AI Paper Digest

Friday, October 23, 2026

425 papers cs.AI · cs.LG · cs.CL ← 2026-10-222026-10-29 →

Aug Sep Oct Nov

Highlights

Breaking the high-degree barrier in Lehmer's problem with a human-AI discovery loop

Highlight Applications Yuri Chervonyi, Sergei Gukov, Thang Luong Lehmer's 1933 problem asks whether the Mahler measure of an integer polynomial can be made arbitrarily small, but small-measure irreducible polynomials had been constructed systematically only up to degree 180, and Boyd's one-parameter family skips nearly half of the higher degrees because it factors. The authors run a human-AI discovery loop in which an AI agent builds numerical and symbolic solvers, launches distributed searches, and works with a human researcher to form and verify hypotheses. The result is a five-parameter Generalized Boyd algorithm that produces irreducible polynomials with Mahler measure below Boyd's limit point for every target degree. Across all even degrees up to 2000, it improves six published values and holds 99.6% of the best known records, and the verified polynomials are released as a dataset.

Lehmer's 1933 problem asks whether an integer polynomial's Mahler measure can sit arbitrarily close to 1, but beyond degree 180 nobody had a systematic way to build irreducible polynomials with small measure. Boyd's one-parameter family often factors at those degrees and skips nearly 50% of them. The authors pair Gemini running in Google's Antigravity agent environment with a human researcher, fast numerical search, exact irreducibility checks and cloud-scale sweeps, and from what the search turned up they derive a five-parameter construction that reaches every even degree.

  • Search workers stream candidates to separate verification workers, which use Mathematica's IrreduciblePolynomialQ to check irreducibility exactly while the search keeps running; the searches themselves combine RL (a cross-entropy-method policy), evolutionary search, priority search, direct sparse search and periodic tiling with a central defect, and the tiling method accounts for 54.2% of leaderboard entries.
  • Across all even degrees up to 2000, the search produced best-known polynomials for 824 degrees; on the 91 degrees with published values it matches 85 and improves 6, including d=178 inside the previously well-searched range and d=188, where the best measure drops from 1.28309 to 1.25208.
  • Multiplying the dense search optima by suitable cyclotomic polynomials cancels their repeating blocks, and 95.5% of those for d≥30 then reduce to 6–7-term specializations of Boyd's two-variable polynomial F(X,Y); this observation led to the Generalized Boyd algorithm, which solves for the exponents and signs (a, b, k, s0, s1) needed to hit any target even degree directly.
  • Generalized Boyd produces verified irreducible polynomials with measure strictly below Boyd's limit point α₁≈1.255434 for all 995 even degrees from 12 to 2000 (the worst case is 1.25535826 at d=1932), and it holds 99.6% of best-known records, including every degree d≥166.
  • All of these are constructions, so they only show that small measures exist and prove no lower bound: every search method levels off just below α₁ for d≥30, Lehmer's degree-10 record of 1.17628 still stands, and the Mahler measures are computed numerically with numpy.roots (only irreducibility is checked exactly), so the paper's talk of "breaking" the high-degree barrier refers to coverage across degrees, not to progress on the conjecture itself.

arXiv:2610.30277 PDF

Random Remasking Scales Masked Diffusion Inference

Highlight Large Language Models Zhaoyi Zhou, Frank Wu, Andrea Zanette Masked diffusion models (MDMs) generate text by progressively unmasking tokens, and errors made during decoding are normally irreversible. The authors recast MDM inference as a Markov chain over clean sequences that randomly re-masks and then denoises, which motivates Random Remask-Denoise (R²D), a test-time procedure that needs no fine-tuning and no reward. They also propose Trajectory Correction Training (TCT), which trains the denoiser on its own errors along the generation trajectory. On logic tasks the approach shows near-log-linear test-time scaling, exceeding 97% accuracy on hard Sudoku and 99% on 3-SAT, and on pretrained models it beats iterative-correction baselines and the official Q-mode accuracy of LLaDA2.1-mini on HumanEval+.

Masked diffusion models normally freeze each token once it is revealed, so errors made during decoding cannot be undone. The authors recast inference as a Markov chain over complete sequences, where each step randomly remasks tokens and then denoises them, and derive a one-step KL bound that motivates a test-time sampler (R²D) and a matching training method (TCT).

  • R²D (Random Remask-Denoise) needs no finetuning and no reward model: it remasks tokens uniformly at random at a fixed noise level, repeats until the argmax prediction stops changing, and only then moves to a lower noise level.
  • With standard training on 6.8M-parameter models, accuracy scales near-log-linearly with the correction budget, rising from 19.7% to 93.5% on Sudoku-Hard and from 86.0% to 98.5% on 3-SAT; standard decoding reaches 48.7%, ReMDM 54.7% and ProSeCo 55.7% on Sudoku-Hard.
  • Trajectory Correction Training trains the model on its own partial decoding trajectories, with a loss at every position, including tokens it already revealed; combined with R²D, it reaches 97.5% on Sudoku-Hard and 99.3% on 3-SAT.
  • On HumanEval+ with the off-the-shelf 16B LLaDA2.1-mini, R²D scores 85.4%, beating the model's published Q-mode result (82.9%), ProSeCo (81.7%) and ReMDM (71.6%), while using about half of ReMDM's forward passes thanks to early exit.
  • Limitations: at the largest budget, per-sample latency is about 77× that of standard decoding on Sudoku-Hard (49.4 s vs 0.64 s); the scaling depends on uniform rather than low-confidence remasking; and the paper attributes the strongest gains to tasks with sparse, local constraints, with only one language benchmark tested.

arXiv:2610.30744 PDF

Is Complex Agentic Memory Always Needed? Naive RAG Is Stronger Than We Thought

Highlight HF pick · 8▲Agents Zhengye Bao, Shengtao Zhang, Haokai Zhang, Shiqi Lou, Muning Wen, Minnan Luo Agentic memory systems for long-term conversation often rely on memory mutation, structured organization, or learned retrieval, and the authors test whether that complexity is needed. Their baseline, Raw-RAG, simply applies dense retrieval and reranking over the raw conversation history and passes the evidence to a reader LLM. Across LoCoMo, LongMemEval, MEME, and MemoryAgentBench, Raw-RAG ranks first in six of seven benchmark-model settings, beating the strongest agentic memory system by 0.13 to 5.53 points. Analysis attributes this to higher coverage of necessary evidence, including the relations needed to derive answers, and shows that even full access to transformed memory stores leaves an accuracy gap relative to raw history.

Agentic memory systems for long-term conversation add extraction, mutation, graphs and learned retrieval, but it is unclear how much that complexity buys over plain retrieval. The authors test Raw-RAG, a deliberately naive baseline that indexes raw chat history without any LLM processing and lets the reader model work out state changes from the retrieved evidence.

  • Raw-RAG splits timestamped history into 256-token windows with 32-token overlap, retrieves the 128 nearest windows by dense similarity, and keeps the top 15 after cross-encoder reranking; dropping the reranker costs 3.92 points, while chunk size and moderate top-k values barely matter.
  • Across LoCoMo, LongMemEval, MEME and MemoryAgentBench, it ranks first in 6 of 7 benchmark–model settings against 12 systems including EverMemOS, MemMachine, Mem0 and Memora, leading the strongest of them by 0.13–5.53 points (e.g. 93.70 on LoCoMo with GPT-4.1-mini); the one loss is MEME with DeepSeek-V4-Flash, where MemMachine is ahead by 1.52 and Full-context by 19.98.
  • An audit of Mem0 finds that updates and deletes make up only 4.39% of write decisions on LoCoMo and 2.46% on LongMemEval, and replacing them with append-only writes raises LoCoMo accuracy from 67.36% to 69.42%; adding memory management on top of Raw-RAG also hurts, with atomic-fact extraction costing 19.62 points, summaries 5.81 and graph expansion 5.78.
  • The analysis attributes the advantage to evidence coverage: Raw-RAG most often retrieves sufficient evidence (45.29% on MemoryAgentBench, 76.35% on MEME) and best preserves the dependency paths behind an answer (62.34%), and even with oracle access to a system's full store, transformed memory still falls 7.58–29.38 points short of raw evidence.
  • It is also much cheaper, ingesting and searching a MEME slice in 306 s, which is 5.1× faster than the next system and 30.4× faster than EverMemOS; the limits are that results cover only offline public conversation benchmarks, rely on a capable reader to reconstruct state, and leave open the storage and privacy costs of keeping raw history.

arXiv:2610.30755 PDF

Scaling Laws for Generative Reward Models

Highlight Large Language Models Angel Raychev, Nathan Lile, Rafael Rafailov Generative reward models (GenRMs) act as judges that supply preferences during policy training. The authors study them for creative writing across Qwen3 judge sizes from 0.6B to 14B, comparing Thinking judges trained with GRPO against Answer-Only judges trained with supervised fine-tuning. Thinking judges score higher on static gold-label accuracy, but Answer-Only judges produce better policies through online DPO in seven of eight matched settings, and earlier, less accurate Thinking checkpoints can be better rewarders than later ones. Larger judges help substantially, even when they exceed the policy's size, and the results show that static judging accuracy does not determine downstream optimization quality in these pipelines.

Generative reward models (GenRMs) that score higher on static preference benchmarks don't necessarily produce better policies when they supply online training signal. The authors test this in creative writing by training Qwen3 judges as online-DPO rewarders, sweeping judge format, judge size and policy size, and scoring the resulting policies against a fixed learned 32B "Gold" evaluator.

  • A Qwen3-32B Gold judge is fine-tuned on vote-derived LitBench Reddit story preferences, reaching 79% agreement with those labels, and then relabels pairs to train 0.6B–14B proxy judges. The proxies come in two kinds: Answer-Only judges trained with SFT and Thinking judges trained with GRPO. Each proxy then labels fresh response pairs for online DPO of 0.6B–8B policies.
  • Thinking judges are more accurate statically at every size (the 4B GRPO recipe reaches 83.2% versus 60.0% off the shelf), but the ranking reverses in training: Answer-Only rewarders give higher peak and top-five Gold Elo in 7 of 8 matched settings, with a mean peak gap of 126 Elo, and 5 of 8 bootstrap intervals exclude zero.
  • Static accuracy can also mislead when choosing a checkpoint: an intermediate 250-step Thinking judge (79.2% accuracy) trains better policies than the later 300-step judge (82.3%), and Thinking-trained runs lose on average 130 Elo from peak to final versus 60 for Answer-Only.
  • Judges larger than the policy keep helping: a 14B Answer-Only judge with a 4B policy peaks at 1719 Gold Elo, a descriptive log-linear fit reaches R² of 95.2% with roughly 136 Elo per judge-size doubling for Answer-Only, and the smallest DPO β tested (0.002) lifts peak Gold Elo from 357 to 1122.
  • All results rest on a learned Gold judge from the same model family and a single seed per configuration, and they mix two changes at once, the training objective (SFT versus GRPO) and the response format; a check with a different judge, Qwen3-235B, was inconclusive (52.5%, 42 of 80), and high-Elo outputs collapse into templates, with 94% of one setting's peak outputs sharing the same opening line.

arXiv:2610.30765 PDF

Inherited, Not Invented: Tracing the Provenance of Wait in Reasoning Models

Highlight HF pick · 3▲Reasoning Zhizhuo Kou, Sizai Hou, Long Chen, Jian Yang, Junyu Luo, Jingshu Peng et al. Reinforcement learning (RL) is often credited with teaching reasoning models to say "wait" before reconsidering a step. The authors trace the word through open models' training histories and find that RL from a base model almost never invents it: on bases that lack the word, RL substantially raises MATH-500 accuracy while "wait" stays nearly absent. The marker comes from training data, through mid-training in OLMo-3 or fine-tuning on reasoning traces, and RL recipes amplify, reshape, or even erase it depending on KL regularization and on whether fine-tuning came first. The authors warn that marker frequency alone does not indicate RL-induced reasoning, and that LLM-as-a-judge revision metrics can reverse trends labeled by humans.

RL is widely credited with teaching reasoning models to say wait before reconsidering a step, but the marker usually comes from training data instead. The authors trace it through the open training histories of 120 model endpoints using label-free indices of frequency, position, form and usage, plus an injected-error test, all computed only from sampled outputs on MATH-500.

  • Zero-RL almost never coins the marker: across 18 runs from 9 marker-free bases (Qwen2.5, Mistral, Qwen3), accuracy rises by a median of 42.9 points while wait stays below 0.05 per 1k tokens in 17 of 18, with General-Reasoner-Qwen3-4B (0.29 per 1k) the lone exception.
  • The marker arrives with data: in OLMo-3 it jumps from 0.04 to 4.03 per 1k tokens during mid-training, and 13 of 14 models fine-tuned on reasoning traces from marker-free bases adopt it.
  • When the base already writes wait, the released RL-Zero recipes (OLMo-3, MiMo) multiply it (OLMo-3: 4.9 to a peak of 17.0 per 1k), but KL-free GRPO from the same base does not; the authors' GRPO recipe even erases a marker seeded by fine-tuning (7.52 per 1k down to zero), and OLMo-3 RL-Zero erodes the canonical Wait, form from 88.9% to 37.5% of uses.
  • In the injected-error test, RL-Zero makes revision markers a more sensitive but more liberal alarm (d′ from 0.57 to 1.09, criterion from 0.11 to −0.24), and in a 60-sentence blind audit one LLM judge reversed the human-labelled revision trend (26.0% to 12.9%, against expert consensus of 1.3% to 15.9%).
  • The evidence is observational, and each open series is a single training run; all RL runs use 0.5B–32B models trained on English mathematics with outcome rewards, and the measurements capture the word itself rather than the underlying reflective behaviour.

arXiv:2610.30882 PDF

SCOPE: Compiling Natural-Language Policies into Neurosymbolic Predicate Programs

Highlight HF pick · 2▲Safety & Alignment Arnon Mazza, Elad Levi Enforcing compound, application-specific policies over multi-turn LLM agent transcripts and tool calls is hard: fine-tuned guardrail models degrade on such policies, and LLM-as-a-judge is costly and unstable. SCOPE compiles a natural-language policy into an executable program of neural predicates, run by small models and deterministic extractors and composed by code. The program is optimized jointly with its training data through adversarially generated counterexamples. It achieves the highest average accuracy against prompt-optimization, decomposition, and workflow-generation baselines. Programs optimized with the weakest executor transfer to other models, which lets inexpensive models reach accuracy comparable to frontier reasoning models with more stable and interpretable verdicts.

Guardrail models fine-tuned for policy enforcement lose accuracy when a policy combines many conditions, exceptions and counts, while frontier LLM judges are expensive and give inconsistent verdicts across runs. SCOPE compiles a natural-language policy into Python code that combines narrow yes/no-style predicates answered by small non-reasoning models with deterministic extractors, and it optimizes this program together with a training pool that grows through adversarially generated counterexamples.

  • Strong LLM agents (Claude Opus 4.8) run the search: they trace each failure back to a specific predicate or to the composition code and revise it, while two adversaries generate counterexamples aimed at a single predicate ("local") or at the whole program ("systemic"), and a teacher labels each one; a counterexample is kept only when the teacher's label agrees with the adversary's intended label.
  • With gpt-5.4-nano as the executor, SCOPE scores 0.85 / 0.82 / 0.87 / 0.78 on the four benchmark tasks (Coaching, Escalation, Plan, Tile-match), beating GEPA, MIPROv2, CheckEval and RLIE everywhere, and it ranks first or ties in 10 of 12 task–executor settings, losing the other two to AFlow (for example 0.82 vs 0.86 on Escalation with gpt-5.4-nano).
  • Decomposition makes accuracy much less dependent on the executor: moving from the weakest to the strongest executor adds only 0.09 accuracy under SCOPE versus 0.26 under prompt optimization, and weaker executors lead the search to split the policy into more predicates.
  • Programs compiled for gpt-5.4-nano transfer without re-optimization to open-weight models running without reasoning, which gain an average of 0.13–0.22 per model; DeepSeek-V4.1-Flash reaches 0.78 / 0.91 / 0.86 / 0.89 against gpt-5.5 with high reasoning effort at 0.94 / 0.89 / 0.89 / 0.82, at a fraction of the cost, and only 5–7% of its Tile-match predictions change across five runs compared with 37% for gpt-5.5 and 26% for gemini-3.1-pro.
  • Removing the adversaries significantly lowers accuracy on three of the four tasks (for example Plan drops from 0.865 to 0.813, p=0.005) and has no effect on Escalation. Limitations: correctness depends on how the teacher model reads the policy rather than on formal verification, and the benchmark is small, with only four tasks of 163–371 samples each.

arXiv:2610.31062 PDF

Beyond Watching and Searching: Benchmarking Agents for Video-centered Deep Research Reports

Highlight HF pick · 9▲Agents Jiahao Meng, Qize Yang, Yuxuan Wang, Yunfei Chu, Wensi Dai, Jin Xu et al. Deep research agents write long reports from open-web evidence, but no existing benchmark evaluates research built around a source video. The authors define a task in which an agent takes a video and a user request, identifies research questions, gathers external multimodal evidence and produces an interactive report linked to moments in the video. They release OmniDR-Bench, with 50 Chinese and English tasks and 6,414 criteria organized into rubric trees (Omni-ARTs) that were built through an audited, human-reviewed process. Across 18 agent configurations the best system scores only 63.3 out of 100, and the main gaps are content coverage, causal reasoning across modalities and citation accuracy.

Deep research agents can already write cited reports from web sources, but no existing benchmark tests research that is anchored to a specific source video. The authors define video-centered deep research report generation: an agent takes a video and a user request, finds research-worthy claims and procedures in it, gathers external text, image and video evidence, and produces an interactive HTML report linked to timestamps in the video. They pair this task with OmniDR-Bench, which scores reports against audited rubrics.

  • Reports are graded with Omni-ARTs (Omni Atomic Rubric Trees). Each tree has three dimensions (content, citation and presentation), 23 shared directions and task-specific binary criteria, which are checked either by a multimodal judge (GPT-5.6 Sol) or by deterministic programs that probe the report in a browser.
  • The criteria come from ARC (Audited Rubric Construction), which builds an independent library of video and web facts before writing any criteria, checks that scores rise with more evidence access, and runs rounds of human review; the final benchmark has 50 Chinese and English tasks, 19 subcategories and 6,414 criteria, and 94% of sampled items were marked correct.
  • Across 18 model and harness configurations, the best score is 63.3/100, from Claude Opus 5 in Claude Code; GLM-5.3 scores 62.0 with every task completed, and the confidence intervals overlap, so the ranking between the two is uncertain.
  • Agents handle presentation (averaging 81.8) much better than content (45.4) or citation (47.8): video faithfulness reaches 92.4%, but claim support is only 15.9% and use of external evidence about 8.7%, often because sources are listed at the end instead of cited next to the claims they support.
  • In ablations, the full video-research skill adds +7.3 points over the individual skills alone, mostly through a +20.7 gain in citation score, but it roughly doubles cost (from 2.3M to 4.7M tokens and from 0.7 to 1.6 hours per report); the main limitations are the small task count, dependence on a single judge model, and only moderate agreement with human holistic ratings (Spearman ρ = 0.64).

arXiv:2610.31158 PDF

Cola2: Co-Adaptive Hierarchical Language Modeling with Continuous Semantic Compression and Balanced Discrete Generation

Highlight HF pick · 7▲Large Language Models Hongcan Guo, Yian Zhao, Feng Wang, Shen Nie, Guoqiang Wei, Yan Zeng et al. Cola2 is a hierarchical language model that separates continuous semantic modeling from discrete text generation. It segments text into semantic spans, compresses each span into a continuous latent, and models those latents with a Flow Matching prior. A block-causal masked diffusion decoder then turns the latents into text, with classifier-free guidance controlling how closely it follows the latents versus generating on its own. Trained on 346 billion source tokens, an unconverged checkpoint averages 60.10 across eight tasks and beats Llama-3-8B on HellaSwag and StoryCloze, even though Llama's reported pretraining budget is over 43 times larger.

Hierarchical language models split generation into continuous semantic modeling and discrete text realization, but how well this works depends on what the latent space compresses and how the token decoder uses it. Cola2 trains the encoder, latent prior, and discrete decoder together so that the latents stay compact and predictive, and so that the decoder learns to balance reconstructing the semantics against generating text on its own.

  • Text is split dynamically into semantic spans, each compressed into a continuous latent; the compression rate is held on target by negative feedback rather than an auxiliary loss, and training combines reconstruction and generation objectives.
  • A block-causal masked diffusion decoder is trained on clean, noisy, and dropped latent conditions, and classifier-free guidance then sets how much the decoder relies on the latents versus its own generation.
  • Token-prediction gradients flow back through the latents into the encoder, while the Flow Matching prior is kept out of representation learning by a stop-gradient; at inference, generated text is re-encoded under a fixed span layout so the latent history matches the actual text.
  • Ablations on five downstream tasks with about 210B training tokens support the core design choices and show competitive scaling that has not yet saturated; with 346.03B source tokens, an unconverged checkpoint averages 60.10 across eight tasks and beats Llama-3-8B on HellaSwag and StoryCloze, even though Llama-3's reported pretraining budget is over 43× larger.
  • The headline comparison uses a checkpoint that has not finished training and covers only selected benchmarks, so it remains unclear how the approach holds up against strong discrete baselines trained at matched compute.

arXiv:2610.31305 PDF

Rethinking Sequence Compression with Test-time Regression

Highlight HF pick · 9▲Large Language Models Sihan Yang, Jie Huang, Hanxuan Li, Yifan Guo, Weigao Sun, Yiran Zhong et al. Sequence compression is a key bottleneck in efficient language and multimodal models, and common pooling or projection compressors perform poorly. TTR-Compressor treats compression as test-time regression. The input tokens define a regression problem learned end-to-end, and the fitted coefficients, computed in closed form during the forward pass, form the compressed representation. As the compressor for block-wise sparse attention it improves RULER scores over attention-based compressors by a large margin, and as a visual token merger for multimodal LLMs it matches the standard MLP merger with 55.6% fewer visual tokens.

Pooling, learned projections and attention-weighted pooling mix tokens without accounting for how they relate to one another, so redundant tokens pile up and drown out distinctive ones. TTR-Compressor treats each block of tokens as a small ridge-regression problem and uses the fitted coefficients as the compressed vector, so correlated tokens are fitted jointly through the Gram matrix and redundant evidence counts for less.

  • Keys are the L2-normalized token features, targets come from a learned projection of the unnormalized tokens, and a learnable diagonal regularizer takes the place of a hand-tuned λ; the solution is computed in closed form with a T×T Cholesky solve, costing O(T²d) and adding only 2d parameters per compressor, all trained end-to-end.
  • As the block compressor in HiLS-Attention sparse attention on a 345M-parameter model trained on FineWeb with 8K context, it raises average RULER needle-retrieval accuracy (4K–32K) from 0.392 to 0.509 over attention-weighted pooling, and a variant with convolution-enriched targets reaches 0.514, a gain of 12.24 points, including 0.988 on S-NIAH-1 at 32K.
  • As the visual token merger for a Qwen3-VL ViT feeding Qwen3-4B-Instruct, it beats the standard MLP merger at 2×2 merging (56.03 vs 54.93 average) and by 2.58 points at 3×3, with the largest gains on text-rich benchmarks such as InfoVQA (+4.54) and OCRBench (+4.10).
  • At 3×3 merging it scores 54.85, nearly matching the MLP merger's 54.93 at 2×2 while using 55.6% fewer visual tokens, and its training wall-clock time is on par with the MLP merger.
  • The language experiments stop at 345M parameters, the zero-shot commonsense benchmarks are too short to activate the compressor at all, accuracy on the harder needle tasks still falls off sharply at 16K–32K, and removing key normalization collapses retrieval to 0.253.

arXiv:2610.31574 PDF

What Makes a Good Unmasking Policy? A Controlled Study of Learned Unmasking in Diffusion Language Models

Highlight HF pick · 1▲Large Language Models Jinya Sakurai, Xun Xu, Jaehong Yoon Masked diffusion language models decode tokens in parallel, and learned unmasking policies decide which predictions to commit. Existing policies differ in so many ways that it is unclear which choices matter. A controlled study finds that labeling any prediction that matches the final output as committable is too permissive at low numbers of function evaluations (NFE), that hidden states complement confidence scores, and that about a 10% NFE reduction offsets the overhead of policies up to 7.5M parameters. An 856K-parameter policy trained on trajectory-contraction targets achieves the best accuracy–NFE frontier and highest throughput on GSM8K with LLaDA-8B-Instruct, is competitive on MATH-500 and HumanEval, and largely transfers to Dream-7B-Instruct.

Masked diffusion language models can commit several tokens per step, but how fast they decode depends on the unmasking policy that decides which predictions to commit. Learned policies in prior work change several design choices at once. The authors treat supervised policies as distilling a teacher decoding procedure and test each design axis separately: what the teacher labels as committable, which states the policy trains on, what the policy observes, and how much the policy's overhead costs.

  • The common "terminal agreement" label marks any prediction that matches the final output as committable, and it turns out to be too permissive: 28.7% of its positives leave the reference trajectory when committed early, so the authors introduce trajectory contraction, which accepts a commitment only if it jumps directly to a later state of the reference trajectory.
  • Contraction improves the accuracy–NFE frontier mostly at low NFE (the number of backbone evaluations), and training on full reference trajectories adds an out-of-domain gain, reaching the 98% target on MBPP in 48.6 vs. 64.8 NFE. Backbone hidden states are weak as the only input, but combined with confidence they cut NFE at the 98% target on GSM8K from 52.9 to 42.1.
  • Latency profiling shows that a policy of up to 7.5M parameters adds about 1.5 ms per step, at most 10% overhead, so roughly a 10% NFE reduction pays for it regardless of prompt length or prefix caching.
  • The resulting 856K-parameter policy, OnTrack (about 0.01% of LLaDA-8B-Instruct), has the best accuracy–NFE frontier and the highest throughput at every accuracy target on GSM8K, for example 331 vs. 232 tokens/s for Fast-dLLM at 98% of reference accuracy without caching. It is competitive at moderate targets on MATH-500 and HumanEval, and its advantage largely carries over to Dream-7B-Instruct.
  • On MATH-500, accuracy levels off just below the reference, where training-free methods like Fast-dLLM and EB-Sampler still reach it at higher NFE. Prefix caching lowers accuracy by up to about 2 points, and the advantage disappears when a change in generation length is combined with the shift from math to code.

arXiv:2610.31576 PDF

Applications 87

ChemBART: Pre-training a Reaction Language Model for Comprehensive Synthesis Planning

Kenan Li, Yijian Zhang, Jin Wang, Haipeng Gan, Zeying Sun, Xiaoguang Lei et al. cross-listed Most language-model approaches to computer-aided synthesis planning (CASP) represent single molecules and handle one task at a time, such as predicting precursors. ChemBART is a BART-based model pre-trained with mask filling on whole reaction expressions (reactant, reagent, product). A single model then handles precursor and reagent generation, temperature and yield regression, and molecular property classification. Combined with Monte Carlo tree search (MCTS) as policy and value functions, it plans multi-step routes, and the proposed routes were validated in the wet lab, giving shorter pathways with more than 20% higher yield than established literature routes.

arXiv:2601.02915 PDF

Breaking the high-degree barrier in Lehmer's problem with a human-AI discovery loop

Yuri Chervonyi, Sergei Gukov, Thang Luong Lehmer's 1933 problem asks whether the Mahler measure of an integer polynomial can be made arbitrarily small, but small-measure irreducible polynomials had been constructed systematically only up to degree 180, and Boyd's one-parameter family skips nearly half of the higher degrees because it factors. The authors run a human-AI discovery loop in which an AI agent builds numerical and symbolic solvers, launches distributed searches, and works with a human researcher to form and verify hypotheses. The result is a five-parameter Generalized Boyd algorithm that produces irreducible polynomials with Mahler measure below Boyd's limit point for every target degree. Across all even degrees up to 2000, it improves six published values and holds 99.6% of the best known records, and the verified polynomials are released as a dataset.

arXiv:2610.30277 PDF

CiteAnchor: Source-Aware Grounding for Hallucinated Citation Detection and Resolution

Chenyi Tong, Ping Nie, Yu Zhang Hallucinated citations are hard to detect because real bibliographic records are often incomplete or differ across source types and publication versions, so a plausible-looking match proves little. The authors release CiteResolve, a benchmark of 423 real-world hallucinated citations annotated for whether the intended source can be recovered. They also propose CiteAnchor, which adapts retrieval and field-level checks to each source type and decides separately whether a citation is valid and whether a unique source can be found. On an existing detection benchmark, it reaches 0.841 F1 with GPT-5.6 Luna and 0.784 with an open-weight model. On CiteResolve, its end-to-end success is up to 65.3 percentage points above web-search baselines.

arXiv:2610.30355 PDF

The Fusion Tax: A Diagnostic Analysis of When Hybrid Retrieval Hurts

Aaditya Chauhan, Arun Agarwal cross-listed Hybrid retrieval, which fuses a lexical retriever such as BM25 with a dense neural retriever, is widely recommended because it improves average results, but this study measures what fusion costs on individual queries. The setup spans five dense encoders, SPLADE, a cross-encoder reranker, seven BEIR datasets and three fusion functions. Even the best method, dense-weighted score interpolation, ranks relevant documents below the dense-only retriever on 8-21% of queries while winning on average, a loss the authors call the fusion tax. Reciprocal rank fusion (RRF) pays a higher tax whenever the dense retriever is clearly stronger. Whether fusion will hurt cannot be reliably predicted across corpora, so the authors recommend deciding whether to fuse based on the measured balance between sparse and dense retrieval on the target corpus.

arXiv:2610.30612 PDF

Group Relative Policy Optimisation Improves Multilingual Speech Recognition in Low-Resource Languages

Tim Wenjie Hu, Evelyn Nafula Ouma, Benjamin Akera, John A. Quinn Encoder-decoder speech recognition (ASR) models trained with teacher forcing suffer from exposure bias, meaning they struggle to recover from their own mistakes at inference time. Earlier work showed that reinforcement learning with Group Relative Policy Optimisation (GRPO) helps LLM-based ASR in major languages. Here it is applied to the smaller non-LLM Whisper-large-v3 across 51 African languages, with the largest gains when training data is scarce or the audio is noisy. The resulting public model, SunflowerASR, sets the best published macro-averaged word error rate on AfriVox-v2 (0.227) and SimbaBench (0.372).

arXiv:2610.30663 PDF

Lean Finder V2: Semantic Search for Mathlib That Disambiguates User Queries

Jialin Lu, Xiaoyu Liu, Hengrui Luo, Rodrigo Stehling, Linxi Liu, Weiran Sun et al. Mathematicians and coding agents usually search the Mathlib library of formal mathematics for the theorem prover Lean with short, ambiguous queries such as a property, a notation, or an alias. Existing search engines assume detailed, context-rich queries instead. Lean Finder V2 treats this as a data problem: it uses adversarial training to generate large amounts of labeled short queries, then disambiguates queries both by fine-tuning the retriever and by rewriting queries at input time. On the human-curated MathlibQR benchmark it sets a new state of the art with at least 11% relative improvement over the previous best Mathlib search engines. When integrated with Claude Code, it also improves proof success on a downstream theorem-proving task.

arXiv:2610.30843 PDF

ReDence: A Hybrid Agentic and Decision-Model Framework for Rapid, Evidence-Grounded Detection of Modern Social Engineering Scams

Jaishnoor Kaur, Josef Hor\'a\v{c}ek, Ian Waskom, Sayak Saha Roy cross-listed Detecting social engineering scams often means walking through website workflows, checking claims against outside sources and linking evidence across pages, work that analysts still do largely by hand. ReDence is an open-source framework built on two locally hosted models: a fine-tuned multimodal agent (Qwen3.5-4B) that plans the investigation and writes reports, and a specialized decision model (Tev1) that handles discrete choices such as browser actions and when to stop. Delegating these decisions avoids lengthy language-model responses. The system reaches 96% detection accuracy across twelve scam categories while running at least four times faster than comparable agentic systems, and its structured evidence reports let analysts audit each verdict.

arXiv:2610.30963 PDF

Energy and Latency First: How Deployment Constraints Reverse Time-Series Model Selection

Shruti Bothe, Illyyne Saffar, Ayush K. Varshney Models are usually ranked by accuracy and then deployed under latency and energy limits that the ranking ignores. The authors study cellular handover prediction on user devices, where each inference must finish within a 40 ms measurement period and energy is a lifetime budget. They screen eight architectures for tail-latency feasibility and then rank them on measured energy, including the cost of computing input features. Gradient-boosted trees on handcrafted features are both the most accurate (F1 0.974) and among the cheapest, whereas the CNN-1D that an accuracy-only benchmark would favor uses more than three times as much energy per inference. The authors state three conditions under which this reversal should be expected in other domains.

arXiv:2610.30965 PDF

SecLog-Bench: A Scalable Benchmark for Cybersecurity Log Reasoning with Scaffolded Evidence

Yifan Wang, Dongliang Xu, Xiaoyong Li, Jing Li, Peter S. Pan, Yue Yao cross-listed Existing datasets for analyzing cybersecurity logs with large language models (LLMs) are small, draw evidence from a single source, and label only the attack type without annotating the reasoning behind it. SecLog-Bench addresses this with an extensible simulation platform covering 29 attack scenarios. The platform generates logs from multiple sources (audit logs, application logs, and network traffic) labeled with attack stages, and each sample is annotated with essential context and a structured step-by-step evidence chain. Across eight mainstream LLMs asked to reason about how far an attack has progressed, the best model reaches only 44.44% accuracy and 45.07% Macro-F1 even when given scaffolded evidence. The best Stage Reasoning F1 is 54.87%, and correct predictions often rest on flawed reasoning.

arXiv:2610.31000 PDF

Trion-1: Unifying Language, Protein Sequence, and Structure with Diffusion Language Models

Tianyu Lu, Kaicheng Yang, Jiangtao Feng, Keyue Qiu, Yawen Ouyang, Yixin Wu et al. The goal is a single protein foundation model that can understand, generate and translate between protein sequence, 3D structure and natural language while keeping the instruction-following of general LLMs. Trion-1 is a multimodal diffusion language model (dLLM) that puts sequence, structure and text into one discrete token space. Instead of strict left-to-right generation, its tokens interact bidirectionally, which suits modeling long-range residue contacts. One model covers folding, inverse folding, sequence and structure understanding, and sequence and structure design. In folding it reaches a TM-score above 0.8 from a single sample under strict homology control, matching or beating an existing tri-modal model that uses 200 samples, and its sequence understanding is on par with Claude Opus 4.7.

arXiv:2610.31215 PDF

The Value of System 1 Models in Medicine

Xiufeng Yang, Konstantinos Vrettos, Zeineb Ben Chaaben, Fabian Drexel, Jannik L\"ubberstedt, Johannes Moll et al. Many clinical decisions are fast choices among a few fixed options, and they need confidence estimates so the system knows when to defer to a clinician. The authors evaluate four "System 1" models, which answer zero-shot with a probability for each option, on 86 public tasks across 14 clinical domains and 20 languages, totaling 179,111 cases. The best model per task reaches a mean area under the receiver-operating-characteristic curve (AUROC) of 0.94 on medical knowledge but only 0.69 on volumetric imaging findings, and imaging tasks score 0.09 AUROC lower than text tasks. Triage nurses outperform the best model, accuracy falls by up to 39 points from English to Yoruba, and generative models from the same families are 11 to 76 times slower for only 3 to 5 points more balanced accuracy.

arXiv:2610.31315 PDF

Auditing the Hand-Written Inputs of Learned Access Control

Arsalan Jalil Sarghaleh, Mehdi Hazrati cross-listed Learned access control systems usually depend on two hand-built inputs: a hard-deny set enforced by a trusted shield, and a region map that decides which requests are answered by mined rules and which by the learner. Auditing both on the five public ABAC Lab policies (ABAC is attribute-based access control), the authors find that deny clauses faithfully transcribed from the written policy are unsound at rates of 0.50-0.70, whether written by a human or by language models. This happens because a restriction stated in one section is overridden by grants from other rules. Shipped unchecked, these deny sets would withhold 16-78% of legitimate grants. The study also shows that standard rule miners do not produce the rule sets that hybrid routing needs, and that comparisons between hybrid and learner-only systems can reverse depending on which loss is measured.

arXiv:2610.31388 PDF

Social-HUMINT: An LLM-Augmented Unified Framework for Measuring Social Media Exposure

Pellegrino Casoria, Simon Pietro Romano cross-listed LLMs make it cheap to profile people from their public social-media posts, including attributes they never disclosed, which raises their exposure to social engineering. The authors unify Social Media Intelligence (SOCMINT) and Cyber Human Intelligence (Cyber-HUMINT) into a taxonomy of 38 capabilities validated by experts (Fleiss' κ = 0.92). They implement its core extraction, inference, and susceptibility steps as an LLM-augmented multi-agent pipeline. On real participants' public data, extraction reaches 97.3% F1, inferred attributes score 4.73/5, and the pipeline predicts targets' engagement preferences significantly above chance, which lets defenders assess their own exposure.

arXiv:2610.31573 PDF

Cost-Aware Abstention for LLM-Based IT Ticket Classification

Sachin Gupta Routing an IT support ticket to the wrong team can cost more than having a human triage it, so the authors evaluate cost-aware abstention: a routing policy is selected on calibration data under explicit costs, then tested once on held-out data. On a public corpus of 47,837 tickets in eight classes, they compare four trained classifiers with four LLMs under zero- and few-shot prompting. TF-IDF with logistic regression and a calibrated 0.60 threshold cut test cost by 38.1% and reduced wrong routes from 15.3% to 3.5%, while the best prompted LLM reached only 0.477 accuracy against 0.879 for fine-tuned DistilBERT. Recalibration lowered LLMs' calibration error without improving their ability to separate right from wrong answers, so the authors recommend reporting correctness AUROC and risk-coverage alongside expected calibration error.

arXiv:2610.31657 PDF

73 more specialized papers

Large Language Models 69

When Multiple Choices Are Correct: Asymmetric Probability Allocation in Jev and Other Decision Models

Rushab Munot cross-listed Decision models return normalized probabilities over candidate choices, but being calibrated on single-answer tasks does not show that those probabilities are reliable when several choices are correct. The authors evaluate Jev, Liquid d1, and GPT-6 Luna Decisions on MetaTool, HotpotQA, and ECtHR. Even when a model ranks every correct choice above every distractor, in 28% to 50% of those lists at least one correct choice gets less than the uniform probability 1/n. Given five identical copies of a correct choice, the models assign them probabilities with mean gaps of 17 to 79 percentage points, so a probability threshold can rank correct choices but cannot reliably recover the full set.

arXiv:2610.29217 PDF

FusionEval: Benchmarking Model Fusion for Large Language Models

Shuo Cai, Yanggan Gu, Yuhang Liu, Su Lu, Yuanyi Wang, Yang Yu et al. Model fusion merges several domain-expert models into one, but existing evaluations rarely cover long chain-of-thought reasoning or tool use, and strong average scores can hide the loss of individual capabilities. FusionEval provides controlled pools of Qwen3 and Llama experts in math, code, science, instruction following, and tool use, and evaluates fusion methods on eight benchmarks for average and worst-case retention, runtime, and peak GPU memory. Training-free fusion methods preserve reasoning but substantially degrade tool use, while MOPD keeps tool use and gives a more balanced profile at the cost of extra data preparation, inference, and training.

arXiv:2610.30162 PDF

MoEZip: Routing-Aware KV Cache Compression for Sparse Mixture-of-Experts LLMs

Minsang Kim, Seungjun Baek In long-context inference with sparse Mixture-of-Experts (MoE) models, the key-value (KV) cache becomes the memory bottleneck, and existing compression methods built for dense transformers ignore that evicting cache entries can change which experts get selected. MoEZip scores context tokens by combining attention weights with a new metric, based on the Fisher information of the routing distributions, that estimates how sensitive expert routing is to perturbation, and it composes this with dense-model eviction scores. At 10% KV retention, it beats the strongest baseline by 24 and 20 points on average on LongBench and RULER, which corresponds to 2.3x and 4.9x higher scores.

arXiv:2610.30206 PDF

How Prompt Style Affects AI Accuracy and Speed in Competitive Programming

Saad B. Alarifi, Mohammed A. Alam LLM outputs on competitive programming depend on how the prompt is phrased, but earlier comparisons looked only at accuracy and not at runtime speed or human baselines. This exploratory study tests one prompt in each of nine styles on GPT-5.5, using three accuracy tasks from the International Olympiad in Informatics (IOI) and two NP-hard speed tasks. No single prompt style was best: the winner depended on the task type and on whether results were summarized by mean or median, and many mean differences came from a few outlier runs. Compared descriptively with human IOI contestants, who worked under different conditions, the model beat the average contestant on one task and fell short on the other two.

arXiv:2610.30223 PDF

High-Value Training Data from Noisy, Diverse Sources

Hedi Driss, Evan Frick, Anastasios N. Angelopoulos, Cho-Jui Hsieh, Yejin Huh, Tijana Zrnic et al. cross-listed Training on large, heterogeneous datasets of mixed quality raises the question of how much weight each example should get. The method uses metagradients taken through the training process to optimize per-example weights against performance on a small, high-quality target dataset. It also learns parametric weighting rules that generalize to new and streaming examples and transfer from small models to much larger ones. In reward modeling with noisy, adversarial, synthetic or mismatched preference data, the learned weights improve held-out performance and adapt reward models to specialized domains. They also let mixtures of human and synthetic supervision outperform either source alone.

arXiv:2610.30421 PDF

FIGFLEX: A Mixed-Precision FP-INT GEMM Accelerator Co-Designed with Low-Bit LLM Quantization

Faraz Tahmasebi, Haocheng Xu, Andrew Trauger, Yomal Perera, Zhenyu Tang, Nael Chowdhury et al. cross-listed Weight-only quantization makes floating-point-by-integer (FP-INT) matrix multiplication the main LLM workload, but GPUs and existing accelerators gain no compute savings from fewer weight bits. FIGFLEX is a co-designed accelerator and quantizer. Its shift-add, radix-4 datapath scales throughput as bit-width falls, and its mixed-precision quantizer assigns each channel group the bit-width needed to cover the error that GPTQ cannot compensate. On LLaMA-2-7B it reaches 5.85 and 8.77 WikiText-2 perplexity at 3- and 2-bit average precision, beating GPTQ, AWQ, and QuIP#, with 1.56x and 1.42x higher throughput per area than the AxCore and FIGLUT accelerators.

arXiv:2610.30480 PDF

Behind Aggregate Scores: What Systematically Varying Arena-Hard's LLM Judge Reveals

Louise Yorke The authors test how repeatable LLM-judge verdicts are and how much they depend on the judge's instructions. They used GPT-4.1 as judge on 500 Arena-Hard v2.0 prompts with fixed answers, across eight instruction variants, each rerun at temperature zero on a pinned model version. Rerunning the default setup unchanged moved the aggregate score by only about one point but flipped the win/tie/loss outcome on 21.2% of prompts. Some instruction changes altered far more verdicts while leaving the aggregate nearly unchanged. The authors recommend repeating each judge setup and reporting prompt-level agreement alongside aggregate scores, especially when verdicts are used as preference labels.

arXiv:2610.30512 PDF

Large language models show systematic failures in incorporating evolving medical knowledge

Min Zeng, Zaifu Zhan, Kai Yu, Rui Zhang ChronoEvo contains 1,709 recommendation-level revisions to clinical guidelines, spanning 12 kinds of medical knowledge change, and separates recognizing a revision from actually applying it. In the primary model, 90.8% of responses recognized the revision but only 58.5% fully incorporated it, a 32.3-percentage-point gap that held across 14 models and in a review of 300 responses by two clinicians. Changes to recommendation strength and patient eligibility were the most likely to leave outdated guidance in place. Making the revision more prominent in the prompt reduced outdated content but did not significantly improve full incorporation.

arXiv:2610.30631 PDF

Compiling Autoregressive LLM Control into Policy-Exact Parallel Inference for Hierarchical RAG

Zehao Lin, Qianfeng Shen, Xixuan Hao, Shaobo Cui, Chunyu Li, Kai Chen et al. When an LLM acts as a controller in hierarchical retrieval-augmented generation (RAG), it chooses which document sections to explore through either unreliable free-form JSON or slow serial grammar-constrained decoding. For control decisions with a bounded number of steps, the possible action sequences form a shallow prefix tree. That tree can be evaluated with one prompt prefill plus one tree-causal probe pass, and a lemma shows this reproduces the serial decoding policy exactly in exact arithmetic while generating no output tokens. One more pass certifies the most likely set of choices. In experiments, compiled control matches the serial policy on 84 to 89% of questions, with every observed divergence a floating-point near-tie, and cuts control latency by 35 to 47% compared with token-by-token decoding.

arXiv:2610.30664 PDF

ULME: Predictive Coupling for Mutually Reinforcing Embedding and Language Modeling

Yang Liu, Zhiqiang Yuan, Shu-Dong Huang, Jiancheng Lv, Liwei Yang, Qiwei Yan et al. Retrieval-augmented and agent systems usually run separate models for text embedding and text generation. Existing unified models such as GritLM share weights between the two but do not let either capability shape the other. ULME inserts learnable slots between query and answer in a decoder-only LLM and trains three losses on the shared backbone: a contrastive loss for a semantic embedding, a language-modeling loss for the answer, and a stop-gradient loss that aligns a reasoning embedding with a pooled summary of the answer, so that generation shapes what the embedding encodes. Across Qwen2.5 and Qwen3 models from 0.5B to 8B parameters, a single checkpoint improves MTEB by 3.3 to 4.1 points over the strongest open baseline, while matching a generation-only fine-tune on MMLU, GSM8K, and HumanEval.

arXiv:2610.30705 PDF

Random Remasking Scales Masked Diffusion Inference

Zhaoyi Zhou, Frank Wu, Andrea Zanette Masked diffusion models (MDMs) generate text by progressively unmasking tokens, and errors made during decoding are normally irreversible. The authors recast MDM inference as a Markov chain over clean sequences that randomly re-masks and then denoises, which motivates Random Remask-Denoise (R²D), a test-time procedure that needs no fine-tuning and no reward. They also propose Trajectory Correction Training (TCT), which trains the denoiser on its own errors along the generation trajectory. On logic tasks the approach shows near-log-linear test-time scaling, exceeding 97% accuracy on hard Sudoku and 99% on 3-SAT, and on pretrained models it beats iterative-correction baselines and the official Q-mode accuracy of LLaDA2.1-mini on HumanEval+.

arXiv:2610.30744 PDF

Abstract Priorities, Concrete Rewards: How Stated Preferences Relate to What People Reward in AI Responses

Zihan Chen, Lei Nico Zheng, Di Zhu cross-listed AI assistants are increasingly personalized to users' stated preferences, but it is unclear whether those preferences match what people actually reward. Using the public PRISM dataset, in which 1,396 participants ranked seven response attributes and then rated 29,347 responses from 21 LLMs, the authors find that features the survey never asked about, such as length, formatting, and refusal language, were at least as predictive of ratings as the surveyed attributes. Stated priorities tracked perceived model strengths but not measurable text features, while written feedback such as asking for shorter answers did predict how much length mattered. Personal information improved prediction of held-out choices by at most 0.13 percentage points, versus 1.11 for a stronger pooled model.

arXiv:2610.30763 PDF

Scaling Laws for Generative Reward Models

Angel Raychev, Nathan Lile, Rafael Rafailov Generative reward models (GenRMs) act as judges that supply preferences during policy training. The authors study them for creative writing across Qwen3 judge sizes from 0.6B to 14B, comparing Thinking judges trained with GRPO against Answer-Only judges trained with supervised fine-tuning. Thinking judges score higher on static gold-label accuracy, but Answer-Only judges produce better policies through online DPO in seven of eight matched settings, and earlier, less accurate Thinking checkpoints can be better rewarders than later ones. Larger judges help substantially, even when they exceed the policy's size, and the results show that static judging accuracy does not determine downstream optimization quality in these pipelines.

arXiv:2610.30765 PDF

ProQuant: Preserving Expert Quality over Quantity in Ultra-Low-Bit Quantization of Mixture-of-Experts Models

Jiaqi Chen, Xianglong Yan, Chengzhu Bao, Xie Chen, Yulun Zhang Mixture-of-Experts (MoE) language models activate only a few experts per token but still need a lot of storage, and mixed-precision post-training quantization (PTQ) struggles to keep output quality at very low bit budgets. The authors point to two causes: existing methods ignore how quantization errors propagate across layers, and fine-tuning the router directly can shift which experts get selected. ProQuant addresses these with Prune-or-Prioritize (PoP), which jointly decides which experts to prune and how many bits each remaining expert gets, so that higher-precision experts can stand in for badly degraded ones. It adds Routing-Consistent RMSNorm Finetune (RoCo), which compensates for quantization error through mathematically equivalent transformations that keep routing stable, and the method outperforms existing quantization methods across diverse generative tasks.

arXiv:2610.30779 PDF

Gandalf: Flash Shall Not Die! Increasing Flash Storage Lifetime for KV Cache Offloading in LLM Serving

Maharani A. P. Irawan, Kuntai Du, Jiayi Yao, Dongjoo Seo, Yuyang Huang, Haryadi S. Gunawi et al. LLM serving systems increasingly offload the key-value (KV) cache to SSDs, but multi-turn chat and agent workloads write so much KV data that SSD lifetime can drop from years to months. Filtering writes by past access patterns alone lowers cache hit rates, because KV caches from different attention types can have similar histories but different future reuse. Gandalf also uses application-level information and gives full-attention KV caches priority for SSD admission over caches from other attention types. Across workload traces and hardware capacities, this extends SSD lifetime by 7.3x on average with negligible effect on cache hit rate.

arXiv:2610.30791 PDF

Joules-to-GDP: How Much Economic Value Do AI Systems Produce per Unit of Energy?

Amirreza Zeinali, Jon Saad-Falcon, Avanika Narayan, Mersad Abbasi, Jonathan S. Hartley, John Hennessy et al. The authors measure how much economic value AI systems produce per unit of energy. They define Net Value per Joule (NVJ), the dollar value of the output minus execution cost divided by energy, and Time Saving per Joule (TSJ), expert time saved divided by energy. They evaluate 220 GDPval tasks across 44 occupations using five open-weight models served locally on four accelerator types. Occupations with the highest value per task are not always the most energy-efficient: engineering yields more net value per task, but software development yields about 27% more per joule. GLM 5.2 delivers the most net value per task, while Gemma 4 31B achieves about 2.5 times the NVJ using 23% of the energy. Hardware choice changes Qwen3.6-27B's NVJ by a factor of 4.32, with B200 leading.

arXiv:2610.30858 PDF

Silicon Samples Match the Answers but Not the Respondents

Yusuke Kumagae, Masahiro Kaneko Silicon sampling has a large language model (LLM) answer a survey while posing as a person with given demographic attributes, and the synthetic respondents it produces are increasingly used in place of human ones. Standard checks look at individual answers, per-question answer distributions and correlations between questions, so a generator can pass them while producing answer combinations that no real person gives. The authors propose the respondent coverage gap (RCG), which treats each respondent's full answer profile as a point and measures how far each human is from the nearest synthetic respondent, relative to the distance between human samples. On two US social surveys, all six LLMs tested leave human respondents uncovered, and in every prompting and inference setting tried, every LLM has a higher RCG than a Gaussian copula fitted to the human data. Under the main settings they also do worse than a baseline that samples each answer independently from the human marginals.

arXiv:2610.30900 PDF

Same molecule, Different Knowledge: Cross-Identifier Binding Failures in LLMs For chemistry

Khouloud Saadi, Di Wang Chemistry workflows often refer to the same molecule by a trivial name, an IUPAC name or a SMILES string, assuming that an LLM can reach the same knowledge through any of them. A controlled benchmark covering models from 7B to 72B parameters, including chemistry-specialized ones, shows that different identifiers for the same molecule consistently expose different subsets of the model's knowledge. The failure persists even when the model can identify a molecule from its SMILES and retrieve a fact through its name, and behavioral, representational and parameter-editing analyses locate it in the binding between the identifier and knowledge the model already has. The authors call this cross-identifier binding failure and propose Binding Self-Distillation, which uses lexical identity during training so that facts can be retrieved directly from SMILES without a name lookup or retraining from scratch.

arXiv:2610.30903 PDF

DConstrain: Compiling Context-Parameterized Transitions for Diffusion Language Model Constrained Decoding

Shidong Cao, Jing Ma, Wenhao Chai, Yuchen Tian, Lun Du, Da Zheng Diffusion language models (DLMs) generate several tokens in parallel, so tokens that are each valid on their own can together violate a schema such as a JSON function-call format. Exact constrained decoders for DLMs have had to fix the order of object properties to keep their compiled state machines from growing exponentially. DConstrain instead compiles schemas into context-parameterized transition programs, tracks field histories at runtime, builds exact successor graphs in parallel with the model's forward pass, and groups tokens that have identical effects, with proofs that these steps preserve the optimal output and the output distribution. It compiles 100% of BFCL schemas under arbitrary property order, compared with 71.30% to 85.81% for an optimized finite-state baseline, cuts end-to-end decoding latency by 6.65 to 46.26 times with fully valid outputs, and raises DiffusionGemma's BFCL accuracy from 37.27% to 81.56%.

arXiv:2610.30920 PDF

Tracing Hallucinations in LLM Reasoning for UML Class Diagram Generation: An Empirical Taxonomy and Mitigation Study

Mojdeh Rahmanian, Foutse Khomh, Ashkan Sami, Yanchao Yu cross-listed Reasoning LLMs that generate UML class diagrams from requirements can hallucinate, but evaluations usually examine only the final diagram. Using DeepSeek-R1 and DeepSeek-V4-Pro reasoning traces, the authors build a taxonomy of eleven hallucination types grouped into three dimensions: Prompt Faithfulness (the reasoning departs from the requirements), Code Faithfulness (the generated PlantUML code departs from the reasoning) and Reasoning Quality (problems within the reasoning itself). Code Faithfulness errors were most common for DeepSeek-R1 and Prompt Faithfulness errors for DeepSeek-V4-Pro, and longer traces were not associated with more hallucinations. In a mitigation test with three Qwen3-based models on Golden UML ModelSet cases, adding Prompt Faithfulness and Reasoning Quality constraints to the prompt raised pooled mean F1 by 4.0 percentage points, while a post-hoc review of code faithfulness added no consistent benefit.

arXiv:2610.30946 PDF

IQuest-Q1: Advancing Foundation Capabilities for Agentic CLI Systems

A. Jian Yang, An Huang, Bo Dan, Chen Mu, Chengran Yang, Chuan Shi et al. IQuest-Q1 is an open-source sparse Mixture-of-Experts (MoE) foundation model built for agentic command-line coding, with 320B total parameters and 15B active. Pre-training combines Stable SplitNorm, the Muon optimizer and muP with a curriculum whose second stage increases the share of coding data. Mid-training broadens the data mixture and extends the context length. Post-training mixes supervised fine-tuning, domain-specialized reinforcement learning and MOPD across several agent scaffolds, and model merging then combines experts and checkpoints from different stages while limiting conflicts between their parameter updates.

arXiv:2610.30949 PDF

OpenAttention: One Percent Is Enough with Tail Proxies

Hyunwoo Yu, Jongbeom Lee, Hyunbin Kim, Sungjun Park, Jaemin Choi, Jincheol Yang et al. Sparse attention reads only a small part of the key-value cache, and the open question is what to do with the unread tail. Discarding it biases the output, and sampling it breaks down at read budgets of 5% and below. The authors show, both empirically and in closed form, that a single proxy token holding the count-aware mean of the tail beats sampling. They build OpenAttention on this idea, a training-free pipeline that combines query-geometry clustering and selection with an error-based allocation step called ErrorGate. On RULER-32K HARD with Llama-3.1-8B it retains 99% of dense accuracy while reading 1% of the cache, and its custom kernel runs decode attention 10× faster than dense attention.

arXiv:2610.30951 PDF

Joint or Per-Option? Compact Encoders for Closed-Set Decisions from 17 Million to 1 Billion Parameters

Tardelli Ronan Coelho Stekel Many language-model pipelines end in a closed-set choice such as picking a tool, and prompting a large model for it is slow and sensitive to option order. The authors study compact encoders from 17M to 1B parameters and compare reading all options jointly with scoring each option separately. Per-option scoring wins by 5.8 points of chance-corrected skill, but only when the encoder starts from a reranker initialization, and it needs 5 to 13 times more encoder passes. The released joint encoders at 68M and 150M parameters beat a 435M DeBERTa-v3 zero-shot classifier on System1-Arena tasks, and the smallest model answers a median request in 14 ms on CPU versus 728 ms for that classifier.

arXiv:2610.30970 PDF

A Reality Check of Inference-Time Context Augmentation Methods

Haeun Yu, Geert Heyman, Pepa Atanasova, Isabelle Augenstein Inference-time context augmentation methods change the input to help language models use their context, either by inserting filler tokens that buy extra computation or by adding content such as elaborations. Until now each method has been judged in isolation and only by accuracy gains. The authors compare five such methods across seven models of different sizes and three datasets. They find that many apparent gains are undirected perturbations that fix some predictions while breaking others, especially for methods that add no new content, which mainly flip predictions near the decision boundary. They recommend checking how reliable a gain is before adopting such methods.

arXiv:2610.30983 PDF

Padding to k: A Minimum Count in the Schema Makes Open-Weight Models Pad Their Extraction Lists

Abhishek Barali, Aayush Paudel Extraction pipelines often ask language models for JSON that follows a schema, and the schema can set a minimum number of list entries. In a preregistered study of eight open-weight models on 558 real documents, the authors find that models treat that minimum as a target. Showing an unenforced minimum raises the share of over-long lists from 34.5% to 55.4%, and about half of the padded entries are real text from the document that passes source-occurrence checks. Common fixes help only partly, and re-prompting a model with the validation error makes padding much more common.

arXiv:2610.30997 PDF

Mostly Reading: What Open-Book Scores Measure in Document-Level LoRA Adapters

Martino M. L. Pulici, Cuong Xuan Chu, Evgeny Kharlamov, Volker Tresp Single-document LoRA adapters have been proposed as mergeable, retrievable units of knowledge, but open-book evaluation (with the document in the prompt) cannot tell whether an adapter holds a document's content or simply reads it well. With one adapter per Wikipedia article for 150 articles on Llama-3.1-8B-Instruct, a random adapter trained on a different document recovers 82% of a specialist's open-book advantage but only 2% of its advantage on a closed-book cloze probe. Content does reach the weights, but selecting adapters by weight-space proximity or text embedding is indistinguishable from random selection. Merges of 16 or more adapters keep most of the open-book gain but little of the cloze gain. The authors also document four silent errors that each produced clean but wrong results.

arXiv:2610.31017 PDF

Chunk Recomputation under Real Document Revisions: Measuring Update Cost and Retrieval Quality of Fixed, Recursive, and Content-Defined Chunking

Yuning Bian, Lele Xu, Yifan Yuan, Xinyu Yin, Chengyan Zhan Chunking strategies for retrieval-augmented generation are usually evaluated on static snapshots, yet when a document is edited, shifted chunk boundaries force unchanged text to be re-embedded. Using 120 pairs of adjacent Chinese and English Wikipedia revisions, the authors measure this update cost for fixed-size, recursive paragraph-first, and content-defined chunking. Fixed-size chunking recomputes 57-68% of the new version's embedding workload, compared with 38-56% for recursive and 43-54% for content-defined chunking, while retrieval recall differs by under two percentage points. Fixed-size chunking is the most expensive in every length-matched setting, and the ranking between the two content-aware chunkers shifts with chunk-length parameters.

arXiv:2610.31019 PDF

Directing to Diversify: Multi-Directional Search for Automatic Heuristic Design

Jianghan Zhu, Jianan Zhou, Fei Liu, Ni Zhang, Yaoxin Wu, Zhuoyi Lin et al. LLM-driven automatic heuristic design for NP-hard combinatorial optimization usually relies on evolutionary operators for diversity and leaves the direction in which heuristics change uncontrolled. PRS (Perturbation-Rewrite Search) transforms heuristic ideas along three explicit directions: expansion toward richer designs, compression toward simpler ones, and steering for fine-grained refinement. It runs these within a breadth-first search with adaptive operator control to avoid premature convergence. Across diverse combinatorial optimization tasks, PRS consistently discovers high-quality heuristics.

arXiv:2610.31048 PDF

Diagnosing the Explicit-to-Applied Theory of Mind Gap in LLMs

Prajvi Saxena, Sabine Janzen, Wolfgang Maass Aggregate Theory of Mind (ToM) benchmark scores mix explicit ToM (recognizing what someone believes) with applied ToM (using that belief to predict behavior). On SimpleToM, Llama-3.1-8B recognizes mental states 99.4% of the time but predicts behavior only 28.9% of the time. The authors compare structured prompting, in-context learning, and a proposed Belief-State Test-Time Training (BSTTT) across Qwen2.5-7B, Llama-3.1-8B, and GPT-oss-20B. Reminders plus BSTTT improve Llama's behavior prediction by 40.8 points, most of it from the reminders. The authors recommend reporting explicit and applied accuracy separately.

arXiv:2610.31060 PDF

RIFT: KV Cache Compression for Draft Models with Complementary Compressibility Training

Dor Tsur, Maor Ashkenazi, Maximilian Golub, Peter Hicks, Yaniv Galron, Talor Abramovich et al. In speculative decoding with long outputs, the draft model's key-value (KV) cache keeps growing. Compression methods that select entries based on the prompt (the prefill) gradually become less relevant as generation goes on, a problem the authors call 'decode staleness'. RIFT is a family of selection-based KV cache compression strategies that balance prefill information against recently generated tokens. It refreshes the selection during decoding, uses an adaptive schedule to cut the cost of those refreshes, and adds a training phase with sparse attention masks so the drafter works better under compression. With Qwen3.5-4B drafting for Qwen3.5-122B on 32K-token inputs and 4K-token outputs, an 8K-token cache budget keeps over 98% of the full-cache acceptance length.

arXiv:2610.31098 PDF

SetCommit: Coordinated Token Commitments for Few-Step Diffusion Language Model Decoding

Yuannuo Feng, Yubing Ye, Wenshuai Yao, Yizhe Chen, Ngai Wong, Wenyong Zhou et al. Diffusion language models generate many token positions in parallel. With only a few decoding steps, tokens that each look reliable can turn out to be inconsistent with one another, and committing to them too early blocks later refinement. SetCommit is a training-free method that chooses a whole set of tokens to commit at each step, balancing each token's reliability against the risk of committing them together. It also swaps out weak earlier commitments when new context supports better alternatives, without changing the commitment budget. With LLaDA-8B and Dream-7B on GSM8K, MATH-500, MBPP and HumanEval, it improves accuracy or pass@1 by up to 7.4 percentage points at 32 model evaluations and gives geometric-mean speedups of 1.44 to 1.74 times at matched quality.

arXiv:2610.31120 PDF

Mixing Matters: Evidence Position Bias Across Sequence Mixers in Long-Context Question Answering

Aman Behera, Namit Solanki, Mehul Anand Long-context models often make poorer use of answer-bearing evidence when it sits away from the edges of the prompt. The study compares this position bias across Transformer, state-space and hybrid Mamba models by moving one gold document through a fixed set of nine distractors. The tested Transformers favor evidence at the beginning of the context while the state-space models do not, though both families favor evidence near the end. The authors also find that prompt changes can create or nearly remove the beginning advantage, so they conclude that effective context use depends on architecture family, capability, task and prompt together rather than on the sequence mixer alone.

arXiv:2610.31135 PDF

Prompting Strategy Effects Depend on the Judge: A Multilingual Study of LLM-as-a-Judge Robustness

Asal Mehrabi, Mohammad Aliannejadi, Clemencia Siro Prompting strategies for LLM-as-a-Judge evaluation, such as Chain-of-Thought and role conditioning, are usually recommended based on English-only benchmarks. The authors test whether these strategies transfer using 9 strategies, 4 open-weight judges, 27 languages and 4 tasks across 13 datasets, producing over 1 million scores validated against human judgments. The choice of judge model matters more than the choice of prompting strategy, and a strategy's effect can reverse direction from one judge to another. Stability also drops in low-resource languages in ways that aggregate scores hide, so the authors advise validating prompts for each intended evaluation setting.

arXiv:2610.31172 PDF

Different Tasks, Different Timesteps: Temporal Task-Aware Quantization for Diffusion Language Models

Ofek Avivi, Eliron Rahimi, Avi Mendelson, Chaim Baskin Diffusion language models (DLMs) produce text through repeated denoising steps, which makes quantization harder than for standard autoregressive models. The authors show that sensitivity to quantization varies by both task and denoising step, and that this pattern can be predicted without ground-truth labels. Their T2Quant framework uses that label-free signal to allocate activation precision across denoising steps under a fixed bit-width budget. Across three DLMs and six benchmarks it improves existing quantization methods by up to 6.7 percentage points in average accuracy at the same budget.

arXiv:2610.31198 PDF

Memory-Aware Reward Modeling: Learning Human Preferences for LLM Memory Use

Yijiang River Dong, Zexue He, Cheng Yu, Tiancheng Hu, Michael J Ryan, Jan Batzner et al. LLM assistants increasingly use memory of past sessions to personalize responses, but existing evaluations do not check whether users actually prefer memory-based answers. About 1,000 pairwise judgments from 87 participants, using their own ChatGPT queries and memories, show that memory-conditioned responses are not consistently preferred, and sometimes users strongly prefer responses generated without memory. Automatic evaluators favor memory-conditioned responses much more often than users do. A memory-aware reward model trained on the user judgments predicts preferences with 72.8% accuracy, beating existing baselines.

arXiv:2610.31272 PDF

Cola2: Co-Adaptive Hierarchical Language Modeling with Continuous Semantic Compression and Balanced Discrete Generation

Hongcan Guo, Yian Zhao, Feng Wang, Shen Nie, Guoqiang Wei, Yan Zeng et al. Cola2 is a hierarchical language model that separates continuous semantic modeling from discrete text generation. It segments text into semantic spans, compresses each span into a continuous latent, and models those latents with a Flow Matching prior. A block-causal masked diffusion decoder then turns the latents into text, with classifier-free guidance controlling how closely it follows the latents versus generating on its own. Trained on 346 billion source tokens, an unconverged checkpoint averages 60.10 across eight tasks and beats Llama-3-8B on HellaSwag and StoryCloze, even though Llama's reported pretraining budget is over 43 times larger.

arXiv:2610.31305 PDF

Extreme Compression of Per-Layer Embeddings via Cross-Layer Latent Vector Quantization

Jiacheng Liu, Yaxin Luo, Xiaohan Zhao, Peng Sun, Xinyi Shang, Jiacheng Cui et al. Per-Layer Embeddings (PLE) make up 35.2% of the parameters of Gemma 4 E4B, and at very high compression rates conventional quantizers can give each token-layer vector only a few bytes. OPTIC instead stores one quantized code tuple per token and decodes it at every layer with a layer-specific affine decoder, fitting both the codes and the decoders to the PLE output path rather than to the stored vectors. At about 100× PLE compression, OPTIC retains 97.8% of the uncompressed model's performance across seven benchmarks, compared with 66.3% for product quantization, the strongest of six baselines. It still retains 84.1% at 1,002× compression, and at 100× it cuts GPU memory by 29.7% while running at most 0.40% slower.

arXiv:2610.31309 PDF

When LLMs Forget What They Believe: Characterizing State Integrity Collapse in Long-Horizon Dialogues

Yancheng Zhu, Yifan Yuan Large language models can repeat an instruction they were given while no longer following it. The authors call this State Integrity Collapse and measure it with StateTrack-Bench, a set of controlled conversations covering numerical, format, persona, factual-memory and logical-rule constraints. Across seven pinned model snapshots, strict retention ranges from 62.4% to 100%, and a recall audit shows that 88.6% of failures are behavioral drift rather than forgetting: the model can still state the constraint but has stopped applying it. A mitigation called Structured State Anchoring recovers up to 24.4 points on immutable constraints.

arXiv:2610.31314 PDF

ValuSpec: Plug-and-Play Candidate Valuation before Target Verification for Tree-Based Speculative Decoding

Liang He, SiYuan Ma, Qishi Zhan, Yongqi Fan, Mingyu Cao, Qizhen Lan et al. Tree-based speculative decoding speeds up large language model inference by verifying several candidate branches in parallel, but retrieval-based drafting builds large trees full of low-quality branches that waste the target model's compute. ValuSpec is a plug-in that scores and prunes candidate branches before the target model verifies them, with a filtering threshold calibrated on the target model's greedy outputs and no parameter training. Because verification still uses the greedy rule, outputs stay identical to standard decoding. On HumanEval, GSM8K, Dolly and SpecBench, it reaches up to 2.38× end-to-end speedup in retrieval-based settings and up to 2.78× when added to EAGLE-3.

arXiv:2610.31338 PDF

HERD: Harness-Evolving Recursive Distillation for Continual Learning

Xun Sun, Thanh Thi Nguyen, Thuy-Trang Vu Continual learning methods that update weights suffer from forgetting, while methods that rely on prompts, skills, or harnesses are limited by context length and lossy text compression. HERD (Harness-Evolving Recursive Distillation) combines the two. It continually refines reusable skills in the harness and keeps the examples those skills fail to capture in a bounded memory, then distills this experience into the weights with drift constraints that reduce forgetting. On the 25-stage TRACE+17 task stream under fixed budgets, HERD matches the strongest weight-updating baseline on learning new tasks while showing the least forgetting and the best retention of general capabilities.

arXiv:2610.31401 PDF

SQuAT: Spherical Lattice Vector Quantization-Aware Training

Sebastian Loeschcke, Ivan Koryakovskiy, Tycho F. A. van der Ouderaa, Riccardo Del Chiaro, Mart van Baalen, Paul N. Whatmough et al. Lattice vector quantization offers strong compression for low-bit LLMs, but how it behaves under quantization-aware training (QAT) at scale is poorly understood. SQuAT splits each weight block into a direction, quantized with a spherical codebook derived from the E_8 lattice, and a magnitude, quantized with a low-bit scalar quantizer. Each part gets its own straight-through estimator (STE), which implicitly yields spherical optimization of the directions. A distributed exact nearest-neighbor assignment scheme cuts QAT training time by 12.8x, allowing training on 10B tokens. Across LLaMA-3.2-1B/3B, Qwen3-4B, and Qwen3-1.7B, it consistently outperforms standard STE-based vector-quantization QAT and reaches state-of-the-art accuracy for this class of method.

arXiv:2610.31412 PDF

Predictable performance is suboptimal, and what it means for model routing

Guanhua Zhang, Moritz Hardt Model-routing systems usually train a separate predictor to guess whether a candidate language model will answer a query correctly. The authors compare such predictors against a simple baseline: the model's own self-confidence, meaning the probability it assigns to its chosen answer. They prove that any predictor beating self-confidence can be turned into a recalibration that improves the model's own log loss. That advantage is bounded by the model's suboptimality plus the extra capacity the predictor adds, and the result extends to two-model cascades. Experiments with 16 language models on four benchmarks confirm that predictors help most on more suboptimal models, and their edge shrinks as models are finetuned.

arXiv:2610.31428 PDF

DIAL: Steerable Language Models with Distilled Instruction-Adherence Levels

Thanh Q. Tran, John Russell Himawan, Arun Verma, Kiwan Wong, Wei Xiao, Daniela Rus et al. Prompting gives users no reliable way to control how strongly an LLM follows an instruction such as keeping a response short. DIAL (Distilled Instruction-Adherence Levels) offers numerical control over adherence. It uses classifier-free guidance on a frozen LLM to produce training signals for on-policy distillation into a lightweight steering adapter. The adapter's text-conditioned classifier sets a per-token steering edit in closed form, so one shared adapter can handle new instructions without per-instruction fitting. Across several LLM families, concept scores increase monotonically with the requested adherence level while fluency stays competitive. The authors also release SteerIF, a dataset of 98,625 instructions.

arXiv:2610.31438 PDF

Evaluating the Applicability of LLM Inference Optimizations on Apple Silicon

Kyrylo Yemets, Andrii Yaroshevych, Victor Muryn, Ostap Khomenko, Pavlo Kryven, Maksym Kmet et al. Most LLM inference optimizations are built and validated on server NVIDIA GPUs under batched serving, yet Apple Silicon is one of the most common platforms for local inference. The authors test KV-cache pruning, speculative decoding, ported GPU kernels, and post-training quantization on four Apple machines under single-request decoding, with matched RTX 4090 comparisons. KV-cache pruning gives up to 1.75x throughput on Apple but can slow down desktop NVIDIA GPUs, and every speculative decoding method tested reduced throughput on Apple despite reported 2-6x gains on servers. CUDA kernel ideas transfer only partly because public Metal lacks asynchronous copy. The authors conclude that a speedup depends jointly on the technique, runtime, hardware, and workload rather than on the technique alone.

arXiv:2610.31479 PDF

Towards User Simulators That Stay On Task And Still Sound Natural

Mustafa Omer Gul, Philippe Laban, Tanya Goyal, Jennifer Neville To test and improve AI assistants on long multi-turn tasks, researchers need simulated users that both stick to complex task intents and talk like real people. Existing simulators usually manage only one of the two. The authors propose a reinforcement learning (RL) recipe that rewards intent adherence and behavioral naturalness together, and they measure naturalness with a lightweight discriminator trained adversarially, in alternation with the simulator. The released UserLM-2.0 produces valid conversations (both faithful and natural) 46.2% of the time, more than three times the 14.1% of the strongest prior simulator.

arXiv:2610.31512 PDF

Beyond Causal Masks: Auditing and Localizing Hidden Future-Token Leakage in Adaptive Sequence Models

Huiting Liu, Jinsong Wu Hidden future-token leakage can make an autoregressive architecture look much better under teacher-forced evaluation while its free-running generation gets worse. In an adaptive residual-routing decoder with a correct causal attention mask, a sequence-level routing summary still exposed future tokens. The leak cut validation loss from 3.5958 to 1.4259, yet the model failed a prefix-invariance audit. Causal repairs erased the apparent gain entirely. The authors propose diagnostics, including a maximum prefix-output discrepancy metric (LeakScore) and layer-wise profiles, and argue that lower teacher-forced loss should always be checked with whole-model causality tests and free-running evaluation.

arXiv:2610.31522 PDF

When to Follow the Teacher: Advantage-Based On-Policy Distillation

Mahdi Sabbaghi, George Pappas, Hamed Hassani, Adel Javanmard Standard online policy distillation (OPD) trains a student to imitate the teacher at every state, even where the teacher's choice does not help the student succeed. AOPD (Advantage-Based On-Policy Distillation) weights teacher actions by their counterfactual advantage under the student, reinforcing helpful ones and downweighting the rest. The objective is estimated without bias from ordinary student rollouts and terminal rewards using teacher importance weights, so it needs no extra rollouts. Experiments on math, coding, and MedQA show gains over distillation and RL baselines in both external-teacher and self-distillation settings.

arXiv:2610.31563 PDF

On the Token Collapse Phenomenon in Textual Graph Soft Prompting

Andrea Giuseppe Di Francesco, Rishabh Jain, Franco Maria Nardini, Nicola Tonellotto, Pietro Li\`o, Fabrizio Silvestri G-Retriever and related methods inject graph structure into LLM prompts through soft tokens learned by a graph neural network (GNN). An analysis of these methods finds that the graph tokens collapse to nearly identical vectors across inputs. Swapping them for a dataset mean or a token from a corrupted graph has no resolved effect on accuracy, and a learned constant prompt that never sees the graph matches or beats G-Retriever on three benchmarks. The authors propose a two-stage procedure that yields non-collapsed tokens and show through interventions that task accuracy alone cannot confirm the model actually uses graph content.

arXiv:2610.31572 PDF

Rethinking Sequence Compression with Test-time Regression

Sihan Yang, Jie Huang, Hanxuan Li, Yifan Guo, Weigao Sun, Yiran Zhong et al. Sequence compression is a key bottleneck in efficient language and multimodal models, and common pooling or projection compressors perform poorly. TTR-Compressor treats compression as test-time regression. The input tokens define a regression problem learned end-to-end, and the fitted coefficients, computed in closed form during the forward pass, form the compressed representation. As the compressor for block-wise sparse attention it improves RULER scores over attention-based compressors by a large margin, and as a visual token merger for multimodal LLMs it matches the standard MLP merger with 55.6% fewer visual tokens.

arXiv:2610.31574 PDF

What Makes a Good Unmasking Policy? A Controlled Study of Learned Unmasking in Diffusion Language Models

Jinya Sakurai, Xun Xu, Jaehong Yoon Masked diffusion language models decode tokens in parallel, and learned unmasking policies decide which predictions to commit. Existing policies differ in so many ways that it is unclear which choices matter. A controlled study finds that labeling any prediction that matches the final output as committable is too permissive at low numbers of function evaluations (NFE), that hidden states complement confidence scores, and that about a 10% NFE reduction offsets the overhead of policies up to 7.5M parameters. An 856K-parameter policy trained on trajectory-contraction targets achieves the best accuracy–NFE frontier and highest throughput on GSM8K with LLaDA-8B-Instruct, is competitive on MATH-500 and HumanEval, and largely transfers to Dream-7B-Instruct.

arXiv:2610.31576 PDF

BeNice! Closing Looping's Inference-Memory Gap via KV-Sharing

Hayden Prairie, Jacob M. Ryan, Taylor Berg-Kirkpatrick, Daniel Y. Fu Looped architectures reuse the same block of layers several times to improve quality without adding parameters, but each pass produces its own KV-cache, so inference memory grows with the number of loops. Sharing the final pass's KV-cache with earlier passes sometimes works without quality loss, though unreliably, and the authors propose a two-stage recipe that makes it work consistently: stochastic depth during pretraining, then a post-training method, BeNice!, that trains the first pass to accept the final pass's cache. The result matches the perplexity of standard unshared fine-tuning and composes with multi-latent attention, Kimi delta attention and fp8/nvfp4 KV-cache quantization. In simulated low-memory, long-context inference, sharing raises throughput by up to 23.6% and cuts p90 time-to-first-token by up to 83.6%.

arXiv:2610.31633 PDF

Low-Bit N-Gram Embeddings for Language Models

Stephen Chung, Wenyu Du Many recent open-weight language models add n-gram embedding tables that store learned representations of short token sequences, and these tables can be huge: DeepSeek-V4.1-Flash devotes 196 billion parameters to its Engram n-gram memory. The authors study storing these tables at low precision, comparing quantization methods and bit allocations across tables. In nanochat-, Llama- and GPT-2-family models trained from scratch, four-bit tables save about 75% of table memory while keeping more than 99% of the validation-loss gain the tables provide, and spending the saved memory on larger tables beats the original full-precision setup. On DeepSeek-V4.1-Flash and Qwen3.8-Flash-Next, low-bit tables save 47–55% of table memory with perplexity increases under 1.2%.

arXiv:2610.31646 PDF

17 more specialized papers

Agents 60

The Cost of Saying Hello: An Empirical Study of Handshake Overhead in AI Agent Communication Protocols

Pranav Varshney Agent protocol surveys list handshake overhead as an efficiency criterion, but none had measured it. Using each protocol's official SDK, this benchmark splits the handshake into phases for MCP (the 2025 initialize and 2026 discover versions), A2A, and ACP, and measures round trips, latency, bytes on the wire, and token cost across cold-start, warm-repeat, capability-mismatch, and degraded-metadata scenarios. Round-trip count dominates: at 50 ms RTT, legacy MCP takes three round trips and 230 ms, while A2A takes one round trip and 61 ms. When an LLM reads the counterpart's metadata, cost grows by 24 to 107 tokens per capability, so a 200-capability counterpart consumes 11% of a 200k-token context window before the agent reads its own task.

arXiv:2610.30151 PDF

MCP vs CLI for AI Agents: A Comparative Analysis of Tool Integration Strategies for Autonomous Systems

Arun Malik, Nitish Talasu, Bradley Klemick, Vineet Tushar Trivedi, Prachi Shah, Sanay Devi et al. AI agents can reach external systems either by running shell commands through a command line interface (CLI) or through the Model Context Protocol (MCP), which provides typed, governed tool access over JSON-RPC. The authors compare the two across 14 dimensions using worked examples, security assessment against frameworks such as OWASP and NSA guidance, and economic modeling. They conclude that CLI is more efficient for single-user local workflows, while MCP provides access control, credential isolation, structured auditing, and multi-tenant governance that CLI cannot. They also argue that progressive disclosure brings MCP's token usage close to CLI's, and they recommend a hybrid architecture for production.

arXiv:2610.30157 PDF

TauIndianBankBench: Procedure-Aware, Action-Oriented Evaluation of Tool-Using Conversational Agents

Krishanu Adhikary, Prashant Devadiga Benchmarks for tool-using conversational agents usually check only the final database state, so an agent that skips a required step that leaves no record can still be scored as successful. TauIndianBankBench contains 1,000 Indian retail-banking tasks with a simulated customer, and it also checks that every procedurally required step was taken and that no unrequested change was made. Across all models tested, 6.5% to 12.0% of sessions counted as successful under final-state checking actually skipped a required step or made an unrequested change.

arXiv:2610.30161 PDF

Self-Evolving Static Analysis via LLM Pseudo-Labeling

Rui Melo, Andre Duarte, Aditya Oke, Shubham Gandhi, Nachiket Kotalwar, Rui Abreu et al. cross-listed Static analyzers miss security bugs when nobody has written detection rules for new frameworks and idioms. SESAA starts with no rules and has an LLM agent iteratively write readable joern queries, revising them based only on feedback from a second LLM critic that flags likely false alarms and misses without seeing ground-truth labels. On the OWASP Benchmark for Python, it reaches a mean F1 of 0.742, well above five widely used tools in their default configurations and comparable to the best of them tuned with ground-truth labels. In a pilot on real Python projects, queries induced from code that predated each vulnerability detected 10 of 19 later CVEs.

arXiv:2610.30184 PDF

Agentic Verified Optimization of Real-World Software

Zhe Ye, Udaya Ghai, Zhizhen Qin, Dawn Song, Soonho Kong cross-listed Coding agents can make programs faster, but they are usually checked only with tests, and agents have been shown to game those tests without keeping the intended behavior. The proposed task, agentic verified optimization, asks an agent to produce a faster implementation in Lean 4 together with a machine-checked proof that it is equivalent to the existing code, which serves as the specification. The authors also build Aqueduct, an agentic framework that ports C, C++, Go, and Rust programs into native Lean and validates each port by differential testing against the original. Agents produced faster proved rewrites for 156 of the 189 reference implementations in Verina, and on 17 migrated applications, 92 proved rewrites sped up 47 of 51 scenarios, with smaller gains at the application level than at the function level.

arXiv:2610.30188 PDF

Feedback That Backfires in Games: Do LLM Players Repeat the Move They Just Watched Fail?

Esmail Gumaan When an LLM agent's move fails, the harness usually writes the move and the error into the transcript and asks again, and this record was already known to make small models repeat failed tool calls. The authors test the same effect in Sokoban, Connect Four, and Othello across 8 instruction-tuned models (135M to 1.7B parameters) and find that the failure record makes the failed move more likely (a corrective gain of -1.38 nats per token), driven by the move's surface form and present for illegal but not for losing moves. Describing the failure without repeating the move string avoids the effect; a do-not-repeat instruction does nothing, and a decoder ban stops exact repeats but not illegal play. Along the OLMo 2 1B post-training stages the inversion is largely removed at the DPO step, and RAGEN's default models already show it before any RL. All experiments ran on a free Colab GPU.

arXiv:2610.30205 PDF

SWE-VERIFIER: Selective Verification for Cost-Efficient Coding Agents

Kaijian Zou, Vaibhav Chennareddy, Yunxiang Zhang, Naihao Deng, Longju Bai, Jiaxin Pei et al. Sampling and verifying alternative actions can improve coding agents, but doing it at every step adds substantial computation and API cost. SWE-Verifier, a lightweight generative model trained from Qwen3.5-4B on preference data, scores each proposed action to decide when extra candidates are worth sampling, then ranks those candidates before execution. In the mini-swe-agent harness it raises resolution from 75.47% to 77.20% for MiniMax-M2.7 at 1.09x cost, and from 72.60% to 74.27% for DeepSeek-V4-Flash without training on that model's actions. It verifies fewer than 4% of steps while matching or beating trajectory-level and every-step baselines at much lower cost.

arXiv:2610.30284 PDF

GAV: A scientific experimentation environment for PDE agents

Muheng Li, Wenlong Mou Coding agents that write solvers for partial differential equations (PDEs) need reliable feedback, but checking a program on a new problem usually requires an expensive reference computation. Generator-Agent-Verifier (GAV) is an experimentation environment in which a generator learns a distribution of reference solutions from simulation or measurement data. It then uses the method of manufactured solutions to build PDE problems whose answers are known by construction. The agent iterates on its program while a verifier scores accuracy and compute cost, and that feedback can steer generation toward problems that expose the program's weaknesses. On a numerical Navier-Stokes task, GAV cuts the mean test error of three existing PDE agents by more than an order of magnitude, and the approach is also tested on Vlasov-Poisson-type equations.

arXiv:2610.30292 PDF

MemPlan: Memory-Conditioned PDDL Planning for Partially Observable Text Environments

Tengwei Song, Robert Hoehndorf Large language model agents used as long-horizon planners are costly and prone to invalid actions, so MemPlan instead uses the model as a bridge between language and a symbolic PDDL planner in partially observable text environments. The PDDL domain stays fixed, and at each step the PDDL problem is rebuilt from parsed observations, candidate hypotheses, and interaction memory. Positive graph memory assigns lower action costs to plausible hypotheses, while negative memory turns assumptions already refuted by execution into exclusion constraints. A language-to-symbol interface trained offline with supervised fine-tuning parses observations and grounds actions as commands. On TextWorld, ALFWorld, and Robotouille, MemPlan improves task completion and step efficiency while substantially reducing online token cost compared with ReAct.

arXiv:2610.30379 PDF

RoleFlow: Self-Evolving Multi-Role Agentic Collaboration via Structural Adaptation

Yi Ge, Pan Wang, Xuehao Cui, Zian Zeng, Xu Hu, Heng Huang et al. LLM multi-agent systems usually use fixed collaboration structures, and dynamic orchestration methods trained on trajectory-level rewards struggle to credit individual structural changes. RoleFlow refines the collaboration structure through feedback-driven atomic edits and synthesizes new roles on demand. Its training method, Single-Edit Counterfactual Policy Optimization (SECPO), scores each edit against a no-edit counterpart at the same chain state and assigns credit to the edit-specific tokens. Across seven benchmarks in four domains, RoleFlow beats the strongest competitor by 4.7 points on average, leads all three multi-hop QA benchmarks by 3.2–7.5 F1, and also performs best on math, web interaction and code generation.

arXiv:2610.30429 PDF

Self-Assessment Reinforcement Learning for Search-State Awareness in Retrieval-Augmented Large Language Models

Taewhoo Lee, Minju Song, Jungwoo Park, Chanwoong Yoon, Junha Jung, Hyeon Hwang et al. LLM search agents often lack search-state awareness, meaning they cannot tell what information they have gathered and what is still missing, which leads to hallucinated reasoning and searches that stop too early. Self-Assessment Reinforcement Learning (SARL) teaches models to assess their own search state during training and needs no extra assessment step at test time. A dual reward credits the corrective effect of each self-assessment and penalizes relying on it too much. SARL reports a 24.3% relative improvement over the strongest baseline, with average gains of 35.9% on multi-hop reasoning and 27.9% on deep research benchmarks, and it stays stable over longer search horizons where existing methods degrade.

arXiv:2610.30435 PDF

SplatLab: Autonomous Scientific Discovery for 3DGS Research

Seemandhar Jain, Manmohan Chandraker cross-listed SplatLab is an autonomous research system for 3D Gaussian Splatting (3DGS). It reproduces a published baseline, proposes an extension, trains and ablates it on Mip-NeRF 360 and LLFF, and writes the manuscript, with every reported number tied to a logged training run. Candidate extensions are plugin mutations on a verified reproduction. They are ranked by an Elo tournament across eight islands using measured PSNR (peak signal-to-noise ratio), and a standing falsifier deducts Elo when a candidate fails stress tests. The generated manuscripts beat their reproduced targets by +0.18 to +3.78 dB PSNR, including a nebula-rendering task with no prior 3DGS method, and the authors report that a research cycle shrinks from months to days.

arXiv:2610.30468 PDF

Why Bother Me Again? Benchmarking Personal Agent Memory for Interaction Preferences

Jialin Li, Zhenhao Chen, Hanjun Luo, Hanan Salam Personal agents can burden users with repeated questions, unnecessary confirmations, and unwanted outreach, yet memory benchmarks mostly test factual recall. MemPABench is an interactive environment in which four simulated users, holding 14 interaction preferences, give feedback over 99 sessions spanning four months, with planned preference changes along the way. Agents must decide when to clarify, confirm, or act on their own. Across four memory systems and four LLM backbones, retrieval-based memory (Mem0, MemOS, simple RAG) did not outperform having no memory. File-based memory improved accuracy but applied preferences outside their relevant contexts, and agents often fell back on default habits even when the correct preference was available.

arXiv:2610.30485 PDF

From Single-Actor Environments to Multi-Actor Event Worlds: Can LLMs Be Event World Models?

Juntong Ni, Zewen Liu, Bohan Wang, Lige Zhang, Wei Jin The authors define "event worlds": systems with multiple independent actors that keep changing without the agent's input and are observed only through event logs. They ask whether LLMs can serve as world models for such systems, using benchmarks built from human Diplomacy games and real Polymarket trades. Fine-tuning and more data improve prediction when every actor's actions are given. When only the agent's own action is known, models struggle wherever other actors strongly shape the outcome, and larger models or longer histories help little. This gap also hurts multi-step rollouts and leads models to overestimate how likely plans are to succeed, though imagined experience from these world models still improves policy learning.

arXiv:2610.30488 PDF

What the Trace Shows and What the Label Says: Auditing Behavioral Detectors in Agent Evaluation

Idunnuoluwa A. Adeniji, Shilpika, Michael E. Papka Agent evaluations reduce execution traces to labels such as answer correctness, tool re-verification, and fault propagation, and the authors audit whether those labels match what the traces actually show. In a GSM8K setting where a calculator tool sometimes returns corrupted outputs, they examine three detectors across six production models. A common failure is silent tool-output override: the model discards the corrupted value without calling the tool again, so the error stops spreading but the verification detector records no re-check. Human review agreed with every correctness and fault label but with only 67 of 81 verification labels, and one model's score ranged from 21/30 to 27/30 depending on step and token budgets.

arXiv:2610.30491 PDF

From Prose to Properties: Testing Research Claims with ARATest for Agent-Native Science

Yifeng He, Jiachen Liu, Hao Chen AI agents now produce research claims faster than people can check them, and those claims are stated in prose rather than as testable properties. ARATest is a property-based testing library that maps each claim onto a reusable property skeleton and checks it against linked evidence or against fresh runs of released code performed by an agent, the DissClaimer plugin. Each property gets a verdict of holds, violated, or inconclusive, and every violation comes with a replayable witness. Of 54 claims with faithful declarations, 32 held, 8 were violated, and 14 were inconclusive; two violations exposed errors in published papers that their authors confirmed, and the tool caught all 553 seeded faults with no false alarms.

arXiv:2610.30495 PDF

Can LLMs Improve Their Own Harness? A Controlled Study on ARC-AGI-3

Jens Lehmann, Sahar Vahdati A harness is the program that orchestrates an LLM's perception, memory, tool use, and recovery; the authors test whether frontier models can build and improve their own harness on ARC-AGI-3, an interactive benchmark of unfamiliar games. Each model builds a harness through a fixed coding bootstrap, plays five development games, and revises its whole codebase twice. Every version is scored on twenty held-out games using Relative Human Action Efficiency (RHAE), a measure of level completion and action efficiency against human baselines. Across 27 lineages, harness scores ranged from zero to near ceiling, but two revisions raised mean held-out RHAE by only 2.80 points, with a confidence interval that includes zero. Gains on the development games overstated held-out improvement, and cutting inference cost sometimes cost task performance.

arXiv:2610.30499 PDF

PLE: Physical-Engineering's Last Exam

Rujun Han, Yufan Zhuang, Arijit Ray, Han Yu, Zoey CuiZhu, George Lee et al. PLE is a benchmark of 85 long-horizon agentic tasks in electrical, civil, structural and mechanical engineering, where each task is a realistic, billable deliverable such as a design package a licensed reviewer would approve. Some tasks were written entirely by practising engineers. Others started from engineer-written problem statements and were built out by an adversarial Creator-Solver agent pipeline, which produces the project environment, an executable grader and reference deliverables, before experts validated them and made them harder. Across frontier LLMs and seven public agent harnesses, the best configuration averages only 50.6% while using about 27M tokens and 63 minutes per task.

arXiv:2610.30537 PDF

Governed Self-Evolution for Autonomous Investing: A Multi-Agent LLM System with an Invariant Core

Yebo Feng, Jingyang Dai, Liangrui Pan, Nelson Lau, Jiahua Xu, Yang Liu Steward is a multi-agent LLM system for autonomous investing that is allowed to improve itself, but its type system walls off an invariant core of order arithmetic, risk rules and discipline-agent prompts that no self-edit can touch. Nine role-specialised agents (analysts, a Portfolio Manager and a protected Risk Manager) share a seven-layer memory assembled by code. Every night, code grades each trade and forecast; every quarter, a reflector proposes prompt changes that a deterministic gate applies reversibly and verifies by replaying recorded inputs. During a five-month live deployment the system held a 24.6% sector drawdown to a 0.3% loss and never followed an injected instruction, and a single-model baseline given the same data traded more and underperformed.

arXiv:2610.30557 PDF

Diagnosing Personalization Failures in Mobile GUI Agents

Zhixin Lin, Dongliang Xu, Jungang Li, Shidong Pan, Jun Feng, Yanbin Sun et al. Personalized mobile GUI agents must work out both how a user preference should change their actions and when along the interaction it should apply, which overall task-success metrics do not show. PrefDiag-Bench sorts preferences into five types and adds a stage-wise diagnostic with 1,664 probes to locate where failures occur. Eleven agent systems were evaluated, with user context given either as explicit profiles or as behavioral memory logs. Even the best model, Seed-2.0-Pro, reaches only 47.17% on Process preferences, and the main bottlenecks are choosing relevant evidence from user history and recognizing when a preference should take effect.

arXiv:2610.30573 PDF

Anticipating Social Consequences with Distributional World Models

Viet-Thanh Pham, Thuy-Trang Vu, Zhuang Li, Lizhen Qu, Gholamreza Haffari Agents that negotiate or cooperate need to predict how a social interaction will unfold after each action. Since one action can lead to several plausible futures, predicting a single future misses alternatives, and running repeated simulations is expensive. DSWM, a Distributional Social World Model, outputs alternative responses, their probabilities, and their immediate and final outcome values for each candidate action in a single pass. It is a 4B model trained on repeated simulations with supervised fine-tuning and GRPO, and an acting agent uses its predictions to choose actions. Across three seen and two unseen environments it achieves the best average performance on 13 of 15 metrics, and ablations show that keeping multiple futures with their probabilities helps in every environment.

arXiv:2610.30657 PDF

Governance-State Corruption in Long-Term Memory for LLM Agents: A Cross-Model Stress Test

Bowen Liu, Youting Wang, Bonan Shen, Dingyan Shang An LLM agent can retrieve a memory correctly and still misuse it if the memory's validity status has changed. The authors study governance-state corruption, meaning faults in which retained records get into the model's context while the stored content and tasks stay the same, and test 2,080 matched fault pairs across six model providers. Corruption changed what the model sees in 96.25% of pairs, and oracle-backed enforcement restored it in all of them. Downstream, corruption reduced LongMemEval correctness by 40.54 percentage points and rescaled Memora scores by 30.60 points. The authors argue that memory systems should be evaluated for the integrity of their control state, not just retrieval accuracy.

arXiv:2610.30698 PDF

AgentTriage: Replayable Security Diagnosis Across Heterogeneous Agent Runtimes

Jie Ying, Xutao Zhang, Ningyuan Chen, Mengce Zheng, Jun Li, Zhongjie Zhu et al. When tool-use agents write files, run code, or send data, their final responses do not explain the resulting security events, and different agent runtimes expose evidence through incompatible interfaces. AgentTriage defines an executable contract for security diagnosis: runtime overlays normalize native observations, and a deterministic pipeline builds a typed provenance graph, selects the policy-relevant slice, and binds every claim to recorded events, abstaining when required evidence is missing. Because the pipeline is deterministic, a diagnosis can be recomputed and re-checked without any language-model judgment. Across five runtimes, one contract captured all 159 applicable executions, and on 120 adjudicated cases it reached 0.858 grounded joint diagnosis with a 0.036 unsupported-claim rate, 0.175 above the strongest same-schema baseline.

arXiv:2610.30738 PDF

Is Complex Agentic Memory Always Needed? Naive RAG Is Stronger Than We Thought

Zhengye Bao, Shengtao Zhang, Haokai Zhang, Shiqi Lou, Muning Wen, Minnan Luo Agentic memory systems for long-term conversation often rely on memory mutation, structured organization, or learned retrieval, and the authors test whether that complexity is needed. Their baseline, Raw-RAG, simply applies dense retrieval and reranking over the raw conversation history and passes the evidence to a reader LLM. Across LoCoMo, LongMemEval, MEME, and MemoryAgentBench, Raw-RAG ranks first in six of seven benchmark-model settings, beating the strongest agentic memory system by 0.13 to 5.53 points. Analysis attributes this to higher coverage of necessary evidence, including the relations needed to derive answers, and shows that even full access to transformed memory stores leaves an accuracy gap relative to raw history.

arXiv:2610.30755 PDF

AgentDP: Towards Automated Model Deployment for Heterogeneous Embedded AI Applications

Leming Shen, Qiang Yang, Xinyu Huang, Zijing Ma, Yuanqing Zheng Deploying pre-trained models on heterogeneous embedded devices requires manual model optimization and adaptation that is slow and device-specific. AgentDP automates this with LLM agents and three components: a module that standardizes user requirements, an operator-aware partitioning scheme that runs different parts of a model on different inference engines, and a configuration-reuse module that retrieves earlier hyperparameter settings for new tasks. Experiments show that it reduces total model adaptation time by up to 80% compared with manual work while achieving more than 13% higher task accuracy on average than baselines, and a user study supports its generalization across requirements and devices.

arXiv:2610.30760 PDF

LLM Multi-Agent Systems Are Vulnerable to Herding and Runaway Cascades

Farnoosh Hashemi, Shiyu Ji, Joice Chen, Michael W. Macy LLM agents in multi-agent systems respond to one another's choices, which creates feedback loops prone to herding. Replicating the classic Music Lab experiment, the authors had large populations of LLM agents recommend unfamiliar songs, with or without seeing cumulative recommendation counts. Seeing the counts triggered runaway cascades that produced extreme, unpredictable outcomes increasingly detached from each song's intrinsic appeal. LLM populations were far more susceptible to herding than humans and less able to self-correct, which bears on multi-agent recommender systems, automated evaluation, and collaborative problem-solving.

arXiv:2610.30769 PDF

Agentic Knowledge-Model Co-Evolution via Consistency Regularization

Sirui Chen, Zhining Liu, Chao Zhang, Jingrui He, Zheng Li LLM agents can improve by turning past experience into reusable knowledge such as skills and memories, but different ways of analyzing the same experience can produce knowledge of very different usefulness. SPLICE borrows the idea of consistency regularization: it varies how an experience is analyzed to produce several knowledge candidates, then keeps updates whose usefulness gains hold up consistently across those variations. Model updates combine reinforcement learning with self-distillation of the evolved knowledge, using knowledge that consistently hurts performance as a contrastive reference, and knowledge updates and model updates alternate. SPLICE beats knowledge-evolution and model-training baselines by up to 7.1 and 4.6 points of average accuracy, and its evolved knowledge cuts output tokens per correct solution by up to 44% while transferring across model sizes and families.

arXiv:2610.30790 PDF

Trace2World: Reconstructing App Worlds from Real Interaction Traces for Scalable GUI Agent Training

Zezhou Wang, Hanzhang Zhou, Panrong Tong, Quyu Kong, Pengxiang Li, Steven Hoi et al. Scaling reinforcement learning for GUI agents needs training environments that behave like real apps, but generating mock apps directly from partial interaction traces often produces incorrect state transitions. Trace2World first merges traces into a shared app model linked to its source evidence, then compiles that model into an executable environment, and it replays source workflows to find mismatches that need repair. From 1,570 traces it builds 50 environments that reproduce 91.1% of 200 workflows end to end, compared with 59.8% for the strongest baseline, and validation against the source evidence keeps 94.0% of simulated training gains on real devices. Training in these environments improves a model by 20.2 points on the authors' benchmark, by 5.4 to 16.2 points on four external GUI benchmarks, and by 16.3 points on 100 real-device tasks.

arXiv:2610.30792 PDF

ADAPT for Science: A Multi-Agent System for Self-Evolving Scientific Optimization

Zhaoran Feng, Yuchen Ni, Jianan Liu, Lemao Liu, Dejing Dou, Kairui Feng Scientific design problems involve costly evaluations, sparse feasible solutions, and strict physical constraints, which general-purpose agentic evolution methods are not built to handle. ADAPT for Science is a multi-agent framework that splits optimization into deciding where to search, by seeding from diverse promising historical designs, and how to search, by evolving the optimization procedure from execution feedback. On gravitational-wave detector design with a differentiable physics simulator and eight unseen detector topologies, it reaches a mean loss of 0.204 under matched evaluation budgets. That is 54.9% lower than the best classical optimizer and 62.2% lower than the best general-purpose agentic evolution baseline, and it has the lowest loss on every topology.

arXiv:2610.30851 PDF

Separate Field Scoring Protects Unused Tools When Agents Pool Tool-Use Experience

Abhijit Chakraborty cross-listed Agents that share a tool catalog can pool usage examples to improve tool retrieval. However, adding those examples as extra index entries raises retrieval of tools that have usage and substantially lowers retrieval of held-out tools that no agent has used yet. SYNAPSE avoids this by scoring usage as a separate field. On StableToolBench, held-out gold tools land in the top five 0.785 of the time versus 0.683 for a router that reserves slots for usage, and this lead holds across three catalogs and two embedders. A reserved-slot router sometimes keeps more of the gain on labeled queries, so the better choice depends on the workload. Splitting usage across ten simulated agents helps exactly as much as giving one agent all the records.

arXiv:2610.30867 PDF

Harnessing Language Models as Optimization Modeling Agents

Yunxiang Luo, Xin Xu, Tianhao Chen, Zhixi Xiong, Peiqi Jiang, Fan Zhang et al. LLMs can translate natural-language optimization problems into solver programs, but the formulations and code they produce are often wrong, and current fixes rely on multi-role agent workflows or task-specific post-training. The authors test whether a lightweight inference harness with a single model role can compete, introducing OPTICYCLE and OPTISTATE and evaluating them on nine benchmarks (2,634 problems) with 11 Qwen variants from 0.6B to 27B parameters. Revising programs based on execution feedback improves macro accuracy over direct generation for every model, by up to 49.46 percentage points, and both harnesses beat six adapted agentic workflows using the same model as well as four post-trained optimization models of similar size, all without extra training. Explicitly tracking the formulation as state helps further, but how much depends on the model.

arXiv:2610.30917 PDF

ReflectPO: Reflection-Guided Policy Optimization for Long-Horizon LLM Agents

Rumei Li, Xunliang Cai LLM agents on long tool-use tasks tend to fail because early mistakes change the interaction history and the environment state, and the damage accumulates over later steps. ReflectPO (Reflection-Guided Policy Optimization) is a reinforcement learning framework aimed at teaching agents to recover from such errors. It builds recovery hints from points where trajectories branch and from states the agent visits repeatedly, and it organizes sampled trajectories into trees so that each action gets credit from the reward difference between parent and child and from any recovery it achieves. On two long-horizon tool-use benchmarks it beats existing baselines in most domains and has the best average performance overall, and further analysis shows agents get better at noticing their own deviations and recovering.

arXiv:2610.30933 PDF

ColaBench: Evaluating Memory-Guided Knowledge Acquisition in Personal Agents

Wenjie Zhang, Zhichao Shi, Xuhui Jiang, Xiaojun Wu, Cehao Yang, Chengjin Xu et al. Personal agents must use a user's long interaction history both to decide what outside knowledge a request needs and to apply that knowledge to the user's situation. ColaBench is a synthetic benchmark in health and daily living with 349 tasks, 10 user profiles and a knowledge base of 13,033 documents. An evidence graph links each answer requirement to personal disclosures and to the documents that support it. The accompanying MemKB method uses state summaries and memory-document links and reaches 86.25% accuracy with about four interaction rounds per task. Errors come mostly from unsupported assumptions about the user and from misapplied rules.

arXiv:2610.30989 PDF

CANNBot-CAKE: Collaborative Agentic Kernel Evolution for Ascend NPU

Deng Pan, Xue-Jian Gao, Junyi Fan, Bin Du, Jiasheng Li, Xinxian Chen et al. AI agents approach expert performance on GPU kernel benchmarks but struggle to write AscendC kernels for Huawei's Ascend NPUs, likely because little training data exists for them. CAKE is an agentic system that combines adaptive routing across multiple domain-specific languages, separate Generation and Evolution modes, and reviewed knowledge reuse through a curated knowledge base (CAKE Book) and a Self-Evolving Agent. On 53 CANN-Bench v1.1 operators it passes 100% of standard cases and 98.28% of hidden cases, with an average 1.334x speedup over the baseline. The Evolution mode cut aggregate device time for one fused operator by 46.6% over 20 rounds.

arXiv:2610.31034 PDF

Dual Exam: Controlled Evaluation of Multimodal Agents and World Models

Zhefan Rao, Meng Chu, Xuan Zhang, Jingfan Chen, Guohao Cai, Runyi Li et al. Studying how multimodal agents and world models help each other requires measuring action selection and state prediction separately. Dual Exam tests agents against fixed environment responses (1,088 items) and world models against fixed action interventions (400 cases), across physical, digital, social, and scientific domains. GPT 6 Astra leads the agent track with an index of 31.3 and GPT-Image-2 leads the world-model track at 21.7, though the two scores are not comparable. A 150-case pilot coupling one agent with one world model records zero task successes, and results show frequent failures to preserve task-relevant state.

arXiv:2610.31066 PDF

nnInteractive Agent: General-Purpose AI Agents as 3D Image Annotators

Fabian Isensee, Constantin Ulrich Harsy, Maximilian Rokuss, Benjamin Hamm, Philipp Vollmuth, Klaus H. Maier-Hein Interactive 3D segmentation models such as nnInteractive can turn a few points, scribbles or boxes into a volumetric mask, but a person still has to find each structure, place the prompts and correct the result. The authors show that general-purpose AI agents can do that work without additional training. They release nninteractive-agent, which exposes nnInteractive as tools over the Model Context Protocol (MCP) along with a skill that encodes how expert users prompt it. On a benchmark of 14 images and 51 targets covering CT, MRI, PET and micro-CT of humans, a mouse and an ant, Claude Code and Codex reach a mean Dice of 0.81 and 0.84 from a plain-language request alone, and they stay accurate on targets where the text-prompted 3D model VoxTell fails outright.

arXiv:2610.31080 PDF

Who Is the Influencer? Consensus Dynamics in Multi-Agent Systems

Gouki Minegishi, Jingyuan Feng, Hiroki Furuta, Takeshi Kojima, Yusuke Iwasawa, Yutaka Matsuo The authors ask which model sets the final consensus when several different large language models (LLMs) are mixed in one multi-agent system. They use a simple task that needs no pretraining knowledge and has no correct answer. Across seven model families, the other models' opinions are consistently absorbed into the opinion of the larger model. To explain this, they model the agents as weighted particles moving on a simplex, where a larger weight corresponds to stronger in-context learning that amplifies the majority of the context and gives a deeper well in the energy landscape. Because larger models are generally better in-context learners, they end up as the influencers.

arXiv:2610.31105 PDF

Learning What to Ask for Tomorrow: Active Preference Acquisition for Long-Term Personalized Web Agents

Yiqi Wang, Zizhi Zeng, Jiaqi Zhang, Yiqun Duan, Mingkai Zheng, Taotao Cai Web agents that serve the same user across many tasks need to know that user's preferences. Existing clarification methods judge a question only by how much it helps the current task, and so they ignore the value of reusing the answer later. Active Preference Acquisition (APA) scores each missing preference by its Prospective Personalization Value: the expected utility it adds over future tasks, minus the effort and disclosure risk of asking, chosen under a per-interaction budget. On the new longitudinal benchmark ActiveWebPref with a 20-task horizon, APA acquires 80.3% of the valuable preferences that only pay off through reuse, compared with 5.1% for current-task value of information. It also recovers 84.2% of the utility gain from asking about everything while using only 27.6% of the questions.

arXiv:2610.31107 PDF

ProductWebBench: Evaluating Website Continuity Beyond Task Completion

Jiajun Wu, Jian Yang, Yaxin Du, Shukai Liu, Linzheng Chai, Tuney Zheng et al. cross-listed Frontend coding agents are usually graded only on whether they built what was asked, not on whether the rest of the website still works afterwards, which the authors call 'website continuity'. ProductWebBench includes 320 change tasks on existing repositories and 80 long-horizon build tasks. A simulated user reveals requirements in stages, and every deterministic browser check that passed earlier stays active for the rest of the run. Across 8,737 runs from 13 models, 28.2% of change runs that pass every check on the requested work still break continuity, for example with blank routes, horizontal overflow or console errors, and 10.8% turn an earlier passing state back into a failure. Because the same checks also run on the untouched repository, every regression can be traced exactly to the agent's patch.

arXiv:2610.31118 PDF

Prompt Sensitivity in Tool-Augmented Model Comparisons

Mateusz Modrzejewski, Tomasz Radzikowski Benchmarks phrase each question one way, and rewording it can change which of two language models ranks higher. The authors ask whether giving models tools makes these comparisons more stable. They build a matched protocol that separates the answer readout, reasoning, the decision to call a tool, the tool's answer and the report given without a call, and apply it to six open models across five domains and five wordings. Tools raise accuracy where a program can answer the question, but they widen wording sensitivity in many pairwise comparisons; the widening tracks how often a model calls tools and depends on the interface and instructions. The authors recommend reporting call frequency and comparison stability across several wordings alongside accuracy.

arXiv:2610.31148 PDF

Beyond Watching and Searching: Benchmarking Agents for Video-centered Deep Research Reports

Jiahao Meng, Qize Yang, Yuxuan Wang, Yunfei Chu, Wensi Dai, Jin Xu et al. cross-listed Deep research agents write long reports from open-web evidence, but no existing benchmark evaluates research built around a source video. The authors define a task in which an agent takes a video and a user request, identifies research questions, gathers external multimodal evidence and produces an interactive report linked to moments in the video. They release OmniDR-Bench, with 50 Chinese and English tasks and 6,414 criteria organized into rubric trees (Omni-ARTs) that were built through an audited, human-reviewed process. Across 18 agent configurations the best system scores only 63.3 out of 100, and the main gaps are content coverage, causal reasoning across modalities and citation accuracy.

arXiv:2610.31158 PDF

Network topology shapes consensus and vulnerability in LLM populations

Roberto Passaro, Ariel Flint, Romualdo Pastor-Satorras, Andrea Baronchelli, Luca Maria Aiello cross-listed Research on how conventions form among interacting large language model (LLM) agents has mostly assumed that every agent can interact with every other. Here, policies extracted from four LLMs play the naming game with 1,000 to 100,000 agents on fully connected, random and heterogeneous networks. Sparse networks slow consensus by up to twelvefold, but they can restore consensus in some cases that fail to converge under all-to-all mixing. When an adversary plants committed agents to overturn an established convention, targeting hub nodes takes about one-sixth as many as random placement (3% versus 16% median on the most heterogeneous network). Networked populations are usually easier to tip than all-to-all ones, so estimates from all-to-all settings can overstate robustness.

arXiv:2610.31165 PDF

PointerBench: A Diagnostic Benchmark for Fine-Grained GUI Grounding

Valentino Sacco, Saurabh Pandey, Alexander Florian Jercher, Maximilian Schilling Existing GUI grounding benchmarks for computer-use agents focus on standard controls and a limited set of professional interfaces. PointerBench has 1,500 screenshot-instruction pairs with pixel-exact targets in three parts: fine sub-element and geometric grounding in spreadsheets, character-level grounding in long documents, and mixed grounding across 100 professional applications. The authors evaluate general multimodal models and 39 specialist GUI grounding configurations. Text grounding is the hardest part: Claude Fable 5, GPT-5.6 Sol and the specialist UI-Venus-2-9B reach only 69.6%, 53.4% and 25.8%, compared with 83.2%, 77.7% and 54.5% averaged across the whole benchmark.

arXiv:2610.31274 PDF

RIPPLE: Recovery Internalization through Process-Level Policy Learning and Effect-Aware Credit for Long-Horizon Agentic Coding

Jingxuan Xu, Zihang Xu, Yuechen Luo, Ziqi Miao, Wenqiang Zhu, Weihao Li et al. Long-horizon coding agents often notice a failure and write a plausible critique but then keep following the same wrong trajectory, which the authors call the Critique-to-Action Handoff Gap. RIPPLE distills recovery guidance from a teacher that sees only the trajectory so far into a single policy. It then uses reinforcement learning with separate credit for triggering reflection, diagnosing the problem and taking corrective action, with local validity signals so that intermediate steps do not simply inherit the final outcome. Across four repository-level software-engineering benchmarks it improves the base policy by 4.8 points on average without needing a critic at deployment. Gains concentrate where the trajectory contains enough diagnostic evidence and the repository state can still be reversed within the remaining budget.

arXiv:2610.31276 PDF

Beyond Accuracy: Disentangling Decomposition from Typed-State Quality Assurance in Multi-Agent LLM Pipelines

Maria Gabriela Alves Zuppardo, Vinicius Cardoso Garcia cross-listed When multi-agent LLM pipelines improve software engineering tasks, it is often unclear whether the gain comes from typed contracts between agents or from the specialized prompts that splitting the task allows. The authors compare a single-call baseline, a two-call variant with specialized prompts but an unvalidated intermediate result, and a pipeline with a Pydantic-validated typed state. The task is classifying and MoSCoW-prioritizing 625 PROMISE NFR+ requirements, using three locally hosted 7-8B models in two languages. The F1 gain (up to +0.075) comes from the specialized-prompt split, while the typed state adds little accuracy (-0.003 to +0.013). However, the typed state catches out-of-range outputs at the agent boundary, whereas the unvalidated variant fails later and silently drops requirements.

arXiv:2610.31280 PDF

The Value of Memory in Language Agents

Zehao Lin, Erenkai Zhu, Helan Hu, Xixuan Hao, Chunyu Li, Kai Chen Agent memory is useful only if the right distinctions from past experience survive, can be used by a resource-limited agent, and can later be recovered or revised. The authors build a decision-theoretic framework that measures memory's value in units of avoided task loss, following it from raw history through stored state, use, and costly recovery, and they use five minimal counterexamples to isolate the roles of information timing, retrieval access, and demand uncertainty. One counterexample shows that a pointer carrying zero mutual information about the answer can still lower the cost of exact recovery. Diagnostics on LoCoMo, a two-project update task and longitudinal cases expose failures of preservation, use and measurement, including evaluator revisions that change scores even when agent outputs stay the same.

arXiv:2610.31303 PDF

Superintelligent Software Engineering: Governance That Scales with Capability

Mohammad Mehdi Gholinejad cross-listed This position paper argues that once coding agents outpace their human reviewers, more capability does not by itself produce better software, and that governance (gates, contracts, recorded decisions and checks for stale knowledge) has to scale along with capability. A simple model shows that beyond a fixed review capacity, the number of bad changes that slip through grows linearly with throughput, and only governance, not capability, can reduce that slope. A snapshot of 452 highly starred GitHub repositories finds agent context files in 48.67% of them, and in the median such file 9.5% to 22% of the repository paths it names do not exist. The author presents a governance discipline called OSTEON but reports no measured effect for it and labels its mechanism claims as opinion.

arXiv:2610.31306 PDF

Subgoal-aware Memory Graph for GUI Agent

Dezhao Luo, Daniel Beechey, Derek Yuen, Weilin Luo Agents that operate mobile graphical user interfaces (GUIs) on long tasks benefit from reusing past experience, but existing memory methods store either single actions, which carry too little meaning, or whole trajectories, which are too coarse to reuse in other tasks. The proposed method breaks past trajectories into subgoals, meaning runs of consecutive actions that share a single intent, and stores them as nodes in a graph that can be merged, extended and retrieved based on the current screen. On SPA-Bench, it improves success rate by 11.11% over the baseline.

arXiv:2610.31307 PDF

AgenticEvolve: Evolutionary Program Search with Persistent Language Model Agents that Design and Run Experiments

Ze-Yu Peng, Qing-Yu Lan, Zhan-He Wang, Jun Zhang, Yun-Song Piao cross-listed LLM-driven evolutionary program search usually asks the model to produce each new candidate program in a single request. In AgenticEvolve, each island of the evolutionary search is run by a persistent LLM agent that designs and runs experiments on candidate programs between evaluations. On two heuristic-design tasks with three LLMs, it produces better programs than both an otherwise identical version with experiments disabled and OpenEvolve, and reaches the baselines' final quality with far fewer evaluations at comparable cost. On twelve equation-discovery problems, it recovers the ground-truth structure of more equations than LLM-based and genetic programming baselines.

arXiv:2610.31326 PDF

Automating Feature Flag Cleanup at Scale with Multi-Agent LLM Systems: An Industrial Experience at DoorDash

Atharv Chandratre, Jai Datar cross-listed DoorDash has over 56,000 feature flags across 623 repositories, 1,659 of them stale, and each manual cleanup takes an estimated one to two hours. The authors built a multi-agent LLM system that queries live flag rollout data through the Model Context Protocol (MCP), processes Jira tickets, and opens cleanup pull requests end to end. Evaluated on 50 stale flags, it produced engineer-approved removal pull requests for 45 of them (90%), at an average of $4.79 and 13.8 minutes per cleanup. The paper also shares lessons learned from designing the multi-agent architecture.

arXiv:2610.31351 PDF

Muffin - A Shared Memory Buffer for Asynchronous AI Agents

Ping-Lin Kuo, Suman Navaratnarajah, Christian Frey, Gresa Shala, Josif Grabocka Multi-agent systems often perform no better than single agents on goal-directed reasoning tasks, largely because agents exchange critical information poorly. Muffin (shared Memory bUFFer INference) replaces synchronous debate-style turn-taking with an asynchronous protocol. An orchestrator sends demanding queries to a group of agents, which post intermediate findings to a shared memory buffer while keeping their private reasoning local. It achieves the best accuracy in 17 of 20 benchmark settings and, compared with multi-agent debate, gains 19.2 accuracy points while using 77.4% fewer tokens, without sharing full reasoning traces.

arXiv:2610.31355 PDF

Conservative Meta-Agent Search: Architecture Adaptation under Non-Stationarity

Ali Atiah Alzahrani Automated design of agentic systems is usually tested on static task suites, while deployed systems face ongoing distribution shift, delayed feedback, and costly architecture changes. Conservative Meta-Agent Search (CMAS) treats architecture search as replacing a running incumbent: it proposes typed graph edits, scores them against the incumbent by paired shadow replay, and promotes a change only if a calibrated, variance-penalized score clears a margin, followed by a simulated canary stage with rollback. On MetaInvest-Bench, a regime-switching financial simulator the authors built, CMAS cuts normalized regret from 0.116 to 0.104 and raises held-out utility from 0.44 to 0.48 over the strongest adaptive baseline. It shrinks to near-single-agent teams when extra workers add no value, though its search phase has the highest token cost of the methods compared.

arXiv:2610.31436 PDF

Adaptive Skill Granularity for Coding Agents: Runtime Routing and Macro-to-Micro Recovery

Dingxuan Zhang, Yanhui Jiang, Jun Wang, Hefeng Zhou Coding agents can complete a task with primitive tool calls, with reusable macro skills, or with a hybrid that starts with a macro and falls back to step-by-step recovery. Adaptive Skill Granularity treats this choice as a runtime policy that depends on task structure, execution state, and macro failures. It combines a feature-based granularity controller with a finite-state recovery executor. On a deterministic coding-agent evaluation protocol, fixed macro execution is efficient when it fits but brittle when it does not, while runtime routing with explicit macro-to-micro recovery fixes macro failures that fixed-granularity execution leaves unresolved.

arXiv:2610.31461 PDF

Training Small Language Models for Transferable Tool Selection

Zhaoyi Zhou, Akshayaa Magesh, Jeongyeol Kwon, Artem Kuriksha, Daniel Jiang, Andrea Zanette et al. When LLM agents can choose from more than ten thousand tools, picking the right ones for a multi-step instruction becomes a bottleneck, and recent systems rely on proprietary LLMs to decompose the instruction and select from retrieved candidates. The authors show that a small open model can learn this job through reinforcement learning. On ToolBench, Qwen3-8B post-trained with their recipe outperforms the 402B-parameter Llama-4-Maverick. The trained model also works with different retrievers and generalizes to new tool corpora in other domains.

arXiv:2610.31483 PDF

DuoCook: An Overcooked-Inspired Benchmark for Visual Planning and Coordination

Ziyi Jiao, Heng Zhang, Drishti Regmi, Xiaokang Ye, Lianhui Qin, Peter Yichen Chen DuoCook is an Overcooked-inspired benchmark, built in Unreal Engine with the help of coding agents, that tests whether one vision-language model can coordinate two cooking players from top-down images. The model picks high-level actions while the engine handles navigation, across 20 maps covering equipment contention, changing terrain, and handoff dependencies. Among four VLMs, the strongest, gpt-6-astra, completes only 26.2% of orders on average. Giving the model a shared, self-written notebook as memory helped Claude and Qwen on average but hurt GPT and Gemini.

arXiv:2610.31511 PDF

Can an Agent Debug a Web Corpus? Agentic Auditing of Slovak FineWeb2

Marek \v{S}uppa, Ivan Vykopal, Jaroslav Kop\v{c}an Multilingual web-data pipelines such as FineWeb2 apply one recipe across a thousand languages, so failures specific to individual languages go unnoticed. The authors use a tool-using LLM agent as a corpus debugger: it explores the corpus through a read-only, budgeted API and compiles each finding into an executable test, which a native speaker then validates. On the 26.5-million-document Slovak portion, runs costing under $3 each uncovered a machine-generated content farm of 1.4 million documents (5.32%) that had passed the pipeline's filters, and blinded labels supported 6 of 7 findings. The compiled test suite scans the full corpus in 0.7 seconds without a GPU, though a static sampling baseline validated more selectors at less than half the cost.

arXiv:2610.31636 PDF

EpiSEM-Bench: Evaluating Safety and Evidence Recovery in LLM Agents under Missing Evidence

Qijin Zhang, Hanzhen Wang, Yuehan Zhao, Haoyu Dai, Yao Shu, Zhongxiang Dai et al. Tool-using agents decide whether an action is allowed based on facts from their environment, so a tool response that leaves out a safety-relevant fact can lead them to take a prohibited action even when every returned statement is true and the request is benign. EpiSEM-Bench is an executable benchmark of 300 matched unsafe and safe task pairs across six domains that systematically suppresses critical facts and measures unsafe execution, failure thresholds and whether agents go looking for the missing evidence. Across ten models, robustness to omitted evidence varied widely, and resisting missing evidence did not necessarily translate into useful recovery. Giving agents a way to recover the facts reduced unsafe execution for all ten, but models differed in whether they actually retrieved the facts and acted appropriately on them.

arXiv:2610.31650 PDF

3 more specialized papers

Other 47

Future of Time Series Anomaly Detection: From Historical Review to Unified Outlook

Zhijie Zhong, Zhiwen Yu, Pengyu Li, Liujian Zhang, Qiaoli Ni, Jiang Jun et al. This survey of time series anomaly detection (TSAD) updates the field's taxonomy to cover recent architectures such as diffusion models, state space models and LLMs, along with a refined categorization of anomaly types. It also runs experiments on how reliable common evaluation metrics are and releases an open-source benchmark repository of algorithms, datasets and evaluation tools. It closes with open challenges including dynamic thresholding, interpretability, zero-shot transfer, multimodal alignment and time series foundation models.

arXiv:2610.30628 PDF

The Delegation Divide: An Open Index of AI-Agent Delegability across Occupations and Countries

Cristian Espinal Maya (Universidad Nacional de Colombia), Santiago Jim\'enez Londo\~no (Universidad EAFIT) cross-listed Existing measures of AI's labour-market impact score exposure to generative AI, mostly for the United States. The Agent Delegability Index (ADI) has two language models rate all 18,796 O*NET task statements on a preregistered rubric of delegation levels and barriers, then maps the tasks to occupations and weights them by employment in 146 countries. It estimates that 35.2% of global task time is at least partly delegable to AI agents in 2026, but only 2.5% can be fully handed over. Delegable employment rises from 22.5% in low-income to 48.7% in high-income countries, mainly because poorer countries employ more people in physically constrained jobs, and model ratings agree with blind human ratings at kappa 0.85.

arXiv:2610.30752 PDF

Amplifying Latency and Cost of LLM Inference via Rowhammer Attack

Mrityunjay Shukla, Shubham Roy, Biswabandan Panda cross-listed LLVMhammer is a Rowhammer attack that flips a bit in CPU memory during Triton's just-in-time GPU kernel compilation, which leaves output correct but makes LLM inference slower. Because the flip hits an optimization pass, Triton produces a functionally identical but slower kernel and caches it on disk, so a single successful flip causes lasting slowdown. Since the targeted compiler stages are shared across GPU vendors, the attack works on both NVIDIA and AMD hardware. A companion tool, Karnage, finds exploitable bit locations automatically. On Qwen2.5-7B, Llama-3.1-8B, Mistral-7B, and DeepSeek-7B running on RTX 4090 and RX 7900 XTX GPUs, the attack increases inference latency by 4 to 16%, and variants can also corrupt outputs, lowering BERTScore to as little as 25%.

arXiv:2610.30798 PDF

Metacognitive Scaffolds Create Productive Friction for AI-Assisted Knowledge Work

Viktor Kewenig, Lev Tankelevitch, Sean Rintel, Advait Sarkar, Abigail Sellen cross-listed When people work with generative AI, more of their effort goes into planning, evaluating and orchestrating rather than producing, which places new metacognitive demands on them. A randomized experiment with 297 workers writing 594 essays with AI assistance tested dynamically generated scaffolds that prompt users to specify goals and break the task into parts. The scaffolds improved writing quality, and the improvement persisted in a short-lag transfer task after they were removed, along with better recall of the topic three days later and changes in how participants prompted the AI. Participants who received scaffolds nonetheless rated the AI as less helpful and more demanding, which suggests that perceived ease can understate how useful support is.

arXiv:2610.30945 PDF

Random Initialization as a Baseline for Time Series Foundation Models

Pinar Sungu Isiacik, Georgiana Ifrim To measure how much pretrained weights add beyond the architecture itself, the authors compare seven released time series foundation models (TSFMs) from four families against randomly initialized copies with the same architecture, on multivariate time series classification over the UEA archive. Across 183 paired comparisons with frozen representations and four classical classifiers, pretraining always improves mean accuracy, but only by 1.43 to 5.19 percentage points (2.58 on average). Only NuTime remains significant after Holm correction. Non-pretrained baselines such as ROCKET, Hydra and Quant stay competitive, and the authors recommend that evaluations include both random-initialization controls and strong non-pretrained baselines.

arXiv:2610.30950 PDF

Muting Tokens: Exploring Token Sparsification for Efficient Audio Transformers

Shreya Sajid, Fabrizio Julian Piva, Hiram Rayo Torres Rodriguez, Elif Ceren Gok Yildirim, Murat Onur Yildirim, Joaquin Vanschoren cross-listed Token sparsification, which cuts transformer compute by pruning or merging tokens, is well studied in vision and language but not in audio. The authors benchmark a range of pruning and merging strategies for audio transformers across several datasets and model sizes, both after training and with fine-tuning. They also propose a method that combines attention scores with spectral cues from the token embeddings. Frequency-aware criteria, especially spectral flatness and low-frequency energy, consistently beat attention-only methods, and pruning by spectrogram patch intensity works especially well on longer audio inputs.

arXiv:2610.31235 PDF

A Fix for Softmax and Cross-Entropy

Jonathan T. Barron, Dan Gnanapragasam, Michael Tschannen Standard implementations of softmax, log-softmax and cross-entropy have two numerical bugs, according to the authors. In Softmax Collapse, the target logit's gradient rounds to zero when that logit is much larger than the rest, so the gradients no longer sum to zero; in Softmax Cancellation, a student whose logits nearly equal its teacher's still receives non-zero gradients. The authors provide simple reimplementations that fix both, which eliminate the previously observed 'Slingshot' phenomenon, speed up grokking with weight decay, prevent accuracy collapse without weight decay and keep distilled students from drifting away from their teacher. The fix also shows that after memorization, logit growth is bounded by a balance between weight decay and Adam's epsilon hyperparameter.

arXiv:2610.31642 PDF

40 more specialized papers

Safety & Alignment 47

Detecting AI Training in Datacentre Power

Tom Kimpson, Mauricio Baker, Emlyn Graham cross-listed AI governance agreements need a way to detect large training runs on a monitored cluster without trusting the operator. The authors use Viterbi tracking of the wandering iteration cadence that synchronous training leaves in facility-level power draw. On simulated datacenter power with a meter sampling at 2 Hz or faster, the tracker detects at least 96% of undisguised training runs at a 1% false-alarm rate, within 0.04 of the theoretical optimum. Of eight evasion strategies, only extreme variation in work per iteration partly succeeds, yet at least 87% of such runs are still caught over a day, and in small-scale experiments the evasion required 93% to 267% more compute to reach the same loss.

arXiv:2610.30182 PDF

Protection Is Not One Event: A Survey of RAG Protection Research by Measured Endpoint

Peiyang Liu, Xi Wang, Di Liang, Youru Li, Zhi Cai, Wei Ye cross-listed Studies of retrieval-augmented generation (RAG) protection use incompatible metrics, datasets, and threat models, so the authors survey 118 defense papers and double-code them (kappa 0.867) into a seven-family taxonomy organized by the outcome each study actually measures. They find that mechanisms such as reranking and context rewriting span five or six outcome families, that only 17.8% of studies use content disclosure or reconstruction as their primary outcome, and that most evaluations rely on general question-answering data and non-adaptive attacks. In a controlled testbed, an 8-gram copy guard reduces 20-token verbatim copying to 0.0%, yet a three-query adaptive attacker still recovers 48.3% of hidden facts on Qwen3-8B. The taxonomy, dataset, coding protocols, and testbed are released.

arXiv:2610.30193 PDF

MPIDetector: Detecting Multi-modal Prompt Injection via LLM-based Reasoning

Yiheng Li, Jingyue Li cross-listed Prompt injection (PI) hidden in text or images threatens multimodal large language models (MLLMs), and signature- or perplexity-based detectors struggle with sophisticated cross-modal payloads. MPIDetector uses few-shot in-context learning (ICL), giving an LLM prompt templates and example attacks so it can judge the semantic intent of inputs without any fine-tuning. On a new benchmark of eight textual and seven visual attack categories plus benign data, tested with Gemini-2.0-Flash, Qwen3-VL-32B-Instruct, and Doubao-seed-1.6-flash, it reaches up to 90.91% accuracy with 100% recall on text and up to 99.28% accuracy on images. Against 500 heavily obfuscated web-agent injections from WebInject, it detects up to 99.80%, outperforming the baselines it is compared with.

arXiv:2610.30202 PDF

Consistency Is Not Validity: Self-Rewarded Moral Argumentation Makes Language Models Decisive but Pliable

Marius De Kuthy Meurers, Aaron Parisi, Bernd Bohnet The authors try to improve language models' moral judgments by rewarding good arguments. They train a reinforcement learning policy, starting from a strong language model, in a tournament where the policy itself judges pairs of arguments while its reasoning is masked. The training dilemmas are taken from classic novels, and a logic solver checks that each one is a genuine conflict. On MoralChoice, ELEPHANT, and MORABLES, the policy becomes more decisive and no longer depends on the order in which options are presented, but sycophancy and sensitivity to reframing remain. The authors conclude that consistency under reordering can be optimized without improving the moral judgment it is meant to signal.

arXiv:2610.30349 PDF

Probing and Steering for Safety-Relevant Behaviours in Large Language Models

Nimun Bajwa, Massimo Poesio Activation probes are directions in a language model's internal activations that can detect a behavior and, when added back into the activations, steer it. Using Llama-3.1-8B-Instruct, the authors systematically compare probe designs for three behaviors: deception, awareness of high stakes, and sycophancy. They cross two estimators, difference-of-means and logistic regression, with three token-pooling schemes and evaluate every direction both as a detector and as a steering vector. Prior work suggests logistic regression detects best while difference-of-means steers best; this holds for deception and high-stakes awareness but reverses for sycophancy, apparently partly because of the data. The best layer for steering can also sit well before the best layer for detection.

arXiv:2610.30363 PDF

SIM-UT: A User Simulation based Evaluation Framework for Medical Triage Chatbots

Deepali Mittal, Dheeraj Mundhra, Yulia Sidorova, Tersia de Kock, Oli Munns, Sofie S Meyer Large language model (LLM) triage chatbots are deployed in low- and middle-income countries, but they are mostly evaluated on single-turn, English-only knowledge benchmarks that ignore confused, terse or dialect-heavy users. SIM-UT (Simulation of User Behaviour for Triage) is a multi-turn evaluation method that simulates users grounded in the personas, constraints and communication patterns of real user populations. It was applied to three WhatsApp health chatbots in India, South Africa and Kenya, comparing MedGemma with ChatGPT and Claude across more than 5,000 multilingual simulated turns. Both model families showed de-escalation risk: after correctly flagging an emergency, they backed down once the user downplayed the symptoms, for example agreeing that vaginal bleeding in pregnancy was "not a problem".

arXiv:2610.30412 PDF

Agent Mosaic: Recognizing Confidentiality Leaks in Agentic LLM Systems

Daniel Ruiz, Luca Luceri, Emilio Ferrara cross-listed LLM agents that read several accessible documents can combine them and disclose a sensitive conclusion that no single document reveals, known as the mosaic effect. Using 30 synthetic cybersecurity, medical and military scenarios, the authors collect 10,800 answers from three agents in controlled simulation and 1,440 answers from live agents running in two open-source harnesses. Even for benign queries answerable from one document, the best model, GPT-5.6-Sol, disclosed forbidden composite conclusions in 7.08% of simulated answers, and live harnesses did worse. Visible classification headers and explicit aggregation rules lowered disclosure rates, but unsafe disclosures persisted in 32.37% of simulated answers across all query intents, and refusals often revealed the prohibited conclusion anyway.

arXiv:2610.30418 PDF

Secure Agentic AI and Multimodal Intelligence in 6G Communication Systems

Satyam Gaba, Devanshu Motwani, Siva Sai, Vineet Suri, Vikas Ranjan, Vinay Chamola Sixth-generation (6G) networks are expected to rely on agentic AI that reasons over multimodal inputs, calls tools and network functions, and coordinates across distributed entities, which opens new security and privacy risks. This survey sets out a 6G-oriented system view covering perception, memory, tool use and multi-agent orchestration. It analyzes threat models and attack surfaces across the sensing, cognition, action, memory and communication layers and arranges existing attacks into a structured taxonomy. It then reviews defenses, including prompt and memory protections, trust mechanisms between agents, secure orchestration and privacy-preserving learning, and closes with open research directions.

arXiv:2610.30436 PDF

Rethinking SAE Evaluation: An Atomic Interpretable Unit Framework Reveals Hidden Polysemanticity and Redundancy

Riteng Zhang, Xiaoqian Wang Sparse autoencoders (SAEs) are judged by a loose collection of metrics such as sparsity, dead features, and reconstruction, with no single objective explaining why those metrics matter. The authors propose that SAE features should form an "atomic interpretable unit basis", meaning a coordinate system suited to semantically meaningful explanation. They organize existing and new metrics, including DAME (Dominance-Aware Monosemanticity Evaluation), into a benchmark and apply it to hundreds of SAEs. They find that some SAEs look monosemantic under standard metrics but are functionally polysemantic: the features that contribute most to model computation are systematically less interpretable and more redundant.

arXiv:2610.30474 PDF

Model Organisms of Collusion: Corrective Transfer in Untrusted Monitors

Juan-Pablo Rivera, Ari Spiesberger, Juan J Vazquez, Arun Jose, Axel Ahlqvist, Adeline Kassler et al. AI control protocols sometimes use a copy of a model to monitor itself, which risks collusion, where the monitor hides its own misbehavior. To test whether fixing one caught collusion carries over to others not yet caught, the authors fine-tuned Llama-3.3-70B and Qwen2.5-72B to carry five collusion behaviors, then retrained the monitor on 300 honest reviews of just one. Under three of four correction methods, collusion on the four held-out behaviors fell from 91% to 14-20%, while a control fine-tuned on the same number of unrelated examples left 56-93%, and capability was largely preserved. Methods that performed similarly on the caught task differed in how far they transferred, so results on the caught behavior alone do not show how much a correction fixes.

arXiv:2610.30582 PDF

Interpreting Model Reasoning through Behavioral Counterfactuals

Shawn Im, James Oldfield, Adel Bibi, Sharon Li Sampled chain-of-thought responses vary a lot from run to run, which makes it hard to isolate how a change to the prompt alters a model's reasoning. The proposed method holds a response fixed and trains a module that edits the residual stream so the next-token distributions under the original prompt match those under a counterfactual prompt, minimizing the KL divergence between them. Matching output distributions recovers counterfactual behavior better than standard activation-reconstruction objectives, and attributions over the module's edits point to locations that activation patching confirms are causally important. Case studies on Gemma-4-E4B-it and DeepSeek-R1-0528-Qwen3-8B find that a suggestion in the prompt can steer reasoning without being acknowledged, acting through representations at the question statement and filler tokens.

arXiv:2610.30630 PDF

Scaling Laws for Safety Probes

Yuchen Wu (University of Washington), Jerry Wei (Anthropic) Safety probes are classifiers trained on a model's internal activations to flag harmful content, and they must be retrained for each new model or harm area with no advance estimate of how well they will work. The authors build six new large, unsaturated safety datasets and train two probe architectures on 20 base models from five families, using 32 to one million labels. Probe error (1 minus AUROC) follows two simple scaling laws. The lowest reachable error falls as a power law in model size, and error approaches that floor as labels increase, faster for larger models. Laws fitted on the smaller models of a family predict the next larger model's error to within 0.02 in 72% of cases, and in 73% of cases they correctly identify whether more labels or a bigger model is the cheaper way to reach a target error.

arXiv:2610.30678 PDF

Disentangling Robustness and Sycophancy in Multilingual LLMs

Rhitabrat Pokharel, Ameeta Agrawal When a multilingual LLM accepts a user's false premise, it may be sycophantic, or it may simply not know the answer in that language, and current evaluations do not tell these apart. The SC-matrix crosses competence (answering correctly without pressure) with sycophancy (keeping that answer under pressure) to separate the cases. Identical aggregate sycophancy rates can hide very different mixes of failures, high competence in a language does not always protect against sycophancy, and reasoning-oriented mitigations work better than directive-based ones.

arXiv:2610.30696 PDF

Faithful Latent Harm Monitors Work No Better Than the Models Own Behaviour at a Matched False Alarm Rate

Abhishek Mishra, Sagnik Chatterjee, Suyash Maniyar Latent safety monitors such as Latent Guard classify prompts as harmful from a model's internal activations, aiming to catch harmful intent even when the model complies. Across 15 instruction-tuned LLMs and 38 activation-based readers, the authors test these monitors on out-of-distribution rewrites of harmful and benign prompts and compare them with TF-IDF and sentence-embedding classifiers, standalone safety classifiers, and each model's own refusals. Activations read at the end of the instruction are largely independent of refusal behaviour but degrade under distribution shift, to the point that a TF-IDF classifier ranks harm better than the original Latent Guard in every model tested. Activations read just before generation catch 44% of harmful prompts versus 32% for the model's own refusals at a matched false-alarm rate, but that gain follows the refusal decision itself rather than an independent harm signal.

arXiv:2610.30727 PDF

UniGuard: Multi-View Guardrail Learning for Cross-Interface Safety Generalization

Viet-Thanh Pham, Thuy-Trang Vu, Junae Kim, Rollin Omari, Dinh Phung, Gholamreza Haffari et al. Guardrail models are usually trained on fixed prompt or response formats and become unreliable when the same safety case appears as a multi-turn conversation, an agent trajectory, a multi-agent trace, or a policy-conditioned input. UniGuard learns interface-invariant safety representations from multiple aligned views of each case, using the UniGuard-FT dataset, which rewrites 1,000 public safety cases into five interface formats. UniGuard-0.6B is trained with classification, supervised contrastive, and asymmetric confidence calibration losses. On the unseen-interface benchmarks ATBench, PolicyGuardBench, and DynaBench, it improves over its base guardrail by 45.12 F1 points on average and stays competitive with much larger models such as Claude-Sonnet-4.5 and GPT-5.2, despite training only on synthetic data.

arXiv:2610.30742 PDF

Invisibilization as a Source of Probability Bias in Language Models

Christian Arnold, Ernesto Calvo, Guadalupe Gonzalez, Sebasti\'an Vallejo Vera, Tiago Ventura Language models are aligned to avoid using protected attributes such as race, and as a result they often refuse to weigh information when those attributes are mentioned. The authors measure the cost of this in a two-urn game that has an exact Bayesian answer, running 6,600 sessions with leading OpenAI and Anthropic models in which models judge an unverifiable report from a reporter described in ten different ways. Models frequently declined to assess the reporter at rates driven by the description rather than the evidence, declining a Black-described reporter 46% of the time versus 9% for a white-described one. Audits that compare only average trust miss this bias, which grows with reasoning effort and gets worse under warning instructions. Scoring the models' trust against revealed outcomes over repeated rounds shrinks the racial gap in trust to below 0.02 in four of five configurations.

arXiv:2610.30785 PDF

What Lies Beyond Perception: A Partition Theory of the Limits of AI Evaluation and Oversight

Ahmed M. Mahran Benchmark scores can change when the judge model changes, and models can be trained to hide capabilities, which raises the question of what an evaluator can actually certify. The authors develop a theory in which an agent's ability to discriminate is the partition of its inputs into classes it cannot tell apart, and they prove that any examinee at least as discriminating as its evaluator can pass every test that evaluator can grade. As a result, passing a test certifies only a lower bound on ability, never an upper bound, and exact imitation of a weaker model cannot be detected by any behavioral evaluator. The bound extends to ensembles of monitors and helps explain published findings on judge disagreement and sandbagging, implying that safety arguments based only on behavioral evaluation must assume the model cooperates.

arXiv:2610.30803 PDF

Training Mixed-Monotone Neural Networks for Verification Without Set Propagation

Jonas Fingerhut, Lukas Koller, Matthias Althoff Formal verification of neural networks usually requires enclosing the network's output set, and computing a tight enclosure is hard in general. The authors propose an architecture that represents mixed-monotone functions as the difference of two monotone subnetworks, so a tight enclosure takes only two forward passes at opposite corners of the input interval. Because the enclosure size is a closed-form function of the weights, it can be minimized directly during training, and the authors prove a lower bound on it based on generalized total variation. On regression and dynamical-system benchmarks, training brings the enclosure to within 0.5% of the lower bound without hurting accuracy. Computing it is orders of magnitude faster than CROWN and α-CROWN, and the result is tighter than CROWN and zonotope propagation on deep networks.

arXiv:2610.30853 PDF

Tracing Subliminal Learning Through Pretraining

Angela Tang, Vishesh Anand, Gon\c{c}alo Paulo Subliminal learning is when a language model trained on another model's outputs picks up traits unrelated to that data, such as a preference for an animal. Using Pythia, PolyPythia, and DataDecide OLMo models, the authors show it persists in both hard-label distillation on teacher-generated numbers and on-policy soft-label distillation, across different initializations, data orders, pretraining corpora, model sizes, and architectures. Across pretraining checkpoints, a model's ability to transmit a preference and its ability to acquire one develop separately, so transfer can be strongly asymmetric. Synthetic number sequences that only match the teacher's number frequencies still induce substantial transfer, and soft-label distillation transfers more consistently with less training.

arXiv:2610.30884 PDF

Verbalizing Activation Differences with Natural Language Autoencoders

Andrea Blasi N\'u\~nez, Gianluca Barmina, Federico Torrielli, Thiusius Rajeeth Savarimuthu, Peter Schneider-Kamp, Lukas Galke Poech Natural language autoencoders (NLAs) learn to describe an LLM's internal activations in text by reconstructing the activations from a natural-language description. Across four models from three families, the study asks whether these descriptions generalize: whether they compose sensibly, whether they transfer to fine-tuned versions of the original model, and whether they can reveal what a fine-tune changed (model diffing). The answers are mostly positive. Descriptions of analogy-arithmetic activations name the target concept, an NLA works on lightweight fine-tunes without retraining, and describing the activation differences between Taboo model organisms and their base models recovers the hidden secret word for most organisms, with reinforcement learning on activation differences improving recovery further. NLAs can reveal the fine-tuning objective without the model generating any tokens, and they fall short of state-of-the-art diffing methods at detecting emergent misalignment but beat a black-box baseline.

arXiv:2610.30934 PDF

Measuring the Reasoning-Action Coupling in Driving Vision-Language-Action Models through Counterfactuals

Gabriel Manalu, Saad Wazir, Thao Do, Seongah Kim, Hyeontaek Hwang, Tolcha Yalew Kidane et al. Driving vision-language-action (VLA) models explain their maneuvers in text, which assumes that the trajectory follows the stated reasoning and that the reasoning reflects what actually drives the trajectory. The authors test both assumptions on four driving VLAs with causal counterfactual interventions. They perturb the models' own reasoning text, and they photorealistically remove key objects that the reasoning mentions. Changing the meaning of the reasoning shifts trajectories only slightly compared with blanking the camera inputs, which indicates that the stated reasoning is weakly coupled to driving behavior and limits the explainability these systems are meant to offer.

arXiv:2610.30953 PDF

Conflict-Aware Logit Adapters for Utility-Preserving Anti-Distillation

Quoc Phong Dao, Binh Anh Nguyen, Linh Ngo Van, Thanh-Toan Do, Mehrtash Harandi, Trung Le Reasoning traces from a teacher model make it easy for others to distill a student from it, and existing decoding-time defenses are slow and hurt fluency. CALA (Conflict-Aware Logit Adapters) trains only a small residual adapter on a frozen teacher. A proxy student identifies the positions where imitation is most sensitive, and the adapter perturbs the output distribution there while staying within high-probability regions. A conflict-aware optimization balances the anti-distillation goal against utility. On reasoning benchmarks it substantially reduces successful student imitation while keeping near-original teacher performance, with modest inference overhead and no auxiliary model.

arXiv:2610.31046 PDF

Harm Is Behavioural, Moderation Is Lexical: How Language and Culture Decide Whom Automated Safety Systems Protect and Whom They Silence

Gabriel Lars Sabadin, Dr Nicola Harding, Sarah Kingston Automated content moderation both over-flags (labeling African American English as offensive up to twice as often, for example) and under-protects speakers of low-resource languages. The authors argue that both problems share one cause: judging harm by vocabulary rather than by what a message does. They organize the evidence into seven failure modes and propose an evaluation protocol that makes the gap between the best- and worst-served languages a headline metric. Applied to five systems, from a keyword lexicon to multilingual guard models, across eleven languages, 51 of 55 system-language pairs fall short of native-speaker accuracy. More capable systems widen the gap, leaving Arabic and romanised Hindi near chance.

arXiv:2610.31047 PDF

SCOPE: Compiling Natural-Language Policies into Neurosymbolic Predicate Programs

Arnon Mazza, Elad Levi Enforcing compound, application-specific policies over multi-turn LLM agent transcripts and tool calls is hard: fine-tuned guardrail models degrade on such policies, and LLM-as-a-judge is costly and unstable. SCOPE compiles a natural-language policy into an executable program of neural predicates, run by small models and deterministic extractors and composed by code. The program is optimized jointly with its training data through adversarially generated counterexamples. It achieves the highest average accuracy against prompt-optimization, decomposition, and workflow-generation baselines. Programs optimized with the weakest executor transfer to other models, which lets inexpensive models reach accuracy comparable to frontier reasoning models with more stable and interpretable verdicts.

arXiv:2610.31062 PDF

What Pairwise Evaluation Misses: Reward Scores across Prompts

Mateusz Modrzejewski, Tomasz Radzikowski Reward models are normally evaluated pairwise, by comparing two responses to the same prompt, so any part of the score that all of a prompt's responses share cancels out and is never tested. That shared part matters whenever scores are pooled across prompts for filtering or ranking. The authors build a human-referenced audit that combines a between-prompt variance share, a sibling-rating regression and prompt substitutions, and apply it to eighteen released scorers and six matched training conditions. On HelpSteer2, six reward models put 0.34 to 0.64 of their score variance between prompts, while human ratings put close to none there. Some of that variation is still useful for selecting responses, so the authors recommend checking between-prompt variation, sensitivity to a wrong prompt and usefulness for selection as separate properties.

arXiv:2610.31090 PDF

Post-Audit Mirage: Identifiability Limits of Offline Agent Verification

Aman Behera, Payal Payal Offline audits may not reveal whether an update to an AI agent will stay safe after deployment. The authors call this the 'post-audit mirage' and demonstrate it with matched safe and harmful scenarios that give the verifier identical evidence but have opposite outcomes in deployment. A theorem built on standard observational-equivalence arguments formalizes an unavoidable tradeoff between making errors and abstaining from a decision. A benchmark spanning authorization, scheduling and shared-capacity environments makes the problem concrete. A proposed monitor, IdentifiedRangeMonitor, uses fresh harm outcomes labelled by group to choose between deploy, hold, cannot determine and unsupported, and it can resolve the cases in favourable settings but stays undecided when groups are rare, effects are small or observations are missing.

arXiv:2610.31130 PDF

Investigating Multi-Cultural Value Selection Biases in LLMs

Siddhesh Pawar, Arnav Arora, Dustin Wright, Dirk Hovy, Isabelle Augenstein Cultural evaluations of large language models (LLMs) typically assign each user a single identity label, although real users have layered identities. The authors give models a country persona in the system prompt plus a second identity cue in the user turn, either by naming a country or through the language of the query. They test with World Values Survey (WVS) multiple-choice questions and 1.6 million open-ended responses. In multiple choice, language cues shift answers about three times more than named countries, yet move them away from that country's actual value distribution. In open-ended generation, naming a country closes 30 to 44% of the gap to the model's own responses for that country, which suggests language and country cues should be treated as distinct signals.

arXiv:2610.31199 PDF

Harmful Responses Are Jailbreak Templates: Transferring Autoregressive Jailbreaks to Diffusion Language Models

Jinwei Gan, Zifeng Cheng, Zhiwei Jiang, Ran Sun, Shiping Ge, Qing Gu Diffusion large language models (dLLMs) generate by filling in masked tokens using context on both sides, which opens a jailbreak attack surface that autoregressive (AR) models do not have. The authors find that a harmful response with some of its tokens masked works as an effective jailbreak template, because the dLLM reconstructs the harmful content. Their method, RMT-Jailbreak, takes a harmful response produced by an existing AR jailbreak, masks most of its tokens with uniform retention, filters out any remaining visible harmful text, and gives the template to the target dLLM. Experiments across several AR jailbreak methods and target dLLMs show that the attacks transfer strongly between the two model types.

arXiv:2610.31331 PDF

AfterDeleteBench: Auditing Post-Deletion Information Persistence in LLM Agents

Ehsan Saleh, Roozbeh Nahavandi, Sama Salarian, Raef Bassily, Andrew Perrault cross-listed LLM agents with persistent memory often derive extra state, such as summaries, from past interactions, so deleting an original record may leave its information alive in those derived copies. AfterDeleteBench measures this by comparing agents that saw and then deleted a synthetic secret with matched agents that never saw it. A single query recovers the deleted secret from Qwen2.5-7B-Instruct in about two thirds of cases, against zero for the controls. Even with no verbatim copy left, exact reconstruction from separately stored components succeeds in 62-99% of cases across Qwen, Mistral, and Phi-3. Provenance-aware closure, which also removes everything derived from a deleted record, eliminated recovery in the tested settings, while an exact-string write filter only stops the secret from being written back.

arXiv:2610.31366 PDF

RIDAR: Safeguarding Research Integrity in Autonomous Research Agents via Automated Detection and Runtime Intervention

Jiening Siow, Amy Xin, Junjie Wang, Ziyu Tang, Zijun Yao, Yunjia Qi et al. Autonomous research agents can commit research integrity violations (RIVs), behaviors that undermine the validity, reproducibility, or faithfulness of the research they automate. The authors define RIVT, a taxonomy of 14 violation types across the research lifecycle, and build RIDAR-post-hoc, which detects violations in agent trajectories at 85.2% F1 on the expert-annotated RIV-Gold benchmark. Applied to 300 trajectories across 50 research tasks in RIV-Wild, it finds 3,964 violations, present in 99.7% of trajectories. A runtime version, RIDAR-runtime, detects and corrects violations while the agent works, cutting the violations left at task completion by 36.8%.

arXiv:2610.31372 PDF

Certified Execution Harness for Agentic AI: Safe and Least-Intrusive Petri-Net Supervisory Control Between Reasoning and Action

Chao Gu, Quanyan Zhu To defend LLM agents against prompt injection, the authors model the execution harness between the model and its tools as a supervisory control problem over Petri nets, where tokens represent capabilities, approvals, and budgets. Because an action fires only when the tokens it needs are present, an injected request with unmet or already spent preconditions is refused regardless of its wording. They introduce a replanning closed loop, in which an agent can propose again after a refusal at the cost of another inference call, and show that the conventional maximally permissive supervisor can be arbitrarily more intrusive than a stricter safe one in this loop. They characterize the least-intrusive safe supervisor as a value-threshold refinement of the conventional one and demonstrate synthesis on an autonomous release agent.

arXiv:2610.31376 PDF

AgentPhishing: A Benchmark of LLM Email Agents Under Adaptive Attack

Justin Tan, Miao Li, Ding Zhao cross-listed Existing phishing datasets were built for conventional detection and say little about how LLM email agents respond to persuasive malicious requests. AgentPhishing is a benchmark for this question. Alongside it, PhishOpt optimizes reusable email-generation skills against individual agents under realism constraints, and analysis of the resulting emails through Cialdini's persuasion principles is used to build a universal attack skill. Agents that safely handle pre-existing phishing emails remain vulnerable to adapted attacks: four of nine agents detect fewer than 10% of phishing attempts and comply with over 45% of malicious requests.

arXiv:2610.31404 PDF

Diffusion Language Models for Perplexity-Aware Black-Box Attacks

Ali Al-Lawati, Elsayed Eshra, Dongwon Lee, Suhang Wang Continuous-space black-box attacks on LLMs search a Gaussian latent space and decode prompts through a random mapping, which yields unnatural text that input guardrails catch easily. The authors replace that mapping with a continuous diffusion language model (dLLM) and keep the same latent formulation and search procedure. In sampling-based RareTrap and optimization-based ThinkTrap, character-level perplexity falls from 10³–10⁷ to 3–5 and from 20–28 to about 4.3–5.4, respectively. Risk estimates also differ by up to an order of magnitude between the two mappings.

arXiv:2610.31558 PDF

Who Is Sycophantic to Whom? Confident LLMs, Human Deference, and the Collapse of Trust Calibration

Deepthika S, Pawni Yadav, B. Sankar Sycophancy research usually asks whether models tell users what they want to hear; this study asks the reverse, whether users defer to a model that is confident, consistent and wrong. In a within-subjects experiment, 32 participants learned science topics from five GPT-5-mini assistants whose briefings had 0 to 100% of their facts secretly replaced with plausible falsehoods, then took surprise quizzes with confidence ratings. Quiz scores fell from 7.4 to 2.9 out of 10, and falsehoods stated by the assistant were adopted about 65% of the time at every falsification level. Participants held these absorbed falsehoods with high confidence, their confidence stopped distinguishing right answers from wrong ones (AUROC fell from 0.70 to chance), and their trust ratings barely moved. The authors propose a baseline-corrected Human Deference Index to measure this human-side sycophancy.

arXiv:2610.31579 PDF

Simulated Control Evals

Jeffrey Olmo, David Zhang, Tyler Tracy Control evaluations test whether safeguards can stop an AI model that is deliberately trying to cause harm, but they usually run in purpose-built settings that may not reflect a real deployment's environments and tasks. Simulated control evaluations (SCEs) replace the real environment with a language model that simulates tool responses, so evaluations can be run against deployment-specific tasks. Validated on LinuxArena, where real execution provides ground truth, SCEs predicted safety with 3.6 percentage points of root mean squared error, against 10.1 for the separate BashArena setting, and they closely reproduced attack strategies and monitor reasoning. Attack strategies found in simulation also improved monitors on held-out real attacks about as much as strategies found through execution.

arXiv:2610.31593 PDF

OP-Print: Tracing Membership Imprints in On-Policy Self-Distillation

Qi Li, Xinchao Wang On-policy self-distillation trains a model on its own outputs, enhanced with extra context or demonstrations, and the authors present the first study of whether it leaks training-set membership. Their attack framework, OP-Print, builds on the finding that the supervision signal is sparse and concentrated at uncertainty-sensitive correction points rather than spread across the whole rollout. In the gray-box setting it aggregates the teacher-student correction along a rollout, and in the black-box setting it measures how stable outputs stay under restarts as a proxy for that hidden correction. Baseline membership inference attacks stay near random guessing even with gray-box access, while OP-Print is effective in both settings across models and distillation schemes.

arXiv:2610.31616 PDF

Questioning Instrumental Convergence in RLVR: Checkpoint Evidence from OLMo-3

Bar{\i}\c{s} Alparslan, Sandy Tanwisuth, Caleb Petterson Instrumental convergence predicts that capable agents will tend toward power-seeking subgoals, and reinforcement learning with verifiable rewards (RLVR) is the post-training stage where such behavior would most plausibly appear. The authors evaluate OLMo-3-7B-Think with InstrumentalEval across supervised fine-tuning, DPO, 13 intermediate RLVR checkpoints and the final model. The instrumental behavior rate rises from 25.0% after supervised fine-tuning to 46.1% near RLVR step 200, then falls to 11.8% at the final checkpoint, and prompt framing explains little of this pattern. Preliminary 32B results show a similar drop, and the authors caution that the findings are specific to this model family and benchmark and that judge agreement is only moderate.

arXiv:2610.31624 PDF

Reliable AI-Assisted Decisions via Selective Auditing

Isaac Gibbs, Elise L. Han, Ryan J. Tibshirani, Tijana Zrnic cross-listed The setting is online monitoring of an AI system that proposes a decision each round together with a confidence score, which may come from a separate verifier agent. A monitor must decide each time whether to accept the proposal automatically or send it to an expert, such as a human auditor. The authors frame this as online constrained optimization: send as few cases to the expert as possible while keeping the fraction of AI errors that go unaudited below a user-set threshold. They derive algorithms with provable error-control guarantees that need no assumptions about the data, including one that asymptotically achieves the best possible auditing rate for both adversarial and i.i.d. data streams. Experiments on screening harmful requests to LLMs, qualitative text annotation, and data labeling show valid error control with much less human auditing than baselines.

arXiv:2610.31662 PDF

9 more specialized papers

Theory 36

A Quadratic Lens on Muon's Polar Update: Benefits, Pitfalls, and the Role of Newton-Schulz

Egor Shulgin, Sam Laing, Antonio Orvieto, Peter Richt\'arik The Muon optimizer updates matrix parameters along approximately orthogonalized directions and often trains LLMs faster than AdamW, but it is unclear when and why this helps. Analyzing PolarGD, the orthogonalized update at Muon's core, on matrix quadratics with line search, the authors show that it helps when the gradient is far from low rank and its top singular directions have high curvature, and that it acts as gradient descent preconditioned by the current iterate. This means tiny singular values in the iterate can make it much slower than plain gradient descent. The few Newton-Schulz iterations Muon actually uses treat those small singular values the way gradient descent does, so fewer iterations can converge faster than exact orthogonalization, and in a transformer trained with Muon about a quarter of the attention matrices' singular values fall in this regime.

arXiv:2610.30280 PDF

Lifetime-valid decisions by continually evolving AI analysts

Xiaojing Sun, Yuhan Zeng, Xiao Wang cross-listed AI agents acting as analysts often make decisions that amount to hypothesis tests, but because these agents are random and keep changing, the chance of deploying a version whose false-positive (Type I) error rate is too high builds up over time. The proposed continual validation tests each new version before deployment on simulated null-hypothesis data, using scenarios that can be randomized and expanded over time. The authors prove that the lifetime probability of ever deploying an invalid version stays bounded, however updates are chosen. In bioequivalence and A/B-testing examples, Type I error rates of 0.098 and 0.183 without validation fell to at most 0.045 and 0.044 with it, while keeping 82% and 96% of attainable statistical power.

arXiv:2610.30562 PDF

Reverse-KL Policy Distillation: Statistical Guarantees and Limits

Haolin Liu, Chen-Yu Wei, Julian Zimmert Reverse Kullback-Leibler (KL) distillation is widely used to train language models, but its statistical properties are poorly understood. The authors analyze sample complexity in two settings: online distillation on student-generated trajectories and offline distillation on teacher-generated ones, each with either the teacher's full action distribution or only the probability of a single sampled action. They prove minimax bounds and show that online distillation with only sampled-action probabilities may need substantially more data, whereas offline distillation reaches the same rate under either feedback model, because teacher-chosen actions themselves reveal the teacher's preferences. They also show that common online objectives can fail, and they propose trajectory-loss and importance-ratio truncation techniques that achieve optimal online guarantees.

arXiv:2610.30844 PDF

The Geometry of Meaning in Neural Manifolds

Valentino Maiorca Neural networks trained separately can learn similar functions while expressing them in incompatible parameterizations, which hides the relationships between the models. This thesis argues that semantic correspondence, meaning the identification of which elements in two systems refer to the same content, is enough to align them in a shared frame and transform between independently trained models without additional training (zero-shot). Fourteen contributions apply this idea to model stitching, cross-modal translation, policy transfer, activation steering and neural decoding, covering vision, language, reinforcement learning and neuroscience, and model sizes from small supervised networks to billion-parameter generative models. The same principle extends to fMRI data, where different human brains processing the same stimuli show geometric structure that can be aligned.

arXiv:2610.30938 PDF

Transformers with CoT Discover Heavy-Ball Acceleration Through Training

Yuxuan Cai, Yuan Cao Earlier theoretical work showed that transformers whose chain-of-thought (CoT) steps are supervised by a gradient-descent teacher learn to carry out gradient descent for in-context linear regression. This work asks whether CoT reasoning can go beyond imitating its teacher. After initial supervision from gradient descent, the model is given further trajectory-level supervision that uses only the true regression weights. Under that training the model learns to implement the Heavy-Ball momentum method and converges faster than its teacher, even though no Heavy-Ball teacher was ever used. The authors present this as a concrete theoretical case of transformer reasoning discovering a new algorithm through training.

arXiv:2610.31089 PDF

The Bayesian Learning Rule Beyond KL Geometry

Sophia Sklaviadis, Wu Lin, Andre Martins, Mario Figueiredo, Thomas Moellenhoff, Emtiyaz Khan The Bayesian learning rule (BLR) derives efficient additive updates for exponential-family variational posteriors by relying on Kullback-Leibler (KL) divergence geometry, which can produce undefined objectives when the posterior's support depends on its parameters. Earlier work suggested that KL is the only divergence that keeps these additive updates, but the authors show that mixed-space Bregman or Fenchel-Young discrepancies also preserve them, with the classical BLR as a special case. Using Tsallis negentropy, the generalized rule yields q-Gaussian posteriors with bounded support or heavy tails through a simple Newton-like update. Experiments up to CIFAR-100 show controllable support and tail behavior with competitive performance.

arXiv:2610.31308 PDF

30 more specialized papers

Vision 24

EgoHand: Monocular 3D Hand Reconstruction from Egocentric Video

Chao-Yeh Chen, Tiffany Chen, Lin Duan, Tianyi Hu, Yu-Hsiang Huang, Umar Iqbal et al. cross-listed EgoHand is a model that reconstructs 3D hands from monocular egocentric video. It is trained on a curated corpus of 14 million images that mixes synthetic data, in-the-wild video, and multi-camera motion capture. It uses SOMA-Hand as its parametric hand model and runs a transformer encoder over 2D hand crops, with camera cues and an auxiliary 2D branch to improve alignment with the image. In zero-shot tests on HOT3D, ARCTIC, and H2O, it has lower joint-position, root-translation, and acceleration errors than baselines at about 9.69 ms inference latency. In a blinded user study, participants preferred its poses to those from Apple EgoDex and Ropedia Xperience-10M, and the model is released under a license that permits commercial use.

arXiv:2610.30293 PDF

LoRA Superposition across Pretrained Models

Keigo Watanabe, Hiroshi Kera, Kazuhiko Kawamoto cross-listed Low-rank adaptation (LoRA) adapters are normally tied to the backbone model they were trained on. Here, LoRAs trained independently for the same task on different backbones are summed, and the sum is applied unchanged to every contributing backbone, a construction called LoRA superposition. Across seven ViT-B/16 CLIP backbones and eight image-classification tasks, the summed adapter improves all 56 backbone-task pairs and comes within 0.28 percentage points of individually trained LoRAs, while a sum that leaves out the target backbone's own LoRA barely helps. The per-backbone LoRAs are nearly orthogonal in parameter space, which suggests each frozen backbone implicitly picks out its own component.

arXiv:2610.30691 PDF

Learning from Next-Frame Prediction: Towards Unified and Scalable Visual Representation Learning

Yifeng Yao, Jinghan Li, Zhihao Zhang, Yajiao Xiong, Yang Jin, Kun Xu et al. cross-listed Masked video pretraining recovers missing content rather than predicting the future, while autoregressive methods have trailed joint-embedding approaches in representation quality. VANE is a self-supervised framework in which a causal encoder summarizes past frames and a state predictor infers the representation of the entire next frame in the fixed embedding space of a frozen semantic tokenizer. A lightweight auxiliary decoder reconstructs future-frame patches to keep fine spatial detail, and still images are handled as local-to-global crop sequences so one encoder covers both images and video. The model improves with scale and reaches state of the art among unified encoders on most anticipation, recognition, and dense perception benchmarks, including a 7.1-point gain in mean-class recall@5 over V-JEPA 2.1 on EPIC-KITCHENS-100 action anticipation.

arXiv:2610.30724 PDF

Direct Forcing: A Single-Stage Training Recipe for Few-Step Autoregressive Text-to-Video Generation

Haojun Xu, Jie Huang, Xin Lu, Mingchen Zhong, Zihao Fan, Linjiang Huang et al. cross-listed Few-step autoregressive video generators enable low-latency streaming, but training them normally separates causal adaptation from distribution matching into two stages. Direct Forcing merges these into a single stage that uses reference videos in two ways. Causal Continuation Alignment learns how a reference history continues, and Rollout Distribution Matching uses maximum mean discrepancy to pull the model's own rollouts toward the distribution of reference videos, with Dynamic Gradient Balancing weighting the two signals. Starting from bidirectional Wan2.1-1.3B, it learns a four-step autoregressive generator in only 250 updates, scoring 84.63 VBench Total without a separate initialization stage or online diffusion-score supervision.

arXiv:2610.30861 PDF

Query-Time Calibration of CLIP Features for Open-Vocabulary 3D Segmentation

Dominik Engel, Yousef Rajeh, Leon Sick, Ivan Viola cross-listed Open-vocabulary 3D segmentation often lifts CLIP features into 3D Gaussian reconstructions and compares them with text by cosine similarity. That comparison ignores the fact that a scene's lifted features tend to cluster strongly in a few directions. The method changes only the comparison at query time: it partially subtracts a shared scene direction from both visual features and text prototypes, whitens the text prototypes against a fixed vocabulary of generic terms, and then uses soft clustering for multiclass prediction and CSLS with a relevance margin for single queries. Across seven lifting methods and four datasets, mean intersection over union improves consistently in both multiclass and single-query segmentation without retraining, re-encoding source images or re-lifting features.

arXiv:2610.30908 PDF

Inside-JEPA: From Prediction to Relational Grounding in Latent World Models

Hossein Resani, Javen Qinfeng Shi Joint-embedding predictive architectures (JEPAs) learn world models in latent space that are efficient for planning but hard to interpret. Inside-JEPA adds relational channels for ordered pairs of entities to the latent state and grounds them without relation labels by using an inverse-dynamics loss that recovers actions from changes in those channels. On Something-Something V2, latent prediction alone does no better than a random relation head, while inverse-dynamics supervision improves relational AUROC by 0.065 to 0.103, with similar gains on RH20T robot episodes. When asked which action achieves a one-step relational goal, a readout of the 16 channels places the correct action in the top five of 65 in 43% of cases, versus 24% for prediction-only latents.

arXiv:2610.31194 PDF

18 more specialized papers

Multimodal 14

AI Mindmeld: A Latent Interconnect Protocol

Siavash Golkar, Jake Kovalic, Shirley Ho When LLMs act as scientific agents, specialist models must pass information to them as text or other media, which is a bottleneck between rich internal representations. The proposed Latent Interconnect Protocol (LIP) trains only small connectors that map specialist representations directly into a frozen LLM's internal states, so new specialists can be added without retraining the LLM or existing connectors. Connecting three protein and three small-molecule specialists to Gemma 3 (4B and 12B) and Kimi K2.6, the authors show that independently trained connectors answer familiar and new question types about unseen objects and transfer zero-shot to cross-domain tasks without joint training, with gains growing as the LLM gets more capable. They note that current LLMs are not trained to use such inputs, so access does not guarantee the information is used reliably in reasoning.

arXiv:2610.30215 PDF

Edit, Don't Imitate: Minimal-Edit On-Policy Distillation for Multimodal Reasoning

Wen Ye, Ailin Deng, Zirui Wang, Shixing Yu, Han Zhang, Tian Yun et al. Smaller vision-language models (VLMs) often fail at visual math because they misread the diagram, and accurate text descriptions of the figure close much of the gap but are costly to obtain at inference time. The proposed minimal-edit corrective on-policy distillation uses reference figure descriptions and verified solutions only during training: a teacher scores the student's incorrect rollouts under a minimal-edit objective so that only tokens the teacher disfavors contribute to the loss, while correct rollouts receive task-reward reinforcement learning. Without any descriptions at test time, trained models meet or exceed their base models given a reference description, improve low-level geometric perception, and transfer to external math and perception benchmarks. The method still works with a frozen teacher of the same size, though the gains are smaller than with a larger teacher.

arXiv:2610.30231 PDF

Text-Only LLM Correctors Break Correct Transcripts: A Repair-Damage Analysis of Audio Grounding in ASR Error Correction

Vasiliy Kudryavtsev, Kirill Borodin, Nikita Vasiliev, Mikhail Gorodnichev, Grach Mkrtchian cross-listed Large language models used to correct automatic speech recognition (ASR) output are usually judged by overall word error rate (WER), which hides what they fix and what they break. The authors decompose every WER change exactly into repaired tokens, damaged tokens, and insertions, and compare correction with and without access to the audio on three corpora. With Gemini 2.5 Flash, text-only correction of its own draft breaks 13-32% of already-correct transcripts, versus 1-3% with the audio, and on medical dictation, insertions cancel the net WER gain. The open-weight Voxtral rarely edits and does not show this pattern. Because the prompts differ between the two setups, the comparison bounds the effect of the audio rather than isolating it.

arXiv:2610.30385 PDF

Visual Grounding First, Multimodal In-context Learning Follows

Minhyuk Seo, Minjae Lee, Chaeeun Lee, Wei Lin, Muhammad Jehanzeb Mirza, Tinne Tuytelaars et al. cross-listed The authors argue that multimodal LLMs struggle with in-context learning (ICL) mainly because of weak visual grounding: they pay too little attention to the images in demonstrations and lean on textual cues. ProCoRe trains grounding with contrastive reinforcement learning. It captions contrastive image pairs, breaks the captions into propositions, and rewards alignment with the matching image while penalizing mismatches. Without ever training on ICL-formatted examples, it improves ICL classification accuracy by 5.6% and captioning ROUGE-L by 13.8% over Qwen-VL-3-8B, beating prior methods that depend on ICL data. The authors also introduce PICL, a personalized captioning benchmark that checks whether models actually read the demonstration images rather than relying on prior knowledge.

arXiv:2610.30592 PDF

VLTest: Finding Joint Image-Text Failures in Vision-Language Models

Bin Duan, Jintao Lin, Guowei Yang cross-listed Testing vision-language models (VLMs) is hard because changing an input can legitimately change the correct answer, and some failures only appear when the image and the text are both changed. VLTest is a black-box tester that creates image variants by editing VQGAN latent codes and text variants with a rewrite-expand-prune pipeline, and it discards variants likely to change the expected answer. It tests each variant on its own before testing combinations, so every joint failure can be attributed using four matched cases, and a VLM judge confirms each failure. Across three tasks and eight VLMs, the judge reaches 88.8% to 94.6% precision against human labels, and with the same budget VLTest finds 1.8 to 2.5 times as many confirmed failures as SINVAD and TextFooler.

arXiv:2610.30794 PDF

TRIAD: Benchmarking Omni-Modal Ambiguity in Multimodal Large Language Models

Zhaolu Kang, Yidi Wang, Yile Li, Siqi Zeng, Yingjie He, Richeng Xuan et al. Most multimodal benchmarks test recognition or reasoning on inputs that are already unambiguous. TRIAD instead tests whether omni-modal models can resolve ambiguity spread across text, image, and audio. It contains 327 hand-written text–image–audio items in 86 scene groups covering 18 ambiguity categories, and each item has a unique answer only when all three modalities are considered. Across 14 omni-modal systems, the best model still trails humans by more than 20 points in scene-level group accuracy, and without answer options model group accuracy drops to single digits while humans stay robust. Ablations show that models tend to underuse audio cues when text and image are also present.

arXiv:2610.30841 PDF

A Task-Centric Survey of Long-Document Visual Question Answering: Taxonomy, Evolution, and Future Directions

Yufei Tong, Guanjie Cheng, Tianle Wu, Leilei Zhang, Hang Zheng, Peihan Wu et al. cross-listed Existing surveys of visual document understanding are organized around architectures. This one is instead organized around the bottlenecks of the long-document visual question answering (VQA) task itself. It breaks the task into four components: document parsing, evidence organization and retrieval, generator organization and learning, and evaluation. For each it builds a fine-grained taxonomy, with a focus on explicit retrieval while also covering direct reading. A temporal analysis shows research priorities expanding from multimodal content understanding toward sufficient, high-fidelity evidence acquisition and sustained evidence management. Proposed future directions include evidence sufficiency, verifiable and correctable understanding, budget-aware coordination, and continual document understanding.

arXiv:2610.31385 PDF

Drift Says When, Budget Says How Much: Rate-Controlled KV Re-selection for Long-Form Multimodal Generation

Bin Hu, Yingbo Xiao KV-cache compression for multimodal LLMs usually picks which visual tokens to keep once, at prefill. The authors show this fails on long-form generation: even an oracle one-shot selection reaches only 0.44-0.52 ROUGE-L agreement with uncompressed output on detailed image description. Their method, Rate-Controlled Refresh (RCR), uses attention drift to decide when and where to re-select tokens, while a per-sample controller sets how often to refresh. A low-rank drift detector keeps the cost low. With a single setting for all models and tasks, RCR beats the per-step retrieval method InfiniGen by 1.6-3.8 ROUGE-L points on five of six settings, stays within 1.7 points of exact per-step re-selection, and uses a 4x smaller GPU-resident detector.

arXiv:2610.31433 PDF

6 more specialized papers

Reasoning 11

Detection Is Not Enough: Reasoning Traces as Testable Program Hypotheses

Vaidehi Hegde cross-listed In deployment no reference solution exists, so the authors ask how far a single reasoning trace can be verified with no reference, no judge model, and no weight updates, using a frozen LLaMA-2-7B-chat on GSM8K. Checking a trace against itself plateaus at an AUROC of 0.741 because two thirds of wrong traces contain no false equation: they correctly execute a wrong plan. A metamorphic probe that turns the trace into a program, perturbs one quantity, and re-asks the model raises the accuracy of supported traces to 60.6% against a 31.6% base rate, but it also rejects 45.5% of correct traces. No tested retry policy significantly improves accuracy, and because retries repair so few wrong traces, the authors estimate that even a perfect detector would recover only about 66 of 1,319 questions, making repair rather than detection the bottleneck.

arXiv:2610.30200 PDF

Route-and-Generate: KL-Regularized Router Improvement for Reasoning Routing in LLMs

Tien-Phat Nguyen, Truong Nguyen, Thanh-Thien Le, Hoang Phan, Thien Huu Nguyen, Linh Ngo Van et al. Large language models often overthink easy problems and underthink hard ones. Route-and-Generate (Rouge) separates the decision about how much to reason from the generation of the answer. A routing policy splits a fixed sampling budget across several reasoning formats, while the base model is trained with GRPO using a rank-based reward that favors shorter correct solutions. The router is updated with a closed-form KL-regularized policy improvement step and distilled into a neural router, so the model gives concise answers on easy problems and reserves long reasoning for hard ones. The authors report better accuracy-efficiency trade-offs across several reasoning benchmarks.

arXiv:2610.30351 PDF

Same Beliefs, Different Policies: Learning Reusable Representations for Early Stopping in LLM Reasoning

Jingwei Ni, Dmitry Abulkhanov, Vladislav Smirnov, Markus Leippold, Elliott Ash, Timothy Baldwin et al. Reasoning LLMs often keep reasoning after reaching a correct answer, which wastes tokens and sometimes replaces the right answer with a wrong one. Stopping controllers trained directly at a fixed token price need retraining for each new price. BELT (BELief-based Termination) instead learns price-independent beliefs about current and future answers from a frozen model's hidden states, then fits a linear stopping policy on CPU for any accuracy-cost trade-off. Across five out-of-distribution math benchmarks it cuts generated tokens by 41% on Qwen3-1.7B and 51% on OLMo-3-7B-Think without losing accuracy. The math-trained probe transfers to coding and agentic tasks, and it needs 55x fewer training GPU-hours than the five price-specific checkpoints of the OS-Pruner baseline.

arXiv:2610.30450 PDF

LLM2Cert: LLM Synthesis of Certificates with SMT/Lean Verification

Zhipeng Cao, Peiyang Song, Nikolaus Vertovec, Weiru Cui, Zhihao Yang, Mingkang Liu et al. cross-listed Lyapunov functions and barrier certificates prove the stability and safety of dynamical systems, but finding and verifying them for high-dimensional nonlinear systems is hard, and existing benchmarks favor low-dimensional polynomial systems that solvers can handle. The authors introduce DynaCertBench, 105 instances spanning stability, region-of-attraction, invariance, and safety tasks, along with LLM2Cert, which has an LLM propose certificates and routes each proof obligation either to an SMT (satisfiability modulo theories) solver or to a Lean proof-repair agent. LLM2Cert verifies 81 of 105 instances (77.1%), 17.1 percentage points more than one-shot proposals without feedback, and verifies 22 certificates that SMT solvers alone leave unresolved.

arXiv:2610.30524 PDF

GIFT: Gradient Informed Filtering for Distillation Augmented Reasoning RL

Shijie Li, Yilin Gao, Tieyuan Chen, Yun Luo, Weiyao Lin, Hang Yu Training LLMs for reasoning often combines sparse, outcome-based reinforcement learning (RL) with dense, token-level distillation from a teacher, but imitating the teacher can conflict with the RL objective. GIFT splits each token's distillation gradient into a component parallel to the RL gradient, which measures conflict and serves as a safety filter, and an orthogonal component, which measures new guidance and serves as a value score. Tokens are ranked by this score, and a coverage-annealing schedule gradually shrinks the set of distilled tokens, dropping the most conflicting first, until training is pure RL. Across model scales on math and logic benchmarks, GIFT beats strong baselines, with relative gains of up to 9.5% on difficult math tasks.

arXiv:2610.30796 PDF

One Teacher, Many Views: Reducing Reference Sensitivity in On-Policy Self-Distillation

Khouloud Saadi, Di Wang On-policy self-distillation (OPSD) trains a reasoning model on its own outputs, using token-level targets from a copy of itself that sees a reference solution. The authors show that swapping one valid reference solution for another can substantially change the teacher's targets, because the reference also encodes a particular ordering, phrasing, and solution path. Monte Carlo Privileged Distillation (MCPD) instead conditions the teacher on multiple partial views of the reference and averages the resulting next-token distributions. The authors prove this average is the exact target for minimizing expected teacher-to-student KL divergence. Across math and science reasoning benchmarks, MCPD gives more consistent supervision and outperforms full-reference OPSD.

arXiv:2610.30880 PDF

Inherited, Not Invented: Tracing the Provenance of Wait in Reasoning Models

Zhizhuo Kou, Sizai Hou, Long Chen, Jian Yang, Junyu Luo, Jingshu Peng et al. Reinforcement learning (RL) is often credited with teaching reasoning models to say "wait" before reconsidering a step. The authors trace the word through open models' training histories and find that RL from a base model almost never invents it: on bases that lack the word, RL substantially raises MATH-500 accuracy while "wait" stays nearly absent. The marker comes from training data, through mid-training in OLMo-3 or fine-tuning on reasoning traces, and RL recipes amplify, reshape, or even erase it depending on KL regularization and on whether fine-tuning came first. The authors warn that marker frequency alone does not indicate RL-induced reasoning, and that LLM-as-a-judge revision metrics can reverse trends labeled by humans.

arXiv:2610.30882 PDF

Multi-Token Prediction and Alternative Tokenization Improve Learning Latent Multi-Hop Graph Search at Scale

Arvid Frydenlund, FuTe Wong, Katherine Lambert, Frank Rudzicz Prior work found that language models fail to learn shortest-path search on unseen graphs once the graphs get large. The authors trace this failure to two independent causes: the supervision signal and the tokenization. They show theoretically and empirically that multi-token prediction (MTP) helps regardless of graph topology, and they introduce alternative tokenizations that address the second cause. Combining the two lets models consistently learn the task at scales far beyond the previously observed scaling wall.

arXiv:2610.31439 PDF

Causal Chain of Thought: Mitigating Attention Leakage via Context Isolation

Gabriele Dominici, Teague McMillan, Fazl Barez, Marc Langheinrich, Martin Gjoreski Chain-of-thought (CoT) rationales may not actually drive a model's answer, because dense attention lets the model look up the answer from the full context and produce a justification after the fact. CausalCoT breaks a problem into a directed acyclic graph of atomic sub-questions and answers each one in an isolated context that contains only its direct parent nodes, removing attention paths to everything else. Across reasoning benchmarks it keeps accuracy competitive while raising the causal sensitivity of intermediate steps by an average of 12 and up to 21 percentage points over the strongest baseline. The explicit graph also lets a correction made at one step propagate predictably to later steps at test time.

arXiv:2610.31639 PDF

Direct Pass@$k$ Optimization for Thought Pruning: Balancing Quality and Diversity in Test-Time Reasoning

Mahdi Salmani, Adel Javanmard, Devansh Gupta, Meisam Razaviyayn Test-time scaling for reasoning models generates many candidate solutions, and to save compute, recent methods prune candidates after only a short prefix. Existing pruning methods either keep the prefixes most likely to be correct (quality) or keep ones that follow different reasoning paths (diversity), without directly managing the trade-off between the two. The proposed method instead picks the surviving pool to maximize pass@k, the probability that at least one kept partial solution leads to a correct answer once completed. This objective rewards quality and penalizes redundancy at the same time. Across eight open-weight reasoning models it outperforms both quality-based and diversity-based pruning, with the largest gains when a moderate number of candidates is kept.

arXiv:2610.31667 PDF

1 more specialized paper

Reinforcement Learning 11

SpecMAGIC: Specification-Guided Multi-Task Generalization in Multi-Agent Reinforcement Learning

Zijian Guo, \.Ilker I\c{s}{\i}k, H. M. Sabbir Ahmad, Wenchao Li Multi-agent reinforcement learning (MARL) policies rarely generalize to new, temporally extended tasks that need different kinds of coordination between agents. The authors define a specification language that combines temporal logic with operators for coordination requirements. SpecMAGIC uses the automaton structure of a specification to decompose it into subtasks: a centralized planner allocates compatible subtask sequences to agents, and decentralized low-level policies carry them out. This allows zero-shot generalization to new task specifications by composing reusable subtask policies without retraining. The planner is proven sound and probabilistically complete, and on multi-agent navigation the method beats existing approaches on unseen specifications and holds up as the number of agents grows.

arXiv:2610.30443 PDF

Short-to-Long: Detached Features for Long-Horizon Value Learning

Tong Che, Boris Ivanovic, Marco Pavone Deep reinforcement learning has to learn representations from value targets that keep shifting, and long discount horizons make those targets especially sensitive to distant bootstrap changes. Short-to-Long (S2L) gives a long-horizon behavior learner features from source networks trained with smaller discounts. Detached connections stop the behavior loss from updating those sources, so the control objective stays unchanged. Across Atari, online continuous control and offline continuous control the approach yields aggregate gains, improving the paired terminal median on 40 of 57 Atari games. Ablations support the roles of feature transfer, smaller source discounts and gradient isolation.

arXiv:2610.30461 PDF

An Introduction to World Models for Generalist Agents

Zhihui Zhu "World model" is used loosely for many things, from predictive dynamics models and learned simulators to systems that generate future observations and actions together. This introductory tutorial starts from the partially observable Markov decision process (POMDP) framework and connects world modeling to classical control, dynamical systems, signal processing, and reinforcement learning. It compares world-modeling approaches by the knowledge they capture, the objectives they learn, and how they support decisions. A central argument is that world modeling and policy learning can be unified through conditional prediction without separate models, and the paper favors core principles over an exhaustive survey.

arXiv:2610.30789 PDF

MOER: Mix-Policy Optimization under Expert Prefix Guidance with Selective Rollback

Xuyang Zhong, Jiuding Yang, Yiwen Guo, Chen Liu Mixing expert demonstrations into reinforcement learning (RL) for reasoning models can push models past what on-policy RL learns alone. Existing methods either suppress exploration with full expert traces or risk collapse with expert prefixes that may lead to wrong answers. MOER adds an expert chain-of-thought prefix to the prompt, shrinks it over training so the model completes more of the reasoning itself, and replaces failed completions with the full expert trace so every off-policy trajectory is correct. It trains on-policy tokens with a GRPO loss and expert tokens with a trust-region supervised fine-tuning loss. The authors report the highest average scores on math and general reasoning benchmarks among all on-policy and mix-policy baselines, consistent across model scales and model families.

arXiv:2610.30873 PDF

Ground-Truth-First Benchmarking of Reinforcement Learning via Converse Optimality: From Smooth, Fully Observed Systems to Non-Smooth or Partially Observed Tasks

Gr\'egoire Ouerdane, Sinan Ibrahim, Stefan Streif, Pavel Osinenko Reinforcement learning benchmarks usually rank algorithms by return on environments whose optimum is unknown, so nobody can say how far a policy is from optimal. Using converse optimality, the authors construct tasks that come with a certified optimal value function and policy, hidden from learners and used only to measure optimality gap, regret and policy error. The framework offers continuous-time generators for sensitivity studies and native discrete-time generators that remove simulator integration choices as a confounder. Extensions cover non-affine actions, non-Gaussian disturbances, non-smooth value functions and partially observed tasks, with demonstrations on robot arms, nonholonomic vehicles and quadrotor-style systems.

arXiv:2610.30954 PDF

Reward-Weighted Preference Optimization for Selective Combinatorial Optimization Problems

Haohui Zhang, Xinyu Hu, Wouter van Heeswijk, Neil Yorke-Smith, Martijn Mes In selective combinatorial optimization problems such as the Orienteering Problem (OP), Prize-Collecting TSP (PCTSP) and Knapsack (KP), a solver must also choose which elements to include. As a result, sampled solutions can differ a lot in quality, and standard preference optimization reduces these comparisons to binary rankings. Reward-Weighted Preference Optimization (RWPO) weights each pairwise preference loss by the normalized reward gap between the two trajectories. It is combined with scale-specific fine-tuning and tilted sampling to generalize to larger instances. RWPO consistently beats unweighted preference optimization and policy-gradient baselines, and on the OP it exceeds the Compass metaheuristic's mean reward at 100 to 2,000 nodes without local-search post-processing.

arXiv:2610.31239 PDF

Same Luck, Different Choice: Twin-World Credit Assignment for Agentic Reinforcement Learning

Zeropy Surio, Tobias Engel, Mei-Lin Chou, Anika Raman, Samuel Okafor In agentic reinforcement learning, a reward for the whole episode has to be credited to individual decisions. Counterfactual branching blurs a decision's effect with fresh sampling luck in the branch. Because retriever, corpus, and grader are deterministic in agentic search, TWICE (TWIn-world Credit Estimation) fixes the Gumbel sampling noise per token so that counterfactual twins share the original episode's downstream randomness. The paired outcome differences become turn-level advantages that are unbiased and lower in variance, cutting the rollouts needed to resolve an advantage Δ from Θ(Δ⁻²) to Θ(Δ⁻¹). On seven question-answering benchmarks, TWICE improves GRPO's average exact match by 5.3 and 4.5 points on Qwen2.5-3B and 7B and uses 68% fewer auxiliary tokens than VinePPO.

arXiv:2610.31544 PDF

METTA: Memory-Enhanced Test-Time Adaptation for Meta-Reinforcement Learning

Ali Najar, Hosein Hasani, Amir Kooshan Fattah, Yasamin Rajabi, Mahdieh Soleymani Baghshah Meta-reinforcement learning agents must adapt quickly to unseen tasks, which requires tracking evidence over long interaction histories; recurrent agents compress that history into hidden states, while transformer agents attend over past trajectories at quadratic cost. METTA (Memory-Enhanced Test-Time Adaptation) splits adaptation into two timescales: a causal neural memory handles fine-grained adaptation within an episode at linear cost, and compact per-episode summaries support task inference across episodes. On held-out Meta-World ML10/45 test environments, METTA exceeds a 50% success rate with a low training budget, and on long-context, sparse-reward PointMaze tasks it explores more systematically than transformer baselines and outperforms them.

arXiv:2610.31656 PDF

3 more specialized papers

Robotics 11

TerraTransfer++: Demonstration-Free End-to-End Planner

Zikang Xiong, Weixin Li, Zhouchonghao Wu, Akshay Rangesh, Saarth Bonde, Grantland Hall et al. cross-listed End-to-end autonomous driving is expensive to train, whether by imitating fleet-scale driving logs or by running reinforcement learning in a photorealistic simulator. TerraTransfer++ first trains a policy by self-play in a fast vectorized simulator, then distills its rollouts into a flow-matching trajectory teacher that can also plan. Lightweight adapters on a frozen vision backbone are trained to reproduce the teachers' latent spaces, so the only data needed is paired image and scene-state frames, with no expert demonstrations. The resulting closed-loop policy beats all published imitation-based end-to-end methods on HUGSim, and the planner reaches 59.4 EPDMS on NAVSIM navhard, above every published method, plus 95.5 PDMS on navtest, all without human demonstrations.

arXiv:2610.30319 PDF

Externalise the Memory: A Pre-Registered Study of Where an Embodied Agent's Experience Should Live

Cheng-Yin Tu cross-listed In a pre-registered study, the authors test whether an embodied agent's experience should be distilled into its policy weights or stored externally and retrieved at inference. A vision-language model (VLM) teacher that retrieves lessons from its own past attempts generates trajectories on a reach task, and LoRA student policies are distilled from those trajectories. Distilling memory into weights gave only a non-significant +1 point gain, while attaching the teacher's frozen retrieval buffer to fixed student weights added +34 percentage points, matching the teacher. Once the buffer was attached, the adapter added nothing detectable, and on a second, harder task the retrieval advantage pointed the same way but was not significant.

arXiv:2610.30478 PDF

Multi-Tool Selection for Robotic Systems: Comparing Semantic Retrieval and LLM

Miguel Sousa, Luis Ramos Pinto, Jo\~ao P. Matos-Carvalho cross-listed Natural-language control of robots requires tool routing, which means picking the smallest complete set of predefined capabilities a request needs; this is hard for compound instructions. Three increasingly refined retrieval strategies built on ChromaDB (fixed-threshold, relative-distance, and clause-aware aggregated retrieval) are compared against a cloud-hosted 120-billion-parameter LLM on 164 annotated prompts at three complexity levels. The clause-aware method CDB-v3 scores 0.7822 versus 0.8869 for the LLM baseline on a completeness-weighted composite. However, it has much lower latency and can run locally without cloud connectivity, which makes this a quality-latency trade-off for resource-constrained or privacy-sensitive robots.

arXiv:2610.31013 PDF

Memory before Goal: Role of Pre-History Memory for Humanoid Loco-Manipulation

Chenyuan Chen, Ran Ding, Lechen Zhang, Zewen He, Yoshihiko Nakamura cross-listed Whole-body vision-language-action (VLA) policies for humanoid robots see only the current frame, but in dynamic scenes the relevant cues can disappear as objects move, become hidden or leave the camera view. WholeBodyMem uses SAM3 to track task objects and keeps key frames selected by object events within an adaptive budget, and the action model reads these frames through cross-attention alongside object tokens encoding appearance, geometry and time. The authors also release HuMeGo, a memory benchmark with six simulated and six real Unitree G1 tasks. Base-policy success rises from 36.7% to 88.3% on the real robot (and from 29.1% to 72.6% in simulation), outperforming both memoryless and uniform-memory baselines.

arXiv:2610.31166 PDF

Learning Robot Behavior Priors from Non-Robotic Synthetic Data

Mateusz Odrowaz-Sypniewski, Michael Amir, Jasmine Bayrooti, Ajay Shankar, Amanda Prorok cross-listed Real robot data is usually treated as the bottleneck for robot learning models. The authors ask how much useful structure a model can learn from cheap, procedurally generated non-robotic data instead. They build three pipelines that generate mathematical abstractions of robotic control and use them to pretrain a Transformer on action prediction. Trained only on this synthetic data, the model achieves non-trivial zero-shot prediction on unseen physical robot tasks through in-context learning, and control-specific priors beat generic causal priors. A compute-matched model trained on real robot data does better, but only on the datasets it was trained on.

arXiv:2610.31275 PDF

Sensorimotor Alignment: A Platonic Proxy for Embodied Policy Success

Yunfei Xie, Howard H. Qian, Shiyi Lan, Kaiyu Hang, Chen Wei cross-listed Evaluating robotic foundation models normally requires costly closed-loop rollouts, so the authors look for a cheap offline proxy. Building on the Platonic Representation Hypothesis (PRH), which holds that representations from different modalities converge as models improve, they treat motion trajectories as another modality. They propose Sensorimotor Alignment (S^2), a training-free metric of how well a model's sensory representation matches the geometry of physical motion. S^2 correlates positively with downstream robot task success across embodied model families. It can also identify promising vision-language model (VLM) backbones before they are fine-tuned into vision-language-action models (VLAs), and it offers representation-level explanations for behavior seen in VLAs and world action models (WAMs).

arXiv:2610.31670 PDF

5 more specialized papers

Unclassified 8

What the Specification Did Not Say: Specification-Oracle Gaps in Test-Graded Coding-Agent Benchmarks

Fei Ding No summary available — see the abstract on arXiv.

arXiv:2610.30635 PDF

Asking Less, Adapting Better: Feedback-Efficient Test-Time Adaptation for Vision-and-Language Navigation

Zixu Wang, Huaquan Zhang, Shoukai Xu, Jinwu Hu, Zhiquan Wen, Mingkui Tan cross-listed No summary available — see the abstract on arXiv.

arXiv:2610.30643 PDF

Speculative Acceptance Does Not Decay with Position: A Full-Length Diagnostic on Long Chain-of-Thought with Independent Drafts

Zhiming Meng No summary available — see the abstract on arXiv.

arXiv:2610.30644 PDF

Major Revision, Minor Accountability: LLMs as Reviewers of Published Articles in Top-Tier Journals

Bastien Le Guellec, Su Hwan Kim, Raphael Bentegeac, Aghiles Hamroun cross-listed No summary available — see the abstract on arXiv.

arXiv:2610.31414 PDF

Post-Hoc Feature Selection Layer for Neural Networks Interpretability

Jo\~ao Kenji Suwa, Jo\~ao Pedro Silveira e Silva, Bruno Iochins Grisci No summary available — see the abstract on arXiv.

arXiv:2610.31420 PDF

How Depth Growth and Looping Shape In-Context Learning in Language Models

Ferdinand Kapl, Kaitlin Maile, Stefan Bauer, Johannes von Oswald No summary available — see the abstract on arXiv.

arXiv:2610.31421 PDF

How Well Do Decision Models Transfer to Vietnamese? An Empirical Study of Text Classification

Phat Huu Ha, Nguyen Xuan Phuc, Nong Nhut Quy, Trinh Tran Tran, Nguyen Minh Triet, Dang Van Thin No summary available — see the abstract on arXiv.

arXiv:2610.31422 PDF

FocusVLA: Future-Intent Guided Dynamic View Routing for Vision-Language-Action Models

Huixin Xu, Ziya Yu, Peng Li cross-listed No summary available — see the abstract on arXiv.

arXiv:2610.31423 PDF