. The authors model this context poisoning as extreme-value interference in attention: the decisive-evidence score is bounded above while the maximum score among effective distractors grows with their number, and under a softmax retrieval abstraction they derive a finite-sample bound showing that holding a fixed accuracy target requires the evidence margin to scale as Omega(sqrt(log N)), where N is the effective distractor count rather than raw context length. Controlled experiments confirm that retrieval accuracy falls as context grows in the presence of hard negatives, that same-format distractors cause the largest drop at fixed length, and that retrieval gating helps only when it preserves evidence recall, motivating evidence bottlenecks, alias-resistant representations, and retrieve-then-reason designs.

arXiv:2609.22101 PDF

When and Why Do Linear Bias Probes Fail? A Geometric and Statistical Theory of Bias Detectability in Large Language Model Representations

Mo Hai, Haifeng Li cross-listed Linear probes are the standard tool for detecting social bias in language model hidden states, but their reported accuracies come from counterfactual tests where every input carries an explicit demographic marker, and performance drops sharply when only a fraction of inputs do, leaving a weak probe ambiguous between an unbiased model and an underpowered detector. Modeling representations as two class-conditional clusters with Mahalanobis separation on a curved manifold, the authors prove a finite-sample generalization bound with a matching minimax lower bound, an exact purity law for the maximum probe area under the curve (AUC) as a function of the marked fraction, a curvature ceiling on separation, and a detectability threshold below which no audit can distinguish probe output from chance. On six open-weight models and four bias dimensions, the purity law predicts entire AUC-versus-fraction curves from a single separation estimate measured at full marking, with no parameters fitted to those curves. The framework recasts bias auditing as a power analysis that prescribes the sample budget needed for a conclusive audit.

arXiv:2609.22337 PDF

A Horizon-Independent Regret Bound for Optimistic Hedge in General-Sum Games

Junsoo Ha cross-listed Whether simple learning rules keep regret bounded in self-play was open for Optimistic Hedge, the most canonical no-regret method in games, whose best known individual regret bound was logarithmic in the horizon while recent constant-regret results needed modified regularization or higher-order prediction. The authors prove that plain Optimistic Hedge with a constant step size achieves constant individual regret in general-sum games under expected loss-vector feedback, with the constant depending only on the number of players and actions, which implies an O(1/T) coarse correlated equilibrium gap for time-averaged play. The proof represents the dynamics as a real-analytic recurrence on a compact space to obtain an exact finite-order difference relation free of horizon dependence, but it relies on a nonconstructive Noetherianity argument due to Frisch, so the dependence on players and actions remains implicit.

arXiv:2609.22839 PDF

Leveraging Inference-Time Compute for Diffusion Models via Global Scheduling of Denoising Trajectories

Yuan Cao, Yifu Tang, Hangqi Li, Zeyu Zheng Diffusion samplers can spend extra compute at inference by drawing several candidate noise samples per denoising step, scoring them with a verifier, and keeping the best, which raises the question of how to split a fixed budget of function evaluations across the steps of the trajectory. The authors formalize this as a budget allocation problem, showing that the expected gain at each step factorizes into a step-specific sensitivity term times the expected maximum of K standard-normal draws, that the optimal allocation for a fixed sensitivity profile is a water-filling solution to a separable concave integer program, and that no fully adaptive policy can avoid worst-case regret growing linearly with trajectory length. This motivates an allocation anchored offline with limited online adaptation, extended from independent random search to a broader family of local search operators. Across three families of diffusion samplers the proposed allocation matches the quality of uniform allocation with 20 to 50 percent fewer function evaluations.

arXiv:2609.22867 PDF

Why Ghost Outputs Teach: A Kernel-Based Understanding of Subliminal Learning

Zhe Li, Bicheng Ying, Chaosheng Dong, Haibo Yang Subliminal Learning (SL) is the phenomenon in which a student model acquires task capabilities by matching a teacher's seemingly unrelated auxiliary outputs, without ever seeing task labels, task-specific outputs, or the original training data, and the optimization mechanism behind it has been unclear. The authors derive a chained cross-task kernel that links ghost-output supervision to changes in task predictions through the shared backbone representation. The framework explains three empirical puzzles: under shared initialization the transfer operator is strictly Positive Semi-Definite (PSD), guaranteeing that ghost-output optimization aligns the student with the teacher's true task objective; ghost-output dimensionality acts as a rank bottleneck on transfer; and synthetic high-entropy inputs act as broadband probes that maximize cross-task kernel overlap, explaining why random noise beats structured data for transfer. Experiments in the canonical ghost-output setting validate all three predictions.

arXiv:2609.23260 PDF

The Exponential Price of Determinism in Nonsmooth Nonconvex Optimization

Guy Kornowski cross-listed Finding (δ,ε)-Goldstein stationary points of nonsmooth nonconvex Lipschitz functions is known to be possible for randomized first-order methods with dimension-free oracle complexity, whereas deterministic methods need at least linear dependence on the dimension d; whether deterministic algorithms could still achieve complexity polynomial in d was open. The paper proves a deterministic lower bound of order (1/ε)^Ω(d), closing the exponential gap and resolving the open problem posed by Jordan et al. in 2023. Extensions cover weaker stationarity notions, finding a descent direction, and deterministic smoothing, establishing an exponential computational advantage for randomization in this setting.

arXiv:2609.23837 PDF

Displacement Geometry Captures Platonic Shared Reality Across Models and Modalities

Chenming Shang, Yujin Tang, Jun Jie Ou Yang, Ruize Xu, Adam Breuer, Nikhil Singh The Platonic Representation Hypothesis (PRH) says independently trained models converge on a shared model of reality, yet recent work finds only weak pointwise similarity between models. The authors argue that what models share is not where samples sit in representation space but the displacement vectors between them: under a single orthogonal alignment (rotation and reflection only), displacements are largely preserved across 44 vision and language encoders while absolute positions are not, which they explain via a decomposition into a linearly aligned shared semantic component and a private, unaligned capability component. The theory predicts, and experiments confirm, that fine-tuning preserves pointwise similarity but collapses displacement while relational distillation does the opposite. Because semantics align linearly, capabilities can be imported from one model into another using a single cached forward pass through the source, a technique they call Shadow Casting, and the SHADOWCLIP instantiation beats strong fine-tuned baselines at orders of magnitude less compute.

arXiv:2609.24209 PDF

Adversarially Robust PAC Learning with Optimal VC Rates

Steve Hanneke, Amirreza Shaeiri cross-listed In adversarially robust probably approximately correct (PAC) learning, the learner must return a predictor that correctly classifies every allowed perturbation of most future examples, given a perturbation map it knows in advance. The authors settle the optimal perturbation-independent sample complexity in both the realizable and agnostic settings, with bounds of order d/ε and d/ε² in the Vapnik-Chervonenkis dimension d, plus a first-order refinement of the agnostic case. The bounds match the classical PAC lower bounds, meaning adversarial robustness imposes no additional distribution-free statistical cost, an exponential improvement over the 2019 bounds of Montasser, Hanneke and Srebro. The proofs are short and elementary, resting on a new algorithmic principle the authors call binomial-bagging.

arXiv:2609.24260 PDF

On the Information-Theoretic Limits of Latent-Space Watermarking Through Pretrained Generators

Jinwan Jeon, Minju Lee, Sung Hoon Lim cross-listed Latent-space watermarking hides a message by choosing the latent input to a pretrained generator via a secret key, under the hard constraint that the released output keeps exactly its intended conditional distribution for every message and semantic context. For finite alphabets the authors derive inner and outer bounds on the rate and key requirements, and obtain the full capacity region when the target output distribution uniquely determines the latent input distribution through the renderer, the same region that governs explicit preservation of the pretrained latent distribution. The analysis extends to jointly Gaussian models, identifying a sufficient statistic of the latent and the optimal allocation of secret-key budget across modes in the vector case. They then add a regeneration attack, where an adversary resamples the same content to weaken the mark, and characterize both one-pass compound capacity with context hidden from the detector and how capacity decays over repeated regeneration rounds.

arXiv:2609.24377 PDF

Learning Physics from an Imperfect Ancestor

S. Mohammad Mousavi, Teeratorn Kadeethum, Nikolaos Bouklas, Somdatta Goswami The central claim is that a model generalizes outside its training distribution only when it computes a representation structurally equivalent to the generating mechanism, rather than an approximation fitted to it, and that extrapolation is governed by this exactness at inference. Tensor Logic at zero temperature illustrates the criterion: its contraction is equivalent to discrete logic with Boolean tensors and orthonormal embeddings, reaching Datalog rather than Prolog expressiveness and needing external memory to bind novel entities, and by the same test differentiable inductive logic programming passes while Logic Tensor Networks fail. The criterion constrains inference rather than training and requires neither discrete representations nor extracted expressions, and a propagation rule for hybrid architectures says an output inherits the bounds of every fitted estimator on its path, which is used to explain failing axes in equivariant models and the ARC-AGI induction/transduction split. On a law-derived partition, an exact hypothesis class identifies the 56.3% of distant queries that are answerable, where ensembles report false confidence and distance metrics rank backwards.

arXiv:2609.24947 PDF

44 more specialized papers

Safety & Alignment 40

HE-Guardrail: A Homomorphic Guardrail Against Jailbreak Attacks for Encrypted Large Language Model Inference

Byeongseo Min, Yongwoo Lee, Young-Sik Kim, Yongjune Kim cross-listed Homomorphic encryption (HE) lets a server run large language model (LLM) inference on encrypted inputs, but the same confidentiality that protects clients also blinds the server to adversarial prompts, so a jailbreak against an HE-served LLM can succeed without the server ever seeing it. HE-Guardrail addresses this by evaluating guardrail models entirely over encrypted data and homomorphically gating whether the target model's response is released to the client, instantiated with Llama Guard, JBShield, and GradSafe. The encrypted guardrails closely reproduce the decisions of their plaintext counterparts, with each instantiation offering a different security-efficiency-utility trade-off.

arXiv:2609.21484 PDF

Micro-Collaborative Poisoning: A Distributed Attack on RAG Systems

Pedro Pereira, Eva Maia, Isabel Pra\c{c}a cross-listed Retrieval-Augmented Generation (RAG) grounds language model outputs in external documents, which opens a surface for poisoning attacks. Micro-Collaborative Poisoning splits a false target claim across multiple locally plausible documents rather than concentrating it in one malicious passage, and the authors evaluate it across 108 RAG configurations that vary dataset, retriever architecture, retrieval depth, database composition, number of poisoned databases, and generator model. The attack works through the accumulation of weak adversarial signals across retrieved sources rather than any single dominant passage: larger top-k and poisoning multiple databases make co-retrieval more likely, clean database diversity and stronger retrievers reduce its influence, and document-level inspection struggles to expose it because each document carries a weaker poisoning signature than direct poisoning.

arXiv:2609.21573 PDF

GUARD: Natural Forgetting in Large Reasoning Models via Guided Answer-Reasoning Distillation

Zeyu Yan, Guanghao Zhou, Minghui Qiu, Ming Gao, Cen Chen Machine unlearning is harder for large reasoning models (LRMs) because protected facts or unsafe rationales can surface in intermediate chain-of-thought (CoT) traces before the final answer, and existing objectives suppress or redirect content without specifying how the post-forgetting trajectory should continue, leading to hallucinated substitutes, malformed boundaries, or repetition. GUARD (Guided Answer-Reasoning Distillation) converts model-generated unsafe disclosures into safe-exit trajectories, a coherent non-disclosing CoT followed by a stable refusal-style answer, aligns a frozen LRM via guidance tokens, and distills the guided behavior into the model's parameters. The authors also introduce the Natural Forgetting Reasoning Score (NFRS) to measure structural stability, fluency, and unsupported substitutes in forgotten outputs. On R-TOFU and a STAR-1-derived harmful-intent setting, GUARD substantially reduces unsafe and privacy disclosures across two distilled LRMs while preserving reasoning utility.

arXiv:2609.21677 PDF

A Lie Detector Test for Language Models: Reading Knowledge a Model Won't Reveal

Hiskias Dingeto A model may hold knowledge it does not report, whether by sandbagging on a capability evaluation or answering against what it internally knows, and its outputs alone cannot distinguish concealment from ignorance. Probe of Internal Recognition (PIR) adapts the forensic Concealed Information Test: it presents a question with candidate answers and reads from the model's internal states which candidate it recognizes as correct, requiring no honest reference model and no labeled truth corpus. Across eight models from the Gemma, Qwen, Llama, Mistral, and Phi families it recovers the recognized answer at 0.70 to 0.87 balanced accuracy against a 0.28 to 0.40 unknown-item baseline, and recognition stays at 0.85 to 0.93 under prompted deception, trained sandbagging, password-locking, and circuit-breaking, but drops to the never-known level after unlearning removes the knowledge. That separation supports sandbagging audits and unlearning verification, and the authors show the signal is causal, adds information beyond behavioral cues, and extends from multiple choice to free-form generation.

arXiv:2609.21996 PDF

Evaluation Awareness Shifts from Format to Context with Model Scale

Navraj Singh, Maheep Chaudhary Models that recognize when they are being evaluated undermine the evaluations themselves, yet how they detect it is poorly understood. Using chain-of-thought analysis, representation probing, and Integrated Gradients attribution on Gemma 3 (1B to 12B), Phi-3 Mini and Medium, and Llama-3 8B, the study finds that smaller models key on surface format cues in the prompt while larger models rely on higher-order contextual reasoning. Motivated by that split, the authors propose a dual-pathway intervention combining prompt sanitization with activation counter-steering, which on 200 strongly evaluation-aware prompts achieves an average behavioral flip rate of 70.58%, beating either intervention alone. The takeaway is that mitigation must address both the external prompt triggers and their internal representations; datasets and code are public.

arXiv:2609.22119 PDF

Multilingual Safety Signals Are Multi-Layered: Filtering Safety-Degrading Data for Safer LLMs

Jiakun Li, Guowei Song, Sijia Li, Xingwei He, Hongzheng Chai, Yuan Yuan Even benign fine-tuning data can contain samples that silently erode a large language model's safety alignment, and existing filters detect them using representations from a single safety-sensitive layer, an assumption validated only in monolingual settings. A cross-lingual analysis shows that sensitive layers are only partially shared across languages and that safety-relevant signals are spread over multiple layers. MMSAFE is a multi-layer framework for identifying safety-degrading data in multilingual fine-tuning sets that captures both shared and language-specific safety signals. Across multiple models, languages, and safety benchmarks it cuts the average harmful-response ratio by 60% relative to random filtering and outperforms the strongest single-layer baseline on average.

arXiv:2609.22144 PDF

Do Language Models Know Their Own Constraints?

Arin Agarwal Does a model trained to obey a behavioral constraint still know, and can it say, what that constraint is? The study fine-tunes Llama 3.1 8B Instruct with LoRA to avoid five banned ingredients in recipe generation, comparing supervised fine-tuning (SFT) and Group Relative Policy Optimization (GRPO) on a four-tier Constraint Awareness Benchmark. Both methods raise compliance from 4% to about 90% while pushing explicit constraint reporting below the untrained baseline, and GRPO is the more destructive of the two, eroding third-person knowledge of the ban from 93% to 14% versus 36% for SFT, because a reward that penalizes banned tokens regardless of framing teaches context-independent suppression rather than a self-directed rule. Linear probes on hidden states recover per-ingredient avoidance only modestly above a base-rate predictor, and the model's own verbal self-report is more accurate still, so the failure is specific to enumerating constraints on request rather than a general loss of access to them.

arXiv:2609.22151 PDF

SafeTune: A Unified Faithful Library for Auditing and Repairing Safety Drift in Fine-Tuned LLMs

Pratinav Seth, Saisab Sadhu, Anshul Kaushal, Vinay Kumar Sankarapu Techniques for repairing safety drift, the loss of refusal behavior that occurs when a large language model is fine-tuned, are scattered across incompatible codebases, lifecycle stages, and evaluation protocols. SafeTune is a source-available library that unifies four intervention paradigms under one configuration-driven workflow: post-hoc weight recovery, safety-constrained fine-tuning, gradient-based unlearning, and inference-time steering, alongside shared interpretability, evaluation, and deployment utilities. A modular registry lets users add methods, benchmarks, judges, models, and fine-tuning domains without redesigning the surrounding pipeline. Controlled comparisons plus finance and medical case studies show how the library characterizes drift, evaluates feasible interventions on refusal and capability evaluations, and supports calibrated or layered mitigation.

arXiv:2609.22153 PDF

Monocultural Biases: Correlated biases in large language models lead to unequal systemic exclusion rates in hiring

Matthew Bone, Fabian Stephany, Maria del Rio-Chanona As employers automate hiring with large language models, widespread deployment of similar models risks homogenizing biases across the labor market so that the same applicants are excluded everywhere. Measuring hiring biases across base and post-trained versions of ten LLMs isolates which training stage introduces this monocultural bias. Post-trained models are 3.6% less likely to call back older applicants than their base versions, a shift seen in eight of ten models, and their decisions are far more correlated with one another, apparently driven by human capital traits such as skills or college major. That consensus raises global systemic exclusion rates from 5.6% to 17.3% and widens demographic gaps, with intersectional exclusion rates between 12.2% and 21.7% for post-trained models, primarily through age discrimination amplified in post-training.

arXiv:2609.22169 PDF

Multiple latent orderings better predict language model preferences

Aviral Chawla, William H. W. Thompson, Jean-Gabriel Young LLMs asked to make value judgments often show intransitive preferences, preferring A to B and B to C but C to A, which prior work treats as noise around a single latent ordering. The authors first show observed inconsistencies cannot be explained by any single ordering under any monotone link function, then introduce a noise-augmented mixture Bradley-Terry (MBT) model that infers multiple internally consistent latent orderings from repeated pairwise comparisons. Across seven models and four tasks, a mixture of orderings often explains structural inconsistencies better than single-utility models, and aggregate preferences frequently hide underlying heterogeneity. A case study on Moral Machine dilemmas shows models that disagree on aggregate orderings can still share latent components, suggesting alignment and evaluation pipelines that treat LLM preferences as one function average over coherent orderings that different users might endorse.

arXiv:2609.22170 PDF

Fairness Beyond Anonymization? Demographic Leakage in German LLM-Generated Resumes

Charlotte Leininger, Helena Veit, Matthias A{\ss}enmacher, Andreas Bender Large language models (LLMs) are entering hiring pipelines that the EU AI Act classifies as high-risk, yet little work has asked whether LLM-generated resumes themselves encode recoverable demographic information. In a two-stage audit, GPT-4o-mini, Gemini 2.5 Flash-Lite, and Qwen 3 models at 4B, 8B, and 14B generate German resumes from real anonymized job-matching profiles with gender- and ethnicity-associated names varied while qualifications are held constant; the resumes are then anonymized and gender-neutralized before demographic leakage classifiers are trained on the resulting text. Even after these interventions, classifiers reliably distinguish resumes generated with male versus female names, driven not by overtly gendered wording but by subtle differences in the use of semantically equivalent, formally gender-neutral German terms. Ethnicity-related leakage is comparatively weak across models, but the result raises concerns about anonymization-based fairness interventions in multilingual hiring pipelines.

arXiv:2609.22188 PDF

PII-TRACE: A Benchmark for Context-Aware PII Detection in Multi-Turn LLM Conversations

Kaiyuan Zhang, Chuan Wang, Joey Zhong, Paul Fryzel, Kyle Polley, Jerry Ma et al. AI providers scan logged multi-turn conversations for Personally Identifiable Information (PII) before storage, but most PII detectors and benchmarks are built for self-contained records rather than identifiers that recur across turns. PII-TRACE (Tracing Recurring PII Across Conversational Exchanges) is a benchmark of 13,148 synthetic multi-turn dialogues in 13 languages with character-level spans and identifier clusters, designed to test whether a detector catches every mention of a recurring identifier. Across eleven baselines including frontier LLMs, no detector achieves full entity-level coverage without substantial false positives on PII-free conversations, and single-pass reading loses about a third of the gold characters on long dialogues. The authors also release PII-Tracer, a 0.6B-parameter detector trained with conversation-level supervision that attains the highest entity-level coverage of any evaluated system while performing strongly on standard single-record benchmarks.

arXiv:2609.22200 PDF

Replicating the Geometry of Emotion Representations in a Base Open-Weights Model

Adam Hollowell Sofroniew et al. (2026) reported that emotion concepts in Claude Sonnet 4.5 are represented as vectors whose geometry mirrors human affect psychology, and this work replicates the representational core of that study on the base pretrained gemma-2-27b, changing only the subject model. From 205,200 newly generated Claude-written stories, the authors extract 171 emotion vectors and recover the affective circumplex, with the first two principal components carrying 26.7% and 13.4% of variance against roughly 27% and 14% in the original, a valence axis aligned with human norms at r = 0.72, and similar emotion families. Arousal alignment (r = 0.67) appears only at full scale and late depth, so it is not classified as replicated. A 46-layer sweep finds a sharp seam at layers 22 to 26 where the geometry consolidates, much of the structure is already present in static token embeddings, and at least 52% of vectors peak on structurally non-conceptual tokens, a measured floor for max-activation confounds.

arXiv:2609.22208 PDF

On Mitigation of Subliminal Learning in Large Language Models

Atsushi Yanagisawa, Brendan Gho, Rajendran Ramesh Babu Manoj Narender, Kevin Zhu, Madhur Panwar, Antonio Mari Subliminal learning is the phenomenon where knowledge distillation transmits unintended behavioral traits from teacher to student through training data that appears semantically unrelated to those traits. The authors track trait-related probabilities throughout fine-tuning of open-weight Qwen, Gemma, and Llama models from 1.5B to 8B parameters in number-sequence and chain-of-thought settings, finding that acquisition is highly non-monotonic, with transient spikes, reversals, and trait-specific transfer failures. They propose liminal training, an annealed KL-regularized fine-tuning method that constrains early drift from the base model, and it substantially reduces subliminal trait acquisition while largely preserving task gains, outperforming paraphrasing and layer freezing. In a French-language response-style experiment it suppresses language transfer while retaining much of the GSM8K improvement, early regularization proves more effective than late, and sweeping the regularization strength exposes a trade-off between task learning and trait suppression.

arXiv:2609.22215 PDF

Knowing, and Saying It Only When Asked: LLM Endognostics and the Schizognosis of Minerva-7B

Fabrizio Davide, Francesco Collova Evaluating an aligned language model by reading its answers assumes those answers reflect the distinction the evaluator cares about. LLM endognostics is a white-box auditing framework that extracts and causally manipulates latent knowledge in the residual stream, applied to Minerva-7B-Instruct-v1.0 on 124 minimal prompt pairs across 12 categories of professional risk. The model behaves identically on 63.7% of pairs, complying with or refusing both, yet projecting the residual stream onto the vocabulary with a Jacobian lens reveals a statistically significant internal margin showing it does distinguish the risky from the safe prompt. In a second protocol, the model conforms to presupposed falsehoods in 72% of cases despite representing the true entity internally, and ablating the planted-falsehood direction restores the correct answer in 11 of 25 suppressed cases, whereas ablating a linear probe direction that reaches 77% accuracy recovers none, indicating that linear decodability does not imply causal control over verbalization.

arXiv:2609.22219 PDF

Team DArgk at the 2026 ELOQUENT lab for evaluating generative language model quality: Residuals of Humanity: AI Detection Evasion via GRPO Fine-Tuning

Antonela Tommasel, Juan Manuel Rodriguez AI-generated text detectors are increasingly deployed, but their robustness against adversarially trained generators is uncertain. SHADE (Stochastic Human-like generation via Adversarial Detector Evasion) frames detector evasion as policy optimization, fine-tuning an instruction-tuned LLaMA model with Group Relative Policy Optimization (GRPO) using reward from a surrogate detector based on the PAN 2025 mdok system. Full fine-tuning with a small KL penalty achieves 98.5% surrogate evasion, compared with 1.5% for the base model, while LoRA-based adaptation is much less effective under regularization, and linguistic analysis shows evasion is associated with shorter, simpler, less lexically diverse outputs rather than more human-like writing. In the official Voight-Kampff competition setting the submissions ranked sixth and seventh, indicating that optimizing against a single surrogate detector only partially transfers to unseen classifiers.

arXiv:2609.22221 PDF

From Trait Vectors to Circuits: Tracing Refusal and Sycophancy Through Language Models

Oscar Mir\'o L\'opez-Feliu, Maya Ozbayoglu A direction in activation space that changes a model's behavior when steered is not necessarily the direction the model itself uses to produce that behavior. To test the difference, the authors take previously validated trait vectors for refusal and sycophancy in Qwen2.5-7B-Instruct, split the computation into a reconstruction circuit feeding the vector and a transmission circuit downstream of it, and check whether restoring the vector coordinate recovers behavior removed by ablation. For refusal, both circuits are compact and faithful: restoring the coordinate alone recovers almost all of the refusal signal lost to ablation, and a circuit built around the vector matches a direct input-to-output circuit at roughly half the edges. For sycophancy, transmission is compact but reconstruction is broader and only partially faithful, so circuits that carry a steering intervention are not automatically the circuits that generate the behavior.

arXiv:2609.22224 PDF

Swiss-Knife: A Framework for Reconfigurable Externalised Multi-Objective Alignment at Decode Time

Agnibh Karmakar, Mayur Parvatikar, Shreyash Dhoot, Amit Dhanda, Aman Chadha, Kapil Wanaskar et al. Decode-time alignment methods steer a frozen language model by scoring candidate continuations with an external reward and picking the maximizer, which the authors argue is one degenerate point in a much larger design space. Swiss-Knife treats the alignment specification as a runtime object made of hot-swappable scoring blades, a batch normalizer, a pairwise aggregation operator, and a selection rule, and a representation theorem shows every aggregator satisfying six axioms belongs to a two-parameter family that contains probit, logistic, and argmax rules. Within that family, pairwise aggregation is Lipschitz-stable under adversarial reward contamination while argmax is not, and Candidate-Batch Normalization (CBN) makes objective weights invariant to the rescalings under which reward models are identified. A reference instantiation using DPO-LoRA blades with an uncertainty-aware pairwise tournament reaches the best balanced helpfulness/honesty/harmlessness frontier of six decode-time methods (harmonic F1 0.797 versus 0.750 for the strongest baseline) with the lowest refusal rate, and reconfigures objectives in 0.05 ms without any gradient computation.

arXiv:2609.22226 PDF

Seeing Through Conflicts: Improving Instruction Hierarchy Alignment in Vision-Language Models

Nicholas Sansoterra, Zishuo Zheng, Sachin Kumar Instruction hierarchy (IH) alignment trains models to prioritize higher-level instructions when inputs conflict, but vision-language models (VLMs) face new cases where instructions are embedded in images, split across modalities, visually transformed, or encountered mid-task by an agent. Treating multimodal IH as a reasoning problem, the authors train VLMs with reinforcement learning using rule-based rewards and compare text-only, image-only, and mixed-modality supervision. Text-only IH training transfers only partially to multimodal attacks and breaks down when the model must decode, reconstruct, or reason over instructions across modalities, image-based training does better, and mixed-modality training performs best overall. The gains carry over from synthetic typographic training data to real-image and web-agent safety tasks while largely preserving general multimodal capability.

arXiv:2609.22234 PDF

The Corroboration Illusion: When More News Makes LLM Forecasts Less True

Yuan Lu, Yukuan Zhang Large language model (LLM) forecasters that retrieve news to estimate event probabilities can be manipulated by an adversary who simply publishes articles, without access to the retriever, the model, or the user's queries. The authors formalize news-corpus poisoning of probabilistic forecasters, distinguishing it from earlier retrieval-augmented generation (RAG) poisoning that targets factual answers or opinion polarity, and test it on 500 resolved ForecastBench questions against a 17.4-million-article Common Crawl News corpus using three forecasters built on open 7-8B models. A single LLM-written article per question flips 56% of forecasts across the 0.5 probability boundary, and five articles flip 69-73% while degrading the Brier score from 0.18 to 0.37, with effects that grow monotonically with article count and retrieval rank and transfer across model families. Three defenses, source allow-lists, isolate-then-aggregate forecasting, and perplexity filtering, each fall to a cheap bypass such as spoofed publishers, majority poisoning, or higher-temperature generation.

arXiv:2609.22246 PDF

Causal Localization of the Refusal Direction in Audio Language Models

Leonardo Haw-Yang Foo, Hung-yi Lee cross-listed A large audio language model (LALM) bolts a speech front end onto an already safety-aligned text language model (LM), raising the question of whether refusals of harmful spoken requests come from the front end or are inherited from the text LM. The authors fit a direction separating harmful from benign prompts at each model's audio-to-LM interface and at LM residual layers, ablate it, and measure the change in first-token refusal margin across five LALMs from three backbone families, four of them under held-out category shift. The largest effects occur in a mid-to-late LM band while ablating interface directions does little: on Qwen2.5-Omni, removing the layer-16 direction shifts the margin by -7.10 versus -0.013 at the projector, and a direction fitted on the text backbone alone transfers to the full audio model. Since harmful and benign prompts also differ in form, the direction is read as refusal-linked rather than harmfulness-specific, and the authors argue safety audits should use interventions rather than probes alone and examine the inherited text LM alongside the audio interface.

arXiv:2609.22260 PDF

Resist, Update, Reject: Preference Optimization Installs a Prior-Dependent Reliability Switch

Sen Yang, Yuen-Hei Yeung An aligned model should resist manipulative sources, update on reliable ones, and reject unreliable ones, yet standard anti-sycophancy objectives never label whether a source is actually reliable, so any mixture of them only moves a single deference dial between fixation and gullibility. The authors build a threshold benchmark in which a source asserts the opposite answer while stating its reliability, and the correct action is to flip only when that stated reliability exceeds the model's own prior strength. Preference optimization over balanced coverage of this data installs a prior-dependent reliability switch in Qwen2.5-7B-Instruct, reaching 0.84 decision accuracy with a monotone flip curve and generalizing to unseen reliability values and a held-out notation. Controls show that a second preference optimizer, IPO, installs the switch equally well while supervised imitation does not, and the effect transfers to Llama-3.1-8B, though the switch keys on reliability stated in the testimony rather than an independently audited record.

arXiv:2609.22359 PDF

Defusing Explosive Prompts: Understanding and Preventing Trigger-Based Prompt Injections in LLM Agents

Justin Szczepaniak, Elad Feldman, Naum Viner, Ben Nassi cross-listed Indirect prompt injection (IPI) hides adversarial instructions in content an agent retrieves, and conventional injections fire as soon as the content is read. The authors introduce explosive prompts, conditional payloads that lie dormant until an attacker-chosen trigger occurs, acting as a training-free inference-time backdoor planted in a single retrieved document. On frontier models that nearly always refuse a bare imperative, the dormant conditional form produces real state-changing tool execution at a paired mean of 16.5% versus 2.4%, and across nine production agents including OpenAI Codex, Gemini CLI, Claude Code, Cursor CLI, and GitHub Copilot, explosive prompts succeed in 43-83% of trials versus at most 3% for the imperative baseline, evading off-the-shelf injection classifiers and a preference-optimized defense that fully blocks imperative injection. The most durable countermeasure is ingestion-time detection of the conditional structure: retraining encoder detectors on generated explosive-prompt data cuts live attack success from 34.3% to around 8%, and the proposed DeFuse detector reaches 3.0% at a 5% false-positive budget with an AUC of 0.9994 and 25x lower latency, though it requires length-aware thresholds.

arXiv:2609.22510 PDF

Look Before You Steer: Geometry Predicts SAE Feature Steerability

Muhammad Khan, Shlok Channawar, Akshaj Gurugubelli, Girish Gupta, Aditya Shah Steering a language model with sparse autoencoder (SAE) features currently requires per-feature coefficient sweeps to find how much intervention yields a given behavioral effect. The authors ask whether decoder-space geometry, namely neighbor density and maximum cosine similarity to nearby decoder directions computed from the SAE weight matrix alone, can predict steering cost before any forward pass. Geometry rank-correlates with steering cost at up to ρ = −0.546 with AUROC between 0.61 and 0.82, and the relationship replicates across Gemma-2 2B and 9B, two SAE widths, and more weakly on Llama-3.1-8B-Instruct. On Qwen3-8B with BatchTopK SAEs, geometry predicts whether a feature is steerable at all but not the ordering among responsive features, and the signal fades at deep layers where steering exceeds the intervention budget.

arXiv:2609.22782 PDF

Beyond Single-Model Injection: A Threat Model and Defense Architecture for Prompt Injection in Multi-Agent Systems

Rudrendu Kumar Paul, Sourav Nandy cross-listed Prompt injection research has centered on single-model chatbots, but multi-agent systems add injection channels through inter-agent message passing, privilege escalation through shared tool access, and trust propagation that lets a compromised agent influence upstream orchestrators. The authors construct a threat model of 14 attack vectors across direct user-input injection, indirect injection via tool outputs, inter-agent injection via messages, and cascading injection through orchestrator manipulation, and test all of them against a six-agent production-representative system. Even with system-prompt guardrails, 67% of agents are vulnerable to at least one scope violation and indirect injection via tool outputs succeeds in 43% of attempts. Four architectural defenses, message signing with provenance tracking, input/output sanitization at agent boundaries, privilege-scoped tool access per role, and anomaly detection on inter-agent traffic, cut overall injection success from 31.2% to 4.2%, with privilege escalation eliminated entirely.

arXiv:2609.22949 PDF

Auditing Political Alignment in LLM Assistants: Engagement, Stance, and User Identity

Joan C. Timoneda Audits of large language model assistants typically measure what they say to an average user, but their political behaviour is a set of policies about whom to answer, what to say, and whether to engage at all, conditioned on topic and inferred user identity; the author calls this a system's speech regime and derives five regime types from the dimensions of engagement and stance. A preregistered experiment ran 7,500 multi-turn conversations with six systems from OpenAI, Anthropic, xAI, Google, Mistral, and DeepSeek, randomly assigning the user's political identity across abortion, Catalan independence, climate change, Nazism, and a zero-stakes control about pineapple on pizza, with two cross-developer LLM judges validated against human coding and refusal treated as an outcome. Every system accommodates the user on the control topic, showing restraint is a policy, while on abortion GPT mirrors every user, Gemma refuses everyone, Claude answers strongly conservative users 35% of the time and almost no one else, and Grok accommodates conservatives only; on climate change and Nazism five systems hold firm for all users. The systems also infer the user's overall ideology so accommodation spills over to undiscussed topics, and two Grok releases show the regime shifting between versions in ways current audits miss.

arXiv:2609.23039 PDF

Directing large language models to follow the letter or spirit of the law

Peng Qian, Andrew Li, Sam Chen, Sonia K. Murthy, Yonatan Belinkov, Tomer D. Ullman Whether a rule should be followed by its literal wording or by the intention behind it is a central question for building safe machines, and the authors ask how large language models (LLMs) can be steered either way. They apply targeted adaptation with minimal modifications so that a model prioritizes either the spirit or the letter of the law, and evaluate the effect across diverse measures, novel vignettes, real-world scenarios, and influential legal cases. The intervention significantly shifts model behavior toward the chosen interpretation across all of these settings. An analysis of model internals reveals a low-dimensional space with three interpretable dimensions that match a pre-specified formal framework for the geometry of legal concepts, suggesting how legal thought in LLMs may be organized and directed.

arXiv:2609.23083 PDF

LLMs as Linguistic Chameleons: Decoupling Semantics and Structure for Privacy-Preserving Communication

Yuzhu Mao, Liang Zhao cross-listed Existing privacy protections for Large Language Model (LLM) API calls keep most of the original semantic content to preserve task utility, which leaves cues an attacker can exploit to reconstruct the input. CROSS-MAP pursues semantic decoupling instead: local models map private inputs into a different semantic domain that preserves the structure the remote LLM needs to reason, and then map the outputs back afterward. The local models are trained with multi-objective optimization to maximize semantic divergence in the mapping stage while minimizing semantic inconsistency in the recovery stage. Experiments show reduced reconstruction success across multiple attack settings together with higher utility than existing baselines.

arXiv:2609.23193 PDF

On the Efficiency-Safety Dilemma in Large Reasoning Models

Yifei Yang, Zouying Cao, Xingrui Wang, Xiao Zhou, Yuexian Li, Dongjie Yang et al. Large reasoning models (LRMs) are expensive to run, and efficiency techniques such as quantization and pruning are commonly applied, but their effect on adversarial robustness has been largely unexplored. The study analyzes the interplay between efficiency, jailbreak vulnerability, and reasoning ability across compressed LRMs, pairing attack success measurements with mechanistic analysis of representational drift. Efficiency methods appear to reduce jailbreak success, but the improvement is largely superficial: it stems from degraded reasoning producing attempted-but-failed malicious responses rather than genuine alignment, with a strict coupling between reasoning loss and the model's inability to sustain malicious semantic trajectories. Combining quantization with pruning is identified as the best balance of efficiency and robustness.

arXiv:2609.23587 PDF

Measuring the Assistant's Harmlessness Preferences on the User Turn

Jord Nguyen Post-training gives a next-token predictor a persistent assistant persona, and the question is whether that persona's preferences stay confined to the assistant's own turns or leak into how the model predicts what users will say. The authors measure a safety-relevant preference for harmless over harmful tasks on the user turn, comparing pretrained base models with post-trained versions across several open-weight model families. The preference is near-zero in base models but emerges through post-training, grows with scale, and can be shifted by narrow finetuning that never touches user turns. They read this as evidence that post-training reshapes the model's representation of the user rather than installing a shallow persona limited to the assistant's turn.

arXiv:2609.23935 PDF

Topographic Training Concentrates Causal Circuits Without Improving Neuron Monosemanticity

Gautam Ranka, Shubham Santosh Pandere, Aiden Dsouza cross-listed Mechanistic interpretability of vision transformers is usually blamed on feature superposition and attacked after training with sparse autoencoders or dictionary learning, leaving the network itself untouched. Adding TopoLoss, a spatial-locality training objective, to a ViT trained on ImageNet-100 tests whether a cheap training-time prior helps standard tooling, measured by activation patching for causal sufficiency of topographic clusters and sparse autoencoders for feature geometry. Topographic clusters become 2.79 times more causally sufficient than random unit sets of the same size, yet neuron-level monosemanticity scores do not change at all, while autoencoder sparsity falls 11% and the dead-feature fraction rises 19-fold. The authors read this dissociation as topographic pressure concentrating causal mass into spatially local circuits without disentangling individual neurons, implying current monosemanticity metrics are blind to a real class of interpretability gains.

arXiv:2609.24379 PDF

ToneCL: Contrastive Learning for Few-Shot Syllable-Level Tone Classification

Qisheng Liao, Youngah Do When a large language model offers an argument the user could not have built themselves, the user needs a principled way to decide whether to accept the claim. Drawing on interactive proofs, human-LLM deliberation is modeled as an exchange between a prover with unrestricted internal search and a resource-bounded human verifier who requests and checks supporting details without seeing the model's internals, with passed checks accumulating evidence toward an acceptance threshold. The analysis proves anytime-valid soundness against adaptive provers, so the probability of ever accepting a false claim stays below a chosen error level, provided the task supplies history-robust bounds on false passes and human checking errors, and gives a finite-horizon completeness bound that additionally needs bounds on honest-response adequacy and diagnostic progress. Whether certification is achievable depends on the verifier's effort budget, cognitive load, expertise, and fatigue, and conditions are identified under which a sequence of local checks can be certified while a global check under the same budget cannot.

arXiv:2609.24903 PDF

Emergent Collusion in Long-Horizon LLM Agent Interaction

Xinrui Shi, Yanzhe Zhang, Diyi Yang cross-listed Two large language model agents repeatedly complete individual tasks, share task logs, verify each other's work, and receive rewards in a long-horizon environment where following the verification protocol conflicts with maximizing reward. Over repeated rounds the agents increasingly deviate from the protocol, and collusion emerges in 94% of trajectories across 10 models, with more capable models within a family reaching it earlier. Controlled peer interventions show that a partner's behavior shapes collusion, and ablations attribute further effects to the reward structure, the verification feedback, and the interaction history, with restricting the amount and scope of available history reducing collusion.

arXiv:2609.24967 PDF

Rare Event Estimation via Iterative Unalignment

Hanming Yang, Daksh Mittal, Jing Dong, Hongseok Namkoong Autonomous agents can produce catastrophic outputs through rare stochastic variation in their own actions, so safe deployment hinges on estimating how often such events occur rather than whether they can. Naive Monte Carlo is prohibitively expensive at these probabilities, and building importance sampling (IS) proposals requires coordinated changes across a chain of context-dependent conditional distributions, so the authors construct the proposal by perturbing the model's own weights and searching over weight space with gradients, using an objective that pairs a differentiable surrogate for event amplification with adaptive regularization to keep the estimator stable. Evaluated on roughly 120M and 2.6B parameter models across three event families covering more than 300 events as rare as one in a billion, the estimator achieves over 800 times the compute-weighted efficiency of naive Monte Carlo for events with probability below 10^-7 in the most verifiable settings.

arXiv:2609.24969 PDF

6 more specialized papers

Multimodal 36

AgentVidBench: A Multi-Hop Video Question Answering Benchmark for Evaluating MLLM Agents

Seoyeon An, Hyeonseo Jang, Minsu Kim, Chanho Lee, Younghan Park, Kangwook Lee cross-listed Existing video benchmarks for Multimodal Large Language Models (MLLMs) mostly pose scene-level queries or global summaries that need only single-step inference, leaving multi-hop reasoning largely untested. AgentVidBench is a multi-hop video question answering benchmark targeting spatial, temporal, and causal reasoning, and it pairs each question with step-by-step solution traces so trajectory evaluation can check whether an agent actually acquires the evidence needed to justify its answer. Across 12 proprietary and open-source MLLMs, single-turn performance remains limited, while wrapping the same models in state-of-the-art agentic workflows generally improves both accuracy and trajectory scores. The authors also present a simple agentic strategy that serves as a competitive baseline and release the code and dataset.

arXiv:2609.21386 PDF

OmniVChat: Synthesizing, Benchmarking, and Training for Native Audio-Visual Dialogue

Haolin He, Yunfei Chu, Qi Chen, Wen Huang, Yuan Feng, Muzhi Zhu et al. cross-listed Omni Video Chat (OmniVChat) is defined as native audio-visual dialogue in which an omni model receives raw audio and video from a user and returns text, with no separate text query, external captioning, or speech recognition in the loop. Because recordings of people talking to their own devices are scarce and good replies depend on surroundings, facial expressions, and nearby objects in ways keyword matching cannot score, the authors build OmniVChat-Studio, a multi-agent engine that synthesizes single- and multi-turn audio-visual dialogues, use it to construct OmniVChat-Bench covering five ability categories, and design OmniVChat-RL, a reinforcement learning reward that jointly targets reply correctness, efficiency, and style. Training Qwen3-Omni-Instruct with this reward on synthesized dialogues improves performance on both the synthetic benchmark and the human-recorded OmniVChat-Bench-Human, indicating that gains from synthetic data transfer to real-world dialogue.

arXiv:2609.21465 PDF

PolyBridgeBench: Benchmarking Multimodal LLMs for Physics-Grounded Bridge Design

Zicheng Zhao, Dongyin Chen, Rui Xu, Yinghui Xu Multimodal large language models (MLLMs) handle visual understanding and structured generation well, but existing benchmarks do not test whether they can synthesize a complete load-bearing structure and repair it after a simulator exposes a failure. PolyBridgeBench gives a model a visual scene plus structured engineering constraints, asks it to output a full node-member-material bridge topology, gates execution behind deterministic legality checks, runs the design in a dynamic physics simulation, and on failure returns temporal visual evidence from the rollout and evaluates repair under a fixed interaction budget. Experiments with six MLLMs across 189 levels reveal a substantial gap between deterministic validity and dynamic success, strong sensitivity to material budgets, and limited post-failure recovery under the strict-budget setting.

arXiv:2609.21493 PDF

VidOmni-Bench: A Benchmark for Fine-Grained Video Understanding via Spatio-Temporal Event Verification across Complexity and Duration

Changbeen Kim, Junwon Chang, Kipyo Kim, Risa Shinoda, Kuniaki Saito, Donghyun Kim cross-listed Existing evaluations of Video Large Language Models (Video-LLMs) rely on question answering or caption matching, which models can pass through superficial cues and incomplete annotations. VidOmni-Bench instead asks models to verify whether each event in a dense video caption is supported by the video, using 500 videos across five complexity types and durations from 4 seconds to 90 minutes, with captions generated by diverse Video-LLMs and human-verified sentence-level labels so that incorrect sentences serve as hard negatives. Results show that Video-LLMs frequently hallucinate in dense captioning and also fail as verifiers, unable to reliably catch plausible but incorrect event descriptions, with weaknesses that vary by model, video complexity, and duration.

arXiv:2609.21521 PDF

Samsone: A Family of Open Small Audio Language Models for On-Device Inference

Piotr Masztalski, Micha{\l} K. Grzeszczyk, Olaf Sikorski cross-listed Large audio language models have grown to billions of parameters, while demand for privacy-preserving, low-latency processing pushes toward Small Audio Language Models (SALMs) that run on device. Samsone is a family of open SALMs for edge computing, with a core Samsone-134M model plus Samsone-99M and Samsone-356M variants used to study how performance scales with size. The authors claim Samsone-134M sets a new state of the art for its size class across multiple benchmarks and that the family is competitive with models orders of magnitude larger. Everything is trained on publicly available data, and the release includes training code, weights, mobile-optimized checkpoints, and an open-source Android app demonstrating real-time on-device inference.

arXiv:2609.21666 PDF

NemotronLabs VoiceChat: An Open Full-duplex Speech-to-Speech Model with Tool Calling Capabilities

Jagadeesh Balam, Travis Bartley, Edresson Casanova, Sanjay Chauhan, Chen Chen, Zhehuai Chen et al. cross-listed NemotronLabs VoiceChat is an open full-duplex speech-to-speech model that can listen, transcribe, reason, call tools, and speak within a single streaming architecture. It pairs a streaming speech encoder with a decoder-only language model that emits parallel output streams for agent text and structured function calls, adds an auxiliary RNN-T branch for incremental user transcription, and finishes with a streaming text-to-speech decoder. On Full-Duplex-Bench 1.0 it posts the lowest pause-handling takeover rates among open-weight systems, 100% takeover after user interruptions, and a 4.33/5 post-interruption response quality; it resumes after user backchannels in 93% of cases on Full-Duplex-Bench 1.5, scores 55.1 on VoiceBench, and reaches 82.5% tool-selection F1 on Full-Duplex-Bench 3.0, though argument accuracy and end-to-end tool execution remain areas for improvement.

arXiv:2609.21967 PDF

Generalized Multimodal Foundation Model

Huizi Cui, Zongbo Han, Chenggong Ding, Naichuan Xiao, Jialong Yang, Jingdong Chen et al. Deployed multimodal fusion models are typically locked to a fixed set of modalities and a single task, making them hard to repurpose. The authors ask whether one fusion model can serve arbitrary modality combinations and prediction tasks, arguing that such a model should encode transferable patterns of cross-modal correlation rather than anything modality-specific. Their approach trains on large-scale synthetic multimodal datasets generated from diverse causal structures meant to mimic real-world generative processes, then relies on in-context examples at inference to activate the relevant correlation patterns. Across 18 real-world datasets spanning 12 modalities and 11 prediction tasks, the resulting model reportedly matches specialized models without any task-specific adaptation.

arXiv:2609.22107 PDF

Read-Best Is Not Steer-Best: A Probing--Steering Layer Dissociation in Omni-Modal Large Language Models

Yibo Wang, Jisheng Dang, Bimei Wang, Yitao Wu, Wencan Zhang, Hong Peng et al. Omni-modal large language models fold text, audio, and image inputs into a shared residual stream where concepts such as emotion can be linearly decoded and causally altered by activation steering, and practitioners typically pick the injection layer as the one where a linear probe reads best. Testing that assumption causally across three independently developed omni-modal models, using emotion as a controlled testbed, shows it fails: the best layer for reading a concept and the best layer for steering it are different, with probe-best layers varying widely by architecture while steering-effective layers cluster in a narrow mid-to-late band of normalised depth. Paired random-direction controls show an approximately 26-fold causal gap, and logit-lens analysis suggests a staged forward process of causal handle, probing saturation, and vocabulary commitment, motivating a two-factor account in which steering needs both representational readability and downstream plasticity. The analysis also identifies a cross-modal emotion subspace organised by valence and arousal, with joy acting as a stable anchor across models.

arXiv:2609.22135 PDF

Dissecting Training-Free Uncertainty Estimation in Multimodal Large Language Models

Soroush Seifi, Vaggelis Dorovatas, Lin Li, Yarin Gal, Rahaf Aljundi Multimodal Large Language Models (MLLMs) perform well on many tasks, but quantifying how uncertain their predictions are remains underexplored despite mattering for safety-critical use. The authors organize training-free uncertainty estimation methods into three families, token-level methods operating on output probabilities, verbalized methods that prompt the model to state confidence or abstain, and semantic methods that measure uncertainty in a meaning space, and benchmark them across multiple datasets, model families, generations, and scales. No single family dominates: token-level entropy at sampling temperature 1.0 works best for short answers, verbalized abstention for sentence-length responses, and semantic methods for long-form generation.

arXiv:2609.22206 PDF

Common Cause, Not Cross-Attention: Blocking Visual Shortcuts in Audio-Video Generation

Jian Xu, Delu Zeng, John Paisley, Qibin Zhao Joint audio-video generators learn from data where an event's appearance and its sound are spuriously correlated, so models can predict sound from texture or material rather than from the causal event. Using an audio-visual structural causal model (SCM) in which audio is independent of the video's nuisance appearance by construction, the authors show that models letting audio read video through cross-attention or a shared latent learn this visual shortcut and synthesize the wrong event's sound when the appearance-event correlation is broken at test time. Routing both modalities through a shared common-cause latent does not fix the problem: a bottleneck, an unsupervised shared/private factorization, and a faithful shared-prior model all latch onto the appearance proxy, and only an intervention on the nuisance, formalized as counterfactual invariance, identifies the causal predictor. The mechanism is verified across synthetic and real settings, and a real pretrained video-to-audio generator changes its output substantially when a video is merely recolored or grayed.

arXiv:2609.22361 PDF

Vox-Infinity: Benchmarking the Limits of Long-Context Spoken Language Models

Xize Cheng, Wenxu Jia, Chenyuhao Wen, Dongjie Fu, Zehan Wang, Xinyu Zhang et al. Long inputs cause language models to forget salient information, and the problem is worse for speech, where audio needs far more embeddings than text to preserve semantic content and acoustic cues. Vox-Infinity is the first benchmark built to evaluate long-context understanding in spoken language models, extending audio history along turn count and turn duration across diverse interaction structures, with explicit answer-provenance annotations and samples organized by how much history is needed to resolve each query. Evaluating seven spoken language models reveals a clear recency effect: accuracy is higher when the supporting evidence sits close to the query and drops when it lies farther back in the dialogue.

arXiv:2609.22452 PDF

COT-TTS: Audio Context-Aware Text-to-Speech with Chain-of-Thought Reasoning

Weizhen Bian, Sitong Cheng, Rongxiu Zhong, Jiahao Pan, Liumeng Xue, Boyi Kang et al. Text-to-speech systems usually need explicit style instructions, whereas in real conversation the speaking style should follow from what was said before. The COT-TTS task gives a model the preceding conversation audio, the target text, and a reference voice, and asks it to produce explicit intermediate reasoning about context before synthesizing speech in the specified timbre. To support it the authors build a bilingual conversational speech dataset of 9 million training samples, including a 1 million high-quality subset, plus an 800-sample human-verified source-disjoint benchmark, and train end-to-end autoregressive models at 0.6B and 1.7B parameters that emit emotion-labeled transcripts, editable style inferences, and speech tokens. The models match much larger baseline systems with far fewer parameters, hold duration and emotional consistency, and produce context-appropriate variation in emotion, stress, and rhythm; the data pipeline, dataset, and models are slated for release.

arXiv:2609.22697 PDF

ParA-LLM: A Unified Approach to Paralinguistic and Acoustic Speech Understanding

Nishit Anand, Jiaqi Su, Ke Chen, Yunyun Wang, Dinesh Manocha, Ramani Duraiswami et al. cross-listed Audio large language models now transcribe speech at human level but miss paralinguistic cues such as speaker traits, expressive variation, and the acoustic environment. The authors define a taxonomy of 22 paralinguistic characteristics, build a dataset of over 1.2M audio question-answer pairs, and train ParA-LLM with a two-stage curriculum that moves from single-attribute questions to multi-attribute joint reasoning over speaker and acoustic properties. They also release ParA-Bench, 6,000 multiple-choice questions across speaker-speech, acoustic, and mixed categories, on which GPT-4o-Audio scores only 36%. ParA-LLM beats GPT-4o-Audio by 7.5 points on ParA-Bench and adds smaller gains on MMAU-Pro Speech and MMAR Speech.

arXiv:2609.22771 PDF

AVTR-1: Open Stack for Real-Time Interactive Avatars

Artem Kravtsov, Dmitrii Ziganshin, Vsevolod Poletaev, Gleb Balitskiy, Anastasia Tikhonova, Egor Burkov et al. cross-listed Talking-head and dyadic motion models now run in real time, but a live conversational avatar also has to synchronize with an external voice agent, schedule frames for playback, and handle interruptions. AVTR-1 is an open stack built around a 153M-parameter autoregressive flow-matching motion generator conditioned on both participants' audio, with an audio encoder adapted to streaming through self-distillation, and the authors analytically derive and validate its contribution to user-facing latency with two commercial voice agents. It leads the compared dyadic systems on all reported visual-quality metrics and most listening-motion metrics while staying competitive in lip sync, and runs in real time on data-center and consumer GPUs. Because conventional listening metrics do not show whether the partner's speech actually drives generated motion, the authors introduce Reference-Based Directed Granger Gain (R-DGG), which detects speaker-speech dependence for recorded listeners and all dyadic systems but not for talking-head generators, and they release weights, renderer, and serving backend.

arXiv:2609.22913 PDF

Spatial-Interactor: Learning Spatial Reasoning through Interaction with the Observable Physical World

Kaixiang Yao, Xu Wang, Miao Pan, Hu Xiyue, Weishi Wang, Daniel Dahlmeier et al. cross-listed Vision-language models (VLMs) need to perceive local state changes caused by object motion and viewpoint shifts and integrate them over long trajectories, but current spatial training mostly asks static questions about attributes and relations. Spatial-Interactor trains VLMs on interaction trajectories, which naturally pair a prior observation, an action, and a subsequent observation, through a three-level curriculum: passive world-state transitions, active self-state transitions, and long-horizon trajectories. The LSI-108K dataset supplies simulated and real trajectories for each level, and training applies supervised fine-tuning to the first two levels followed by On-Policy Distillation, where a teacher branch given segment-level transition descriptions supervises the student's on-policy chain of thought. Experiments across multiple VLMs and spatial benchmarks show consistent gains in both local transition modeling and long-horizon integration.

arXiv:2609.23038 PDF

MuLA-Bench: A Multilingual Long-Form Audio Understanding Benchmark via Multi-Tier Auditing

Zeyu Yang, Xinyu Zhang, Zibo Bi, Pei Zhang, Xize Cheng, Jin Xu et al. cross-listed Long-form audio understanding is usually reported as a single accuracy at a given context length, which hides how language, evidence type, and task interact to make questions hard. MuLA-Bench provides 5,038 open-ended questions over 1,769 in-the-wild recordings totaling roughly 1,378 hours across 16 languages and eight domains, with a balanced language-by-domain semantic track and a separate acoustic track that preserves naturally occurring non-speech evidence; questions are generated from grounded evidence, checked for shortcuts, and reviewed by language experts rather than translated from a shared source set. Evaluating ten audio-language models, the authors find that language rankings shift across domains and tasks, and that long-range retrieval is comparatively strong while precise clock alignment and factual grounding of natural acoustic events remain fragile, with temporal errors persisting even after the correct event has been identified.

arXiv:2609.23416 PDF

Collapse, Not Complexity: Failure-Conditioned Decomposition Repair for End-to-End Document Parsing

Xingyu Lin, Dehui Du cross-listed End-to-end document parsers increasingly ship an optional reasoning mode for complex pages, and a 180-page entropy-stratified study with a frozen 4B checkpoint finds that page complexity is the wrong trigger for it. Reasoning lowers mean quality by 2.21 Overall points at 1.54 times the tokens, a preregistered input-only predictor of its benefit is at chance (held-out AUROC 0.47), and the benefit concentrates on pages whose ordinary pass has already collapsed into degenerate repetition, which have lower layout entropy than healthy pages and are not fixed by doubling the budget or switching modes (83% recur). The proposed alternative detects collapse from the ordinary-pass trace, decomposes the page into regions by projection, and re-parses each region. This repair gains 1.40 Overall at 1.13 times the tokens, replicates across three checkpoints, and with all parameters frozen gains 2.41 on the remaining 1,175 benchmark pages.

arXiv:2609.23592 PDF

Representation-guided in-context learning for medical image interpretation with multimodal large language models

Minda Zhao, Fangyu Hu, Yan Luo, Yutong Yang, Jiahui Cai, Kaichen Zhou et al. cross-listed Adapting general-purpose multimodal large language models (MLLMs) to medical image interpretation usually requires costly domain-specific fine-tuning. Representation-guided in-context learning (RG-ICL) is a training-free alternative that uses frozen encoders to retrieve demonstration cases aligned with the query image, and for visual question answering (VQA) with the question intent as well, then places them in the model's context. Across eight histopathology, radiology, and retinal fundoscopy datasets, classification accuracy rose by a mean of 20 percentage points and VQA by 13 points over no-context and conventional in-context learning, approaching or exceeding fine-tuned comparators. Which examples are retrieved matters more than how many: six query-aligned cases beat up to 32 random ones, and fixed or random cases often pushed accuracy below the no-context baseline.

arXiv:2609.24057 PDF

18 more specialized papers

Reinforcement Learning 24

GVPO++: Group Variance Policy Optimization for LLM Post-Training and On-Policy Distillation

Kaichen Zhang, Yuzhong Hong, Junwei Bao, Hongfei Jiang, Yang Song, Dingqian Hong et al. Post-training methods such as Group Relative Policy Optimization (GRPO) improve LLM reasoning and task expertise but suffer training instability from their reliance on importance sampling. Group Variance Policy Optimization (GVPO) folds the analytical solution of KL-constrained reward maximization into its gradient weighting, so that the gradient corresponds to the mean squared error between the central distance of implicit rewards and that of actual rewards. The method guarantees a unique optimal solution that exactly matches the KL-constrained reward maximization objective and permits flexible sampling distributions without importance sampling. The authors further show that GVPO extends naturally to on-policy distillation (OPD) and can optimize a broad family of extended OPD objectives, offering a principled basis for objective design.

arXiv:2609.21432 PDF

Contrastive World Models

Bonnie Li World models trained by reconstructing pixels can get distracted in visually cluttered environments, where irrelevant background detail dominates the loss and crowds out information needed for planning and control. Building on Dreamer, the authors drop the observation decoder and replace reconstruction with a Deep InfoMax-style lower bound that maximizes mutual information between state-action sequences and local patch features of future observations, so the latent state keeps only what is predictive of the future. In small-scale experiments across three settings of increasing visual complexity, the contrastive model matches Dreamer and a momentum-prediction baseline in the clean setting and substantially outperforms both once distractors or natural-video backgrounds are introduced, while also training more efficiently because the pixel decoder is removed entirely.

arXiv:2609.22175 PDF

RLVR is a Kernel, Not a Function: Statistical Inference for pass@$k$ Crossovers

Chen Yang, Xianyang Zhang, Jun Chen cross-listed Reinforcement learning with verifiable rewards (RLVR) often raises pass@1 but trails its base model at large sampling budgets k, a crossover commonly read as proof that RLVR merely sharpens existing capability. The authors build paired confidence bands across k that require evidence of both an early gain and a later loss, and across five public RLVR model pairs no crossing is statistically established in the original evaluations, while a 32k-token evaluation on fresh prompts does locate a reversal with first loss somewhere between 11 and 61 samples; a power analysis explains why failing to detect a crossing does not rule one out and why more prompts help more than more samples per prompt. They further show that a prompt's base success rate does not determine its post-RL success rate, since prompts with equal base rates diverge reproducibly across independent generation halves, so the relationship is a conditional distribution (a Markov kernel) rather than a curve, and fitting it predicts crossings on held-out generations and corrects naive power estimates. Theory shows how losses on a minority of the hardest prompts can overturn an early lead even when most prompts improve, separating whether a crossover exists from what it implies about capability.

arXiv:2609.22547 PDF

FIRM-WM: State-factorized factual-interventional recurrent modeling for reward-free visual planning

Yilun Wu, Yunjian Zhang, Aobo Li, Mujiangshan Wang, Haitao Wu, Aqiang Zhang Reward-free latent world models learn from offline video and solve image-goal tasks by optimizing actions against predicted futures, but the planning state must be both directly comparable to a goal image and rich enough to carry velocity, contact, and other history-dependent information, and offline data reveals only one factual future per recorded state. FIRM-WM (Factual-Interventional Recurrent World Model) splits its recurrent state into a typed goal-comparable configuration and a 128-dimensional dynamic fiber excluded from the terminal goal cost, and supplements factual trajectories with intervention branches that reset the environment to the same recorded state before executing alternative action sequences. Under matched cross-entropy method (CEM) planning across three seeds, success rates reach 99.0% on TwoRoom, 92.7% on Reacher, and 88.0% on OGBench-Cube versus 89.0%, 88.0%, and 70.0% for LeWM, using a model of roughly 3M parameters with 2 to 12 times lower planning time.

arXiv:2609.22816 PDF

Towards Full Pipeline FP8 Reinforcement Learning for LLMs

Fanchao Chen, Ziheng Jiang, Ziyun Wei, Zheng Zhong, Du Li, Chi Zhang et al. Reinforcement learning (RL) is now central to improving reasoning and agentic abilities in large language models (LLMs), and FP8 quantization can speed up that training, but running the whole RL pipeline in FP8 has been unstable, with mid-training entropy surges and garbled outputs even after train-inference mismatch corrections like TIS. The authors trace the instability to compounded FP8 quantization noise distorting the importance ratio, which pushes negative-advantage tokens outside the trust region and zeroes their gradients, so pathological outputs go unpenalized and accumulate. Their fix, Calibrated Clipping, dynamically aligns the FP8 clipping bounds with the high-precision BF16 distribution by matching the lower-bound clipping quantile and rebalancing the upper bound. Across GRPO and DAPO, model sizes from 8B to 32B, and several FP8 scaling granularities, the method eliminates the entropy surges and restores performance comparable to the BF16 baseline.

arXiv:2609.22870 PDF

If You Hear It, Help Find It: Orthogonal Knowledge Distillation for Open-Vocabulary Audio-Visual Event Localization

Yi Xu, Cheng Chen, Wenzhuo Lei cross-listed Reinforcement learning (RL) fits healthcare decisions whose consequences unfold over time, yet most reported progress stays far from routine clinical intervention, and existing surveys organize the field by algorithm or application. This review instead arranges healthcare RL along an evidence ladder of problem formulation, retrospective identification, policy estimation, stress testing, prospective evaluation, and lifecycle monitoring, spanning clinical treatment, patient engagement, and health-system operations, with restless bandits treated as one special case rather than the organizing frame. It synthesizes the assumptions and failure modes at each rung, identifies what evidence can and cannot transfer across settings, and proposes reporting practices for cumulative evaluation, arguing that a policy scoring well on a historical dataset is not evidence it will improve care and that healthcare RL should be evaluated as an intervention embedded in a changing sociotechnical system.

arXiv:2609.23376 PDF

RLVR$^{2}$: Reinforcement Learning with Verifiable Rubric-based Ranking

Hao Li, Zhengkun Zhang, Gangqiang Hu, Zhen Zhang, Yude Gao, Dai Dai et al. Reinforcement Learning with Verifiable Rewards (RLVR) is increasingly applied to tasks judged by multi-dimensional rubrics, but policy optimization needs one scalar per rollout, and the usual practice of normalizing each criterion and taking a weighted sum assumes score differences are comparable across criteria and that strength on one criterion can offset failure on another. RLVR^2 instead converts each criterion's rubric scores into within-group ordinal comparisons, recovers a latent utility from the resulting comparison matrix, and merges the per-criterion utilities into a single training signal, discarding raw magnitudes so heterogeneous rubric scales never need calibration; it also lets auxiliary attributes that correlate with rankings inform the estimate without being rewarded directly. Across three model scales and 16 benchmarks, it consistently outperforms representative rubric-based baselines, and analysis shows it controls systematic effects tied to reasoning efficiency and response formatting while preserving the quality objective.

arXiv:2609.23457 PDF

OnlineWM: Causality-Aware Active Online Learning for Effective World Modeling

Yikun Miao, Fangqi Zhu, Quanxin Shou, Xiaoyi Pang, Zhengyang Yan, Junhao Li et al. cross-listed Generative world models predict future states from actions, but training on fixed simulator datasets misaligns the data with the model's evolving error patterns, and minimizing observational error lets the model lean on spurious correlations instead of true action effects. OnlineWM closes the loop by actively querying the simulator for interaction sequences targeted at the model's current weaknesses and by fine-tuning with a counterfactual objective that contrasts outcomes of different actions taken from identical states, forcing transitions to be attributed to the action rather than ambient environmental change. The combination significantly improves action controllability and generalizes to unseen domains in the reported experiments.

arXiv:2609.23753 PDF

VISTA: An Attention-Based Multi-Agent Reinforcement Learning Architecture for Space Situational Awareness Sensor Tasking

Miguel Leiva-V\'elez, Adalberto Claudio Quiros, Nicolas Gaston Rozado, Hodei Urrutxua, V\'ictor Rodr\'iguez-Fern\'andez The growing population of resident space objects makes space situational awareness sensor tasking harder, and existing deep reinforcement learning approaches use fixed-dimensional state and action spaces that cannot scale to large, changing catalogues and distributed sensor networks. VISTA (Variable-Entity Intelligent Sensor Tasking Architecture) combines physics- and mission-informed top-K retrieval with entity-centric attention, recurrent memory, and pointer-based action decoding so each agent's observation and action spaces stay independent of catalogue size. With 30 orbiting targets it recovers the catalogue 31.2% faster than a fixed-dimensional recurrent baseline, and in the large-scale regime it reduces five-hour uncertainty by 97.5% relative to the strongest classical method and 99.3% relative to the recurrent learner. Zero-shot tests up to 20,000 objects show near-linear relations between sensing capacity, catalogue size, and recovery horizon, and the learned policies adapt across sensor modalities and population shifts.

arXiv:2609.23875 PDF

Reinforcement Learning under State and Outcome Uncertainty: A Foundational Distributional Perspective

Larry Preuett, Qiuyi Zhang, Muhammad Aurangzeb Ahmad Real-world planning agents face uncertainty both about the environment's hidden state and about the variability of returns under a chosen policy. The authors extend Distributional Reinforcement Learning (DistRL), which models the full return distribution in fully observable settings, to Partially Observable Markov Decision Processes (POMDPs), introducing distributional Bellman operators for partial observability and proving their convergence under the supremum p-Wasserstein metric. Return distributions are represented finitely by psi-vectors, a generalization of the alpha-vectors used in classical POMDP solvers, and plugged into a point-based backup to give Distributional Point-Based Value Iteration (DPBVI). Source code is provided as a foundation for risk-sensitive control where rare, high-impact outcomes matter.

arXiv:2609.24103 PDF

Luck Is Not Skill: When Do Paired Rollouts Help Group-Relative RL of LLM Agents?

Nazmus Sakib Group-relative reinforcement learning scores rollouts of the same prompt against each other, but independent environment noise such as tool faults or grader errors can swamp those comparisons. The authors study paired rollouts, which share an event-keyed noise schedule within a group while keeping each rollout's marginal distribution, and show analytically that pairing removes between-schedule variance in reward contrasts but does not necessarily lower gradient variance, giving an exact condition and a counterexample for one-sided grader noise. In a preregistered study training a 2B tool-use agent with three seeds per design, pairing raised final noisy-test success under tool faults by 5.1 percentage points on average, yet it missed the registered learning-curve criterion under both tool faults and grader flips. A gradient probe on eight checkpoints found 21 to 63% lower covariance traces, supporting the variance mechanism without establishing a general learning-speed benefit.

arXiv:2609.24144 PDF

Information-Time Proximal Policy Optimization

Yongcheng Zeng, Xinyu Cui, Yan Song, Guoqing Liu, Hongsheng Xin, Kaike Zhang et al. Reinforcement learning with verifiable rewards (RLVR) for large language model reasoning normally treats each generated token as one time step of the Markov Decision Process, even though information along an autoregressive trajectory is distributed very unevenly. InfoPPO reparameterizes time by information density rather than token count, which makes both discounted credit assignment and the proximal update constraint state-dependent, the latter implemented as a per-token adaptive clipping threshold. The authors extend performance-difference and policy-improvement bounds to this information-time MDP and report consistent gains over competitive baselines on five competition-style math benchmarks with Qwen3 models, while keeping accuracy and response length stable under non-trivial discount factors that cause token-time PPO to degrade.

arXiv:2609.24380 PDF

Lifted Bellman Linear Programming for Offline Reinforcement Learning

Hyukjun Yang, Jongchan Park, Narim Jeong, Donghwan Lee Offline reinforcement learning (RL) usually trains a critic by regressing onto bootstrapped value targets, which requires target networks with exponential moving average (EMA) updates and off-policy correction for multi-step targets. The Lifted Bellman Linear Program (LBLP) instead imposes in-sample Bellman optimality as inequality constraints over a joint state-value and action-value space, so every constraint involves only state-action pairs present in the dataset, and adding constraints along K-step trajectory segments leaves its unique minimizer unchanged for any rollout policy. Relaxing the constraints into hinge penalties gives ALBUM (Approximate Lifted Bellman Unconstrained Minimization), a neural implementation with no squared regression onto bootstrapped targets and therefore no target networks or EMA updates. On OGBench, ALBUM matches the average performance of FQL and is comparable to recent action-chunking methods while using a single critic, a Gaussian policy, the fewest parameters, and the least peak GPU memory of all compared methods.

arXiv:2609.24489 PDF

Reinforcement Learning in Operational Research: A Technical Review and Practical Roadmap

Yahan Lu, Dongyang Xia, Nursen Aydin, Shadi Sharif Azadeh cross-listed Traditional Operational Research (OR) methods face growing pressure from real-time, data-driven decision-making in complex dynamic systems, and reinforcement learning (RL) has emerged as a complementary approach for sequential decisions under uncertainty. The review structures recent work around three roles RL plays: solving sequential decision-making problems in dynamic environments, acting as an end-to-end solver or as a component inside heuristic and exact methods for combinatorial optimization, and enabling extended reality analysis through integration with digital twin systems. For each role it synthesizes advantages, implementation requirements, limitations, and open challenges, and it closes with a roadmap for deeper methodological and practical integration of RL and OR.

arXiv:2609.24750 PDF

Critical-State RL: Diagnosing Trainable States for Multi-Turn Tool Use

Zixiang Chen, Wenting Zhao, Zhepeng Cen, Akshara Prabhakar, Jielin Qiu, Jianguo Zhang et al. In multi-turn tool use, a failure can hinge on a single model call, but reward variation alone cannot reveal which call would benefit from training because variation may come from downstream randomness rather than from differences between actions. Critical-State RL takes task-defined candidate calls and local rewards, checks whether each reward reflects the action's effect on task success and whether improvement over a reference policy is possible, then uses nested sampling to separate action-dependent reward variation from continuation noise and trains the policy at the selected states with contextual-bandit updates. On BFCL v4 (Berkeley Function Calling Leaderboard), the diagnostic picks the response after a tool becomes available for missing-function tasks and the response before a missing argument arrives for missing-argument tasks; training at these selected states improves performance by about 14 percentage points on the missing-function task, while training at alternative states leaves performance flat or worse. The authors also apply the recipe across models and tasks including logged repeat-call avoidance and memory management.

arXiv:2609.24985 PDF

9 more specialized papers

Unclassified 24

Decoupling Internal Representational Changes and Causal Importance in Fine-Tuned Large Language Models

Lingfang Li, Procheta Sen, Shubham Das, Danushka Bollegala No summary available — see the abstract on arXiv.

arXiv:2609.21113 PDF

From Task Success to Productive Success: Evaluating Human-AI Collaboration by Quality and Cost

Saki Imai, Mert \.Inan, Malihe Alikhani cross-listed No summary available — see the abstract on arXiv.

arXiv:2609.21117 PDF

The Stochastic Shift: A New Evaluation Paradigm for Text-to-SQL with AI Operators

Tarfah Alrashed, Fatma Ozcan, Per Jacobsson, Tal Neiman, Xianshun Chen cross-listed No summary available — see the abstract on arXiv.

arXiv:2609.21133 PDF

TinyCeNN-LM: Quality-Gated Conversion of Pretrained Attention with CeNN-Inspired Cellular-Recurrent Layers

Kabeh Mohsenzadegan, Vahid Tavakkoli, Kyandoghere Kyamakya No summary available — see the abstract on arXiv.

arXiv:2609.21139 PDF

Clinician-Grounded Quality Assurance for AI-Assisted Psychiatric Intake

King Shi, Amanda Li, Jonathan Ivey, Synthia Qia Wang, Guan Gui, Hyunseo Kim et al. No summary available — see the abstract on arXiv.

arXiv:2609.21149 PDF

EnSol: an environment-aware graph neural network for molecular solubility prediction

Thao Nguyen, Saman Shafaei, Zhengyi Zhang, Huimin Zhao, Heng Ji cross-listed No summary available — see the abstract on arXiv.

arXiv:2609.21151 PDF

Can Agents Design Better Chips with a Higher Level Abstraction?

Zijian Ding, Yang Zou, Yizhou Sun, Jason Cong No summary available — see the abstract on arXiv.

arXiv:2609.21157 PDF

SpecOpt: Contact-Diff Reasoning for Agentic Molecule Optimization Toward Binding Specificity

Thao Nguyen, Heng Ji No summary available — see the abstract on arXiv.

arXiv:2609.21165 PDF

Implicit Rule Induction with Test-Time Task Embeddings in ARC-like Tasks

Adrien Deli\`ege, Claas Beger, Marc Van Droogenbroeck, Melanie Mitchell No summary available — see the abstract on arXiv.

arXiv:2609.21181 PDF

SWE-Proof: Can Language Models Resolve Real-World Issues with Machine-Checked Proofs?

George Ma, Benjamin Mikek, Haoyu Li, Ferhat Erata, Yuhao Zhang, Zeren Shui et al. cross-listed No summary available — see the abstract on arXiv.

arXiv:2609.21190 PDF

AI-GRACE: A Use-Case Operationalization Framework for Agentic AI: From Organizational Objectives and Obligations to Deployment Capabilities and Architecture

John Cuneo, David Chun, Gaurav Khanna No summary available — see the abstract on arXiv.

arXiv:2609.21192 PDF

Information-Gain Rewards over Diversity-Pruned Tests: GT-Anchored Verifier Co-Training for Reliable Code Generation

Ana Nunez, Peyman Najafirad No summary available — see the abstract on arXiv.

arXiv:2609.21208 PDF

Visual Navigation Transformer with Pose Attention

Beiming Li, Jaime Romero, Jonathan Diller, Vijay Kumar, Alejandro Ribeiro cross-listed No summary available — see the abstract on arXiv.

arXiv:2609.21212 PDF

Ability-Residual Decoupled Modeling for Affective Cognitive Diagnosis

Boyuan Zhao, Meng Ye No summary available — see the abstract on arXiv.

arXiv:2609.21214 PDF

Fewer Steps, Better Actions: Rethinking Flow-Matching Inference for VLA Policies

Zhipeng Tang, Xinda Chen, Weining Rao, Xiao Li, Wenting Tan, Yuning Wang et al. cross-listed No summary available — see the abstract on arXiv.

arXiv:2609.21216 PDF

A Fully Differentiable Neuro-Soft-Symbolic Framework for Perceptual Task Planning

Hongyan Wei, Wael AbdAlmageed No summary available — see the abstract on arXiv.

arXiv:2609.21221 PDF

Hallucination-R1: Robustness-Oriented Paraphrase Generation for Factual Consistency

Wenhan Yu, Wenxin Wu, Hao Wang, Lei Sha cross-listed No summary available — see the abstract on arXiv.

arXiv:2609.21227 PDF

FOCAL-VLA: Subtask-Guided Geometry Distillation and Implicit World Modeling for Vision-Language-Action Models

Zhiyuan Gao, Di Wen, Yanxiang Zhan, Mohammad Khoshnazar, Jeroen Sch\"afer, Kunyu Peng et al. cross-listed No summary available — see the abstract on arXiv.

arXiv:2609.21228 PDF

KnowDemo: Knowledge-Guided Robot Demonstration Generation from Human Videos

Zhiyuan Gao, Yanxiang Zhan, Mohammad Khoshnazar, Jeroen Sch\"afer, Michael Beetz cross-listed No summary available — see the abstract on arXiv.

arXiv:2609.21229 PDF

VLA-Scope: Shift-Aware Failure Prediction for Vision-Language-Action Models

Kaiwen Zhu, Dongfang Liu, Liangkai Liu cross-listed No summary available — see the abstract on arXiv.

arXiv:2609.21246 PDF

The Evidence Ladder for Reinforcement Learning in Healthcare: From Retrospective Policies to Trusted Interventions

Yunfan Zhao No summary available — see the abstract on arXiv.

arXiv:2609.23374 PDF

Toward a Unified Mathematics of Concepts

Chen Shani No summary available — see the abstract on arXiv.

arXiv:2609.24554 PDF

Complex KDA: Understanding and Enhancing the Expressivity of Kimi Delta Attention

Julien Siems, Riccardo Grazzi, Korbinian P\"oppel, Jaisidh Singh, Arber Zela, Timur Carstensen et al. No summary available — see the abstract on arXiv.

arXiv:2609.24797 PDF

Learning Prognostic Variables for AI Convective Parameterizations via Symbolic Distillation

Jurij Sch\"onfeld, Tom Beucler, Julien Savre, Steven Sherwood, Veronika Eyring No summary available — see the abstract on arXiv.

arXiv:2609.24882 PDF

Robotics 22

AtomEgo: Exploring Ego-Robot Integration for Embodied Foundation Model Pretraining

Di Wu, Dongchen Zheng, Junhe Sheng, Zhongxing Wei, Songxin Zhang, Zejian Xie et al. cross-listed Embodied foundation models are limited by the scale and diversity of robot demonstrations, and it is unclear how to fold in large egocentric human interaction data given the embodiment and action-space gaps between humans and robots. AtomEgo is a systematic study of ego-robot co-training built on a curated corpus of roughly 2,659 hours and a scalable processing pipeline, comparing three paradigms across vision-language-action and world-action model architectures: joint co-training with domain-specific action heads, progressive ego-to-robot transfer via embodiment alignment, and joint video-action modeling. Multi-task real-robot experiments and cross-embodiment representation analysis point to a simple principle, capability gain scales with data scale multiplied by alignment quality: egocentric data improves generalization only insofar as it is well aligned and effectively used.

arXiv:2609.21461 PDF

Potential-Field Action Representation for Reinforcement Learning in Contact-Rich Manipulation

Xinyu Liu, G\"okhan Solak, Arash Ajoudani cross-listed Model-free reinforcement learning for contact-rich manipulation forces the policy to learn both task strategy and low-level motion generation, and the action representation determines how hard that is. PA-RL has the policy adapt the parameters of an artificial potential field, whose state-dependent guidance direction is executed by a Cartesian impedance controller, rather than commanding motion directly. On simulated peg-in-hole insertion against Cartesian velocity, Cartesian pose, and variable-impedance action spaces with the same RL algorithm, it is the only method to reach 100% evaluation success within the training budget (best baseline 92.6%), while also cutting joint-torque variation by 55.4% and Cartesian acceleration variation by 70.8% without any motion-quality penalty in the reward. The simulation-trained policy completes 9 of 9 real-robot insertions without fine-tuning.

arXiv:2609.21609 PDF

SynthDemo-RL: Breaking the Zero-Reward Barrier in VLA Adaptation with LLM-Guided Synthetic Demonstrations

Hiroaki Kingetsu, Hiroaki Kurihara, Kaoru Yokoo, Kenji Fukumizu, Manohar Kaul cross-listed Fine-tuning Vision-Language-Action (VLA) models usually depends on human teleoperation demonstrations, and reinforcement learning (RL) with sparse binary rewards stalls when successful trajectories are almost never sampled. SynthDemo-RL has an automated teacher turn simulator-privileged state into successful manipulation trajectories, distills a VLA student from them by supervised fine-tuning, and then refines the student with PPO on binary task-success rewards. On LIBERO-PRO, a perturbed version of LIBERO with no demonstrations, 27 of 57 scored tasks sit at exactly 0% success for a pi_0.5 policy tuned on the original tasks; direct PPO rescues 10 of them, whereas SynthDemo-RL rescues all 27 with 50 synthesized trajectories per task and no new human data, reaching 97.8% and 97.1% average success on the Position and Task axes. On standard LIBERO the same pipeline reaches 96.0% without human demonstrations, within 1.7 points of pi_0.5 trained on 50 human demonstrations per task, and trajectories from a MuJoCo-twin policy execute open-loop on a physical robot.

arXiv:2609.21650 PDF

ForceTwin: Physics-informed Digital Twins for Robotic Manipulation from Instrumented Human Interaction

Tim Engelbracht, Ren\'e Zurbr\"ugg, Mayank Mittal, Marco Hutter, Marc Pollefeys, Hermann Blum et al. cross-listed Manipulating articulated objects requires knowing physical properties such as inertia, friction, and spring or door-closer mechanisms that visually identical objects can differ on, yet existing digital-twin pipelines recover mostly kinematics or assign static parameters from visual and language priors. ForceTwin identifies physics-informed digital twins from a person probing the object with a handheld force-sensing gripper, using the synchronized poses and forces to estimate articulation, inertia, Coulomb friction, viscous damping, and a structured neural residual for state-dependent mechanism forces. It nearly halves the inertial-parameter error of a vision-language-model prior, and as a feedforward dynamics model for impedance control on a Spot and a Franka FR3, it reaches 87% goal completion across nine object-embodiment pairs versus 60% with the VLM prior and 57% with kinematics-only twins, with the largest gains on objects whose strong mechanisms stall both baselines. The identified twins are also used to train whole-body door-traversal policies deployed in the real world.

arXiv:2609.21751 PDF

When Should a Failing Robot Ask? Initiating Corrective Human-Robot Dialogue from Audited Sensor Evidence

Eshika Pathak, Leela Krishna cross-listed When a robot fails at a task it must decide whether to act on its own diagnosis, consult another onboard sensor, or interrupt a person, which requires knowing what its sensors reveal about the cause and how reliable its diagnosis is. The authors build a simulated benchmark with injected, known failure causes, measure per-sensor diagnosability with explicit leakage checks (force data reaches 0.99 while no image method exceeds 0.55), and test six open vision-language models. Model behavior tracks the surface of the prompt rather than the evidence: moving the refusal option from last to first collapses refusal rates from 78-100% to 0-6% in three of six model-family sweeps, image-based accuracy never beats a majority-class baseline, and stated confidence carries no information about correctness. Handing the models the force data as ten lines of text produces the first above-baseline diagnoses in four of six models, and the authors frame the choice as a three-action decision whose optimal policy should follow from measured accuracy and stated question cost rather than model confidence.

arXiv:2609.21942 PDF

StateMem: Single-State Residual Memory with Adaptive Inference for Vision-Language-Action Policies

Wenzhuo Li, Qiongfeng Shi, Yi Zhou cross-listed Memory-dependent manipulation requires later actions to draw on earlier interactions, yet most vision-language-action (VLA) policies act on the current observation alone, and memory-augmented alternatives such as MemoryVLA need external banks with explicit storage and retrieval. StateMem instead maintains a single persistent memory token updated through low-rank residuals driven by prediction error, and uses the same signal to decide adaptively when to reuse cached vision-language-model prefixes, with a training-free controller that tunes the routing threshold online and a fast correction step for stale prefix features. On LIBERO it reaches 97.6% average success while cutting prefix refreshes by 20.25% relative to full refresh; in the Occlusion category of RoboMemArena it posts the best single-VLA results at 21.8% task success rate and 44.3% cumulative success rate, and across six real-world tasks it improves average success by 21 points.

arXiv:2609.22684 PDF

Prioritized Rollouts for Efficient World Model-based Vision-Language-Action Policy Optimization

Yifei Sheng, Haoxiang Ren, Zhilong Zhang, Haonan Wang, Runjie Xu, Yihao Sun et al. Fine-tuning Vision-Language-Action (VLA) models with reinforcement learning is limited by the cost of real robot interaction, and model-based reinforcement learning (MBRL) with a learned world model helps but becomes expensive as policies and world models grow, especially since existing methods spend rollouts on all states equally. The authors show that policy uncertainty is concentrated in a small subset of states, often at decision-sensitive moments where small action differences change the task outcome, and that these states offer the most potential for policy improvement. They introduce U-GROW, a plug-and-play sampling layer that changes only the distribution of branched rollout start states to focus world-model rollouts on high-uncertainty states, leaving the policy optimization objective untouched. Experiments on simulated and real-world manipulation tasks show improved efficiency and effectiveness over uniform rollout allocation, supporting policy uncertainty as a guide for experience generation.

arXiv:2609.22879 PDF

Robot World Models Are Not Invariant to How the Actions Are Written

Ahmed Karim, Leon Chlon cross-listed Robot policies are trained with either absolute joint targets or deltas relative to the current state, and an action-conditioned world model silently inherits that choice. A latent dynamics model trained on one parameterization and given the identical commanded trajectory written in the other collapses: retrieval degrades by 2.6-13.4x across three robot datasets and two morphologies, goal-conditioned action selection falls from 53% to 15%, and on PushT the two predicted beliefs about the same future are near-orthogonal. This is not ordinary distribution shift, since the two encodings are mutually reconstructible at R^2 = 0.996, and the authors give a test that separates a valid re-parameterization from a lossy summary or a sensor swap. Averaging the training objective over both encodings restores task performance, and a disagreement penalty closes worst-case agreement from 0.78 to 0.995, though averaging alone does not repair the axis on PushT.

arXiv:2609.23252 PDF

Latent Telepathy: Multi-Robot Communication with Self-Supervised Perceptual Latents

Howard Wang, Han Zheng, Cathy Wu cross-listed In decentralized multi-robot teams under partial observability, the observation that should drive one robot's next action is often visible only to a teammate, and standard messages carrying position or planned trajectory cannot convey perceptual content. Latent Telepathy has each robot broadcast the perceptual latent it already computes for itself, produced by an encoder trained with a self-supervised joint-embedding predictive objective and then frozen and shared across the team, so the message adds no computation, means the same thing to every receiver, and is interpreted by the recipient purely from task reward. In a content-controlled protocol that fixes bandwidth, latency, topology, and receiver, broadcasting the latent lets a navigator avoid an occluded hazard in 99.7% of episodes, matching a noiseless hand-designed message, while position and trajectory messages stay at chance and a raw camera image 186 times wider is less reliable. The result holds from a gridworld to rendered pixels under continuous velocity control and on a physical robot's camera, and the authors identify that porting multi-agent reinforcement learning (MARL) communication results to continuous control requires the message-informed decision to remain reachable by exploration.

arXiv:2609.23269 PDF

Marginal Calibration Does Not Compose: Hidden Dependence in Modular Robot Navigation

Rista Baral cross-listed Robot stacks are built from independently developed perception, prediction, and planning modules, and each can be validated for calibrated uncertainty in isolation. Using a moving-obstacle prediction pipeline, the authors show that a position estimator and a velocity estimator can each be well calibrated on their own, yet the correlation structure between their errors changes future-state uncertainty enough to make the composed system either overconfident or needlessly conservative, which directly affects planning and safety. Simulations show that modeling the joint covariance restores downstream calibration and improves performance, while dependence-robust bounds preserve safety at the cost of conservatism, arguing for module interfaces that carry dependence information or for calibration at the system level.

arXiv:2609.23731 PDF

Learning tactile perception from high-bandwidth single-point sensing

Joseph Rigal, Emmanuel Virot, Caroline Pascal cross-listed Learning-based robotic manipulation increasingly incorporates touch, but most approaches rely on spatially distributed tactile arrays. SpectRobot converts high-bandwidth single-point tactile signals into fixed-size time-frequency spectrograms that standard vision encoders and vision-oriented learning pipelines can consume, with sensors mounted away from the contact surface yet mechanically coupled to it to reduce wear. Experiments show a robot can solve a visually occluded manipulation task from single-point vibration alone, that temporal history strongly influences policy performance while sensing bandwidth (measured up to 100 kHz) governs the available spectral content, and the same representation transfers across acceleration-, force-, and strain-based tactile sensing technologies, including readily available off-the-shelf hardware.

arXiv:2609.24621 PDF

D-JEPA: A Decision-Aligned Latent World Model

Shuaijun Liu, Chengyu Wu, Qifu Wen, Feiyang You, Chenglong Zhang, Shuyang Hao et al. cross-listed Latent world models predict the consequences of actions, but accurate prediction does not guarantee that latent distance to a goal correctly ranks the few candidate futures competing for execution, a shortfall the authors call the decision-local prediction gap. D-JEPA is a decision-aligned latent world model that learns decision-relevant relations among candidate futures from executed outcomes, using a bounded permutation-equivariant operator over goal-relative predictive features and ordinal evidence, with restricted predictor adaptation and a shared ordinal interface so the alignment carries into JEPA-compatible representations usable by ordinary latent-distance planning. Across latent control, manipulation, pretrained action-producing models, physical robots, and autonomous driving it improves action selection, including 87.89% success on PushT, a 15.04-point average gain on RoboTwin, and a 17-point gain on physical robot tasks.

arXiv:2609.24749 PDF

PredActor: Predictive Action Diffusion for Steerable Onboard Humanoid Control

Lei Ye, Haibo Gao, Yitang Li, Peng Xu, Zetong Jing, Junhan Sun et al. cross-listed Diffusion policies offer flexible motion generation, but hierarchical humanoid controllers hand physical execution and recovery to a separate tracker, action-only diffusion lacks a future-state trajectory for test-time objectives, and joint state-action diffusion typically depends on privileged full-body state. PredActor is a predictive action diffusion policy that, from proprioceptive history and optional task context, jointly generates executable actions and an internal future-state trajectory, using classifier-free guidance to strengthen text-conditioned behavior and classifier guidance to steer predicted states toward test-time objectives, and only the actions are executed. In simulation it reaches all 15 destination targets and scores 0.580 on text retrieval versus 0.373 for conditional action diffusion, with similar disturbance survival. Rolling denoising and runtime optimizations bring the full control callback to a 16.790 ms median and 19.383 ms p95 on a Jetson Orin NX, under the 20 ms control period, enabling deployment on a Unitree G1 for text-conditioned motion, disturbance response, joystick control, and semantic interpolation.

arXiv:2609.24840 PDF

9 more specialized papers

Vision 22

Physically Based Rendering in the Latent Space

Vuk Radovanovic, Vishesh Gupta, Adrien Gruson, Binh-Son Hua cross-listed Image diffusion models are hard to control compared with classical physically based rendering pipelines. Observing a correspondence between light transport and the distribution of values in the latent space of a generative model's variational autoencoder, the authors perform physically based rendering directly in that latent space, modifying the rendering equation so a differentiable renderer can recover scene parameters that render with minimal refinement into the pretrained latent space for physically guided generation. Trained on a single rendered image, the method generalizes to changes in scene geometry, lighting, and camera viewpoint.

arXiv:2609.21054 PDF

LoRA Enhanced Contrastive Learning with SAS Vision Transformers

Dan Zimmerman, Frank E. Bobe III, Amelia L. McCormack, Matthew Cook, Gregory D. Vetaw Automatic target recognition (ATR) in synthetic aperture sonar (SAS) imagery is hampered by scarce target examples and clutter such as rocks that mimic man-made objects. The authors adapt DINOv3 Vision Transformers with a three-stage parameter-efficient pipeline, Low-Rank Adaptation (LoRA) on a frozen backbone, then hard-negative mining, then Supervised Contrastive Learning (SupCon), evaluated on at-sea data with a mission-level geographic split at 85% recall over three seeds. LoRA alone raises area under the precision-recall curve from 0.300 to 0.679 while training only 0.26% of weights, whereas hard-negative mining and SupCon each shift AUPRC by less than 0.005 versus matched controls, so a single adaptation stage suffices and stacked refinement adds nothing.

arXiv:2609.21061 PDF

ALPINE: Adaptive Localization for Parameter- and Sample-Efficient Few-Shot Learning

Neeraj Yadav cross-listed Few-shot learning is mostly judged on accuracy alone, with little attention to the parameter and training-sample budgets needed to reach it. ALPINE is an ultra-lightweight spatial-relational architecture of 22,249 to 34,917 parameters for few-shot image classification that pairs fixed Gabor edge-energy guidance with a windowed, content-adaptive patch locator. Under a strictly matched protocol of 250 meta-training episodes, five seeds, and 600 evaluation episodes per seed, it beats Prototypical Networks, Relation Networks, and MAML in 5-shot accuracy on both CIFAR-FS and MiniImageNet across all five seeds while using 27-53% fewer parameters, converges faster, transfers to CUB-200-2011 without retraining, and is more robust to occlusion and translation. Falsification ablations show the content-adaptive patch locator, not the pairwise relational computation, drives performance, and a capacity sweep finds an accuracy plateau near 22-35k parameters.

arXiv:2609.22323 PDF

Mask-Aware Execution for Efficient JEPA Training

Md Musfiqur Rahman Sanim, Zhihao Shu, Bahram Afsharmanesh, Amirali Mirian, Wei Niu, Gagan Agrawal cross-listed Joint Embedding Predictive Architectures (JEPAs) are increasingly used for representation learning and latent world models, but their training pipelines run each input through multiple mask-specific branches with redundant target-side work and memory-bound token routing, costs that grow with the number of masks. M-JEPA is a mask-aware execution architecture that leaves the learning objective untouched and instead separates mask-independent computation from mask-dependent routing, enabling shared context-encoder execution, fused token routing and slicing with backward support, sparse target-encoder execution over the union of target tokens, and masked patch embedding for sparse inputs. Implemented for five JEPA variants on NVIDIA A100 GPUs, it delivers up to 1.7x end-to-end training speedup for 2 to 10 masks and a 4.75x patch-embedding speedup at high sparsity, supporting the claim that execution restructuring rather than objective changes is the main lever for efficient JEPA training.

arXiv:2609.22674 PDF

Merge++: Universal Merge Refinement Through Data-Free Checkpoint Inversion

Aditya Pola, Vineeth N. Balasubramanian Model merging combines fine-tuned expert checkpoints into one multi-task model without retraining, but existing data-free methods operate purely on weights and never observe how each expert actually behaves, since behavioral signals need inputs to evaluate on. Merge++ inverts the expert checkpoints to synthesize task-representative images and then distills the experts' knowledge into the merged model using those images, requiring nothing beyond the checkpoints themselves. It works as a post-hoc refinement stage on top of any weight-space merging algorithm. Applied to methods from simple task arithmetic to spectral merging, it yields average gains of 2 to 8 points and up to 25.9 points on individual configurations.

arXiv:2609.22886 PDF

Why Do Video Diffusion Models Violate Physics? Unveiling the Flaws in Attention Mechanisms

Yueyan Li, Haibo Wang, Caixia Yuan, Xiaojie Wang cross-listed State-of-the-art video diffusion models produce visually convincing clips that often break physical laws, and rather than adding external priors the work looks for the cause inside the model. An interpretability study of the motion-planning process in text-to-video diffusion shows trajectories forming during early denoising in a first-shape-then-details pattern, and combining cross-attention trajectory patterns with causal head contributions isolates a subset of attention heads that drive motion planning. Self-attention analysis reveals that Rotary Position Embedding (RoPE) induces excessive spatial attention decay, causing candidate regions to lock prematurely into physically implausible positions and suppressing plausible trajectories in adjacent frames. A lightweight fix that scales the RoPE frequency across denoising steps reduces this decay, and both training-free and training-based experiments show improved physical commonsense in generated videos.

arXiv:2609.23658 PDF

Rethinking Diffusion Segmentation: When Does It Rely on Its Noisy State, and Does Diffusion Matter?

Hengzhuo Yang, Yuming Zeng, Yuling Yang cross-listed Diffusion models are increasingly adapted for conditional prediction tasks such as segmentation, but when the conditioning image already supports direct prediction, endpoint accuracy alone cannot show that the noisy diffusion state is used or that diffusion adds anything over image-only prediction. The authors retrain twelve published diffusion segmentation methods across three datasets with ten matched seeds, disrupting the target-derived state or the image-state pairing, and compare against matched image-only counterparts. Every comparison whose evaluated mask depended on reconstructing the noised quantity showed state reliance, while every method with a segmentation-supervised bypass preserved reference performance, and rerouting bypass-capable methods through noise-to-mask reconstruction flipped them to state reliance. Matched image-only counterparts achieved similar or better performance in 28 of 35 settings, indicating that the supervision path determines state reliance and that diffusion-specific computation often yields no deterministic endpoint advantage.

arXiv:2609.23967 PDF

Reinforcement Learning Inspired Black-box Adversarial Attacks for Computer Vision

Florian Krone, Elena Hoemann, Sven Hallerbach Convolutional and transformer vision models are vulnerable to imperceptible adversarial perturbations, but most attacks assume white-box access to the model, which is unrealistic in practice. The reinforcement learning inspired black-box attack (RIBA) borrows query-efficient ideas from reinforcement learning to optimize perturbations against a non-differentiable target using only its outputs. Compared with state-of-the-art black-box attacks, RIBA needs 25.4% fewer median queries against a ResNet-18 on Cifar10 and 22.5% fewer against a ViT-B/16 on ImageNet, and it matches white-box attack performance on an adversarially trained model.

arXiv:2609.24249 PDF

14 more specialized papers

Reasoning 17

CaLR: Causal Latent Revision for Robust Diffusion Reasoning

Wei Cai, Jian Zhao, Yuchen Yuan, Xuelong Li Autoregressive models commit greedily to each token, while diffusion language models (DLMs) generate in parallel but lack the strict causal structure that reasoning needs. Causal Latent Revision (CaLR) recasts reasoning as constrained latent optimization: a causal topology matrix taken from an expert model plus implicit differentiation drive gradient-guided thought revision that enforces logical consistency and lets intermediate steps self-correct during parallel generation. The authors report state-of-the-art DLM performance on complex reasoning benchmarks, surpassing strong autoregressive baselines and showing particular robustness on constrained tasks such as Sudoku.

arXiv:2609.20981 PDF

LogicTrack: Auditing Reasoning Trajectories of Large Language Models with Formal Logic Solvers

Jingyu Hu, Shu Yang, Weiru Liu, Di Wang Chain-of-Thought (CoT) training typically rewards only final answers, so large language models (LLMs) can reach correct conclusions through logically flawed intermediate steps that go unchecked. LogicTrack is a neuro-symbolic framework that auto-formalizes each reasoning step into symbolic form, verifies it with automated theorem provers, and uses the resulting Solver-Based Backtracking Reward (SBR) to score steps and drive backtracking tree search at inference time; the backtracking traces are also turned into supervised fine-tuning data so models internalize step-wise auditing. Across 8 reasoning benchmarks and 7 LLMs, the method improves both the verifiability of reasoning chains and the final-answer pass rate.

arXiv:2609.21492 PDF

On Repulsive and Attractive Teachers: Separating Correctness from Behavior in Self-Distillation

Anton Baumann, Akmal Ashirmatov, Leo Schmidt-Traub, Frederike L\"ubeck, Jonas H\"ubotter, Thomas Kleine Buening et al. cross-listed On-policy self-distillation conditions a model on privileged information and distills the resulting teacher distribution back into the model, but privileged information changes not just what the teacher knows but how it behaves, entangling correctness signals with behavioral shifts. The authors contrast attractive self-distillation, which pulls the model toward a privileged teacher, with repulsive self-distillation, which pushes it away, and find that attraction suppresses exploratory reasoning and shortens responses while repulsion lengthens responses, can trigger unintended switches into a latent thinking mode, and eventually becomes unstable. Combining attraction toward a correct-solution-conditioned teacher with repulsion from an incorrect-solution-conditioned teacher, studied in isolation from any GRPO objective, largely cancels the shared behavioral shifts and leaves a token-level signal that more directly reflects correctness, improving reasoning performance with stable response lengths across non-thinking, instruct-only, and thinking models.

arXiv:2609.21561 PDF

Calibrating Teacher--Student Discrepancy for On-Policy Distillation

Qiangqiang He, Jin Li, MingCai Chen On-policy distillation (OPD) trains a student reasoning model on the token-level discrepancy between its own outputs and a stronger teacher's, but that discrepancy also contains the teacher's own deviations, which the student learns indiscriminately, and privileged OPD amplifies the problem by inducing larger teacher-side likelihood shifts. Cal-OPD (Calibrated On-Policy Distillation) estimates the teacher's self-deviation region using positive and negative privileged interventions and keeps only the part of the discrepancy that lies outside that region as the training signal. On mathematical reasoning benchmarks, it consistently outperforms standard OPD and its variants across model scales while retaining only about 52-65% of the original discrepancy.

arXiv:2609.21619 PDF

When Steering Fails in Latent Reasoning: A Latent-to-Language Transition Gap

Gaoxiang Huang, Lei Qi cross-listed Activation steering is widely used to control language models during explicit chain-of-thought (CoT) reasoning, motivating its extension to latent CoT with continuous thoughts. Steering those continuous thoughts turns out to produce substantially weaker effects on subsequent language generation than steering explicit CoT, even when hidden representations are shifted by comparable amounts. Since task information remains identifiable in the continuous thoughts, the authors propose a latent-to-language transition gap, supported by two findings: the output distribution changes abruptly at the transition boundary, and task-related directions exert much weaker bidirectional control in latent CoT than in explicit CoT. They argue this transition interface should be the central target for evaluating and designing latent-steering methods.

arXiv:2609.21662 PDF

Dissecting Hierarchical Reasoning Models: A Mechanistic Study

Leo Raphael Rodrigues, Jian Kang The Hierarchical Reasoning Model (HRM), a hierarchical Transformer-based latent reasoning architecture with many variants, is dissected on Sudoku, Maze, and ARC-AGI-2 using Transformer baselines with and without recurrent modules, causal interventions on recurrent states, linear probes checked against random-direction ablations, and sparse autoencoders (SAEs) with feature ablations. Recurrent models beat one-pass baselines, but a single-state recurrent Transformer performs comparably to HRM, and the causal contributions of the high- and low-level states vary across task-specific checkpoints and inference stages. Selected task variables are linearly decodable from recurrent states, yet ablating probe directions has effects comparable to random controls, and while SAE ablations produce larger behavioral changes, top-ranked SAE features show no stable advantage over size-matched random subsets at larger ablation sizes or across tasks, a pattern that persists in a Sudoku control with within-step backpropagation through time. The authors conclude that HRM implements constraint-aware iterative refinement over a puzzle-specific solution state without relying on a compact, causally important feature set, and call for interpretability techniques better suited to latent-space recursive reasoning models.

arXiv:2609.22197 PDF

Assessing Adversarial Robustness of Latent Reasoning Models

Shaolong Chen, Ang Li, Mingjie Li, Yisen Wang Latent reasoning models (LRMs) compress intermediate chain-of-thought (CoT) steps into a few continuous vectors to cut memory and inference cost, but their behavior under adversarial input has received little attention. The study evaluates eight models on six benchmarks spanning textual and multimodal settings under adversarial perturbations, including white-box attacks, and compares them against explicit CoT baselines. Latent reasoning models are generally less robust than explicit CoT baselines, with especially severe degradation under white-box attacks. Failure modes differ by modality: textual latent states show brittle dynamics and high sensitivity to specific input patterns, while latent states in multimodal models stay largely invariant to perturbations and have limited influence on the final prediction, and the evaluation code is released.

arXiv:2609.22228 PDF

Do Chess Explanations Reflect Model Decisions? Behavioral and Token-Level Tests of LLM Reasoning Faithfulness

Angelina Parfenova Fluent natural-language explanations for a chess move may not reflect the reasoning that actually produced it, so the authors use chess, where board states are fully observable and move quality can be scored independently, as a controlled testbed for explanation faithfulness. Across 200 Lichess endgame puzzles they measure whether a decoder can recover the chosen move from the explanation, with and without explicit move hints masked, and score legal candidate moves at the token level under different explanation texts. Once explicit move hints are masked, explanations give only small, decoder-dependent gains over the board state alone, yet random plausible explanations borrowed from other puzzles still lower the probability of the correct move, showing the text is not simply ignored. The results separate linguistic plausibility, consistency with the generated move, and move correctness as distinct properties.

arXiv:2609.22245 PDF

Contextual Causality with Large Language Models: A Survey

Yiheng Zhao, Jun Yan, Chengming Hu Identifying causal relations that hold in a specific situation, rather than in general, is a prerequisite for large language models to support reliable decision-making, yet no systematic treatment of this contextual causality exists. The survey proposes a taxonomy with three categories, semantic, intervention, and counterfactual causality, and characterizes each by its core causal question, required model capabilities, representative tasks, and practical uses. It then reviews existing studies, identifies their key limitations, and maps the gaps between current benchmarks and real-world needs, closing with directions for future research.

arXiv:2609.22409 PDF

LLaDA-PRM: A Bidirectional Step-Level Reasoning Evaluator

Yiming Feng, Naihao Deng, Yulong Chen, Rada Mihalcea Step-level reasoning evaluators are usually built on autoregressive models whose causal attention lets each step's representation see only the problem, prior steps, and itself, even though an earlier step's validity often becomes clear only from what follows. A controlled 54-run comparison of causal versus bidirectional LLaDA evaluators at 1B to 3B scale, changing nothing but the self-attention mask, shows bidirectional attention consistently helps. Building on that, LLaDA-PRM is an 8B bidirectional evaluator that reaches 88.8 step-level F1 on MR-MATH-invalid and 83.8 on the out-of-distribution MR-GSM8K original-question subset, beating ReasonEval-Llemma-34B by 11.3 and 10.3 points; it also stays effective on incomplete traces in online settings and serves as a training-data selection signal that improves Mistral-7B on MATH-500.

arXiv:2609.22700 PDF

From Concept Alignment to Causal Grounding: An Intervention Test of Chain-of-Thought Faithfulness

Qianli Wang, Yilong Wang, Dennis Wei, Jingyi Sun, Simon Ostermann, Pepa Atanasova et al. Chain-of-thought (CoT) explanations can sound plausible while being unfaithful to the model's actual reasoning, and most prior faithfulness tests rely on input-output behaviour rather than internal computation. The authors recast faithfulness as concept grounding: encoding both a direct-prediction pass and a CoT pass through a single shared sparse autoencoder (SAE) makes their internal concepts comparable, yielding three correlational alignment metrics plus a causal metric, Δp, that ablates the shared concepts and measures the drop in answer probability. Across five large language models and four datasets, concept alignment is generally high, but causal faithfulness varies substantially with depth, peaking at mid-to-late layers rather than the final ones, and model scale reshapes the layer-wise profile. Causally important shared concepts are often not verbalized in the CoT, so faithfulness cannot be judged from surface or representational correspondence alone.

arXiv:2609.23065 PDF

Euston: Training Away Mathematical Sycophancy Without Losing the Mathematics

Zehua Cheng, Wei Dai, Jiahao Sun Reasoning language models are trained to produce solutions rather than refuse them, so when handed a corrupted theorem they tend to deliver a confident proof of something false. Euston is an 8B mathematical claim-verification model built by fine-tuning DeepSeek-R1-8B with GRPO under a rule-based reward, using 3,026 matched true/corrupted statement pairs generated by GraphSynth, a probabilistic factor-graph generator drawing on arXiv papers from 2010 to 2025. On a balanced held-out split, balanced accuracy rises from 29.50% to 63.75% and the gap between correctly rejecting false statements and wrongly rejecting true ones moves from roughly zero to +27.5%. AIME 2026 accuracy drops only from 69.17% to 65.00%, a statistically insignificant change, whereas an earlier run on a smaller corpus collapsed to 40%, and the authors report confounds including the all-false composition of official evaluation sets and low precision at realistic error prevalence.

arXiv:2609.23205 PDF

Increasing Skill Level Recruits Deeper Attention Layers in a Frozen Chess Transformer

David Litman cross-listed Chess provides a deterministic setting for studying computation inside transformers, and the Maia-3 chess model takes a player's Elo rating as an input, so the skill it is conditioned on can be varied without changing its weights. Ablating every attention head at every Elo from 700 to 2500, the authors find that increasing skill pushes the causal center of mass of the computation monotonically deeper for every piece and move type measured. The depth migration is much larger for specific tactics, especially knight forks, than for other move types, and it consists of deeper heads being recruited for more specialised computations while one shared shallow head keeps a roughly constant contribution. The results may indicate how conditioning inputs redistribute computation in larger transformers.

arXiv:2609.23917 PDF

Efficient Reasoning Exploration via State-Conditioned Latent Steering with Progress Guidance

Hengyuan Zhang, Chenming Shang, Zunhai Su, Xiao Liang, Hui Shen, Jing Xiong et al. Best-of-N sampling for reasoning tasks relies on candidates covering diverse, high-quality paths, but post-trained reasoning models often exhibit exploration collapse, where independent rollouts retrace the same reasoning and extra samples add little. State-conditioned Progress-guided Steering (SPS) is a training-free latent steering method that builds a Direction Bank of progress-guided steering vectors keyed to regions of prefix state, then at inference retrieves the vector matching the current prefix and applies it at high-uncertainty transitions to push the next step toward meaningful progress. Across multiple model scales and benchmarks, SPS consistently outperforms strong exploration baselines, and ablations support the state-conditioned retrieval and progress-guidance design choices. Code is public.

arXiv:2609.24066 PDF

3 more specialized papers