AI Paper Digest

Wednesday, November 11, 2026

465 papers cs.AI · cs.LG · cs.CL ← 2026-11-10

Aug Sep Oct Nov

Highlights

EarthRecord: From Pixels to Land Change with Multimodal Agents over Four Decades of Satellite Imagery

Highlight HF pick · 3▲Agents Haoran Zhang, Huizhen Sun, Hanbin Wen, Jie Leng, Jiaxin Zhu, Zhaohua Sun EarthRecord is a benchmark that tests whether multimodal agents can turn raw multi-sensor satellite archives into land-change analysis, leaving cloud masking and cross-sensor harmonization to the agent. It covers 56 sites with four decades of Landsat, MODIS, VIIRS, and Sentinel-2 data and poses 238 questions across ten tasks, with two controls: removing the record entirely, and altering one satellite's data while nothing changes on the ground. Among twelve agents evaluated, leading ones complete most analyses end to end but struggle when an answer requires mapping from imagery, and when one satellite's data are altered, half of the right answers follow the alteration rather than the real ground change.

EarthRecord benchmarks whether multimodal agents can turn four decades of raw, multi-sensor satellite archives into reliable land-change analysis, instead of answering questions on curated or preprocessed imagery. Its two controls test whether a right answer rests on regional priors, on unverified pixels, or on change that actually happened on the ground.

  • The benchmark gives agents the unscaled Landsat, MODIS, VIIRS and Sentinel-2 record for 56 sites (1984–2026) in an offline CPU-only sandbox, leaves cloud masking and cross-sensor harmonization to them, and scores 238 questions across ten tasks by pass^3 (right in all three passes).
  • Of twelve agents, the top four (Opus 5.5, GPT-6 Sol, GPT-6.1 Sol, Sonnet 5.5) reach 81.0–85.4% on the 226 computable and recognition questions with overlapping confidence intervals, while GPT-6 Luna and Qwen 3.7 Flash fall below the 23.9% majority-class baseline.
  • Mapping from imagery is the main bottleneck: the six leading models score 92.0% on questions computed from record variables but only 61.1% on those needing a map, where GPT-6.1 Sol, GPT-6 Astra and GPT-6.1 Terra abstain in 46–67% of recognition runs, mostly without viewing an image.
  • Giving agents the record lifts every model on computable questions by 4.6 to 60.8 points over the closed-book control, but when one satellite's data are altered with nothing changing on the ground, 53.5% of previously right answers turn wrong and only 3.5% of those failed runs blame the instrument.
  • Limitations include a sandbox without pretrained vision models, only 22 instrument pairs that test clear-cut breaks instead of subtle drift, and references that rest partly on global products and on 37 AI-drafted, human-reviewed visual interpretations.

arXiv:2611.12934 PDF

Predictive Surprise as Self-Grounding Concept Bottleneck for Interpretable Time Series

Highlight HF pick · 3▲Other Sachith Abeywickrama, Emadeldeen Eldele, Min Wu, Xiaoli Li, Chau Yuen Time-series classifiers in risk-sensitive settings need to be accurate, interpretable, and able to abstain, but defining meaningful patterns requires segment boundaries and vice versa. ConceptTime resolves this with a frozen probabilistic forecaster: the mismatch between its prediction and the observed signal, called predictive surprise, both locates segment boundaries and characterizes each segment, and the segment summaries are clustered into a vocabulary of dynamical-regime concepts. A lightweight head classifies from the concept sequence, with distance to the nearest concept serving as a reliability score. Across 17 public benchmarks it is competitive with black-box accuracy, beats every interpretable baseline, and retains near-98% accuracy on Epilepsy with 1% of labels, without annotations, language model calls, or domain experts.

Concept bottleneck models for time series face a circularity: meaningful concepts need segment boundaries, and meaningful boundaries need concepts. ConceptTime breaks it with one frozen probabilistic forecaster whose next-step negative log-likelihood ("predictive surprise") both places the boundaries and describes each segment by how it deviated from expectation, not by its shape.

  • A 4-layer causal Transformer (ARGDM) is trained once by next-step Gaussian density estimation, its per-timestep surprise drives a length-constrained dynamic program that cuts each series into K patches, and each patch's predictive-statistics signature is clustered by a diagonal GMM into 8–32 label-free concepts read by a classifier of only ~10³–10⁴ parameters.
  • Across 17 benchmarks it is the strongest interpretable method, and on Epilepsy with 1% of labels it reaches 97.91% against 95.41% for the best black-box baseline (TSLANet), though that low-label advantage does not carry over to UCI-HAR (74.60% vs 86.55% for MiniRocket) and at full supervision it trails black-box models by roughly 2–8 points outside Epilepsy.
  • The mean distance from patches to their nearest concept prototype serves as a built-in fidelity score for abstention, reaching OOD-detection AUROC of 0.864 to 0.992 on Gaussian-noise-perturbed UCI-HAR while softmax confidence stays near chance, and giving the lowest deletion-faithfulness AUC on Epilepsy (0.804) but narrowly losing to GradientSHAP on UCI-HAR (0.418 vs 0.411).
  • On cross-condition fault-diagnosis transfer, keeping the source concept vocabulary and training only a classifier on 5% of target labels with a target-trained density model averages 94.37% over 12 directed pairs versus 90.76% for fine-tuned TSLANet, while zero-shot transfer manages only 67.61%.
  • The ablations weaken the single-primitive story: the choice of boundary signal moves accuracy by at most 1.8 points and surprise is not the best signal on UCI-HAR or Sleep-EDF, a vocabulary of 128 beats the default 32 on all three ablated datasets, and the authors acknowledge that vocabulary size is fixed in advance and that each concept's meaning must be assigned by humans after training.

arXiv:2611.12955 PDF

ProTEUS: A mutation-resolved world model for closed-loop protein evolution

Highlight HF pick · 8▲Applications Yikun Zhang, Minghao Guo, Guangxue Xu, Peter Yichen Chen, Wojciech Matusik, Wengong Jin et al. Protein engineering must search a huge combinatorial mutation space while experiments can sample only a small fraction of it. ProTEUS is a world model pretrained on deep mutational scanning assays that uses a mutation-biased attention architecture, MutationFormer, to represent variants as states and mutations as actions with predicted functional consequences, and it is refined through closed-loop experimental feedback. It includes a physics-constrained all-atom structure predictor reported as more than 10 times faster than AlphaFold3 with guaranteed physical validity. Beyond surpassing established models on zero-shot and fine-tuned mutation prediction, two closed-loop campaigns on aldehyde dehydrogenase and levansucrase yielded variants with more than 10-fold higher activity than wild type.

Protein engineering has to search a combinatorial mutation space that wet-lab screening can only sample sparsely, so the bottleneck is deciding which variants are worth making. ProTEUS casts this as a world model: variants are states in a joint sequence-evolution-structure latent space, mutations are actions, and the model predicts their functional consequences, then updates itself on experimental feedback from each design round.

  • The backbone is MutationFormer, a mutation-biased attention architecture pretrained on deep mutational scanning assays, which fuses sequence, evolutionary, and structural signal into a per-variant state representation rather than scoring mutations in isolation.
  • To make in-silico exploration of many mutants tractable, the authors pair it with an all-atom structure predictor that enforces geometric constraints explicitly, reporting a >10-fold speedup over AlphaFold3 with guaranteed physical validity of the output coordinates.
  • On mutation-effect prediction the model reportedly beats established baselines in both zero-shot and supervised fine-tuning settings, though the abstract gives no benchmark numbers, so the size of the margin and the choice of comparators remain to be checked in the paper itself.
  • Two closed-loop, multi-objective campaigns — on aldehyde dehydrogenase and levansucrase — produced variants with >10-fold higher activity than wild type, which is the strongest evidence here because it is wet-lab rather than retrospective.
  • The main caveats are scope and dependency: only two enzymes were engineered experimentally, the pretraining signal is deep mutational scanning data that is itself biased toward shallow, single-mutant coverage, and the closed-loop gains require real assay rounds, so the framework is not a drop-in replacement for screening so much as a way to aim it.

arXiv:2611.13348 PDF

Massive Activations Confound Sparse Autoencoders

Highlight Safety & Alignment Nathan Delisle, Chenhao Tan Sparse autoencoders (SAEs) are usually judged by variance explained and ablation damage, and both measures are distorted when massive activations push up to 99.9% of activation variance onto the first principal component (PC1). Because PC1 has no unique sparse decomposition, each SAE run rebuilds it from different features, so an average of 47 of the 50 most important features have no match in a separate run, and a dictionary can report 99.96% variance explained while leaving 33% of the non-PC1 variance unexplained. The proposed fix trains the SAE on activations with PC1 removed and adds PC1 back to the reconstruction. In affected models this raises top-50 feature sharing from 4-39.5% to 73-99.5%, with up to 44% less model degradation at the same TopK sparsity.

Standard sparse autoencoder evaluations are distorted by massive activations, which push up to 99.9% of residual-stream variance onto the first principal component (PC1), so both variance explained and ablation-based feature importance mostly measure how PC1 gets reconstructed. The proposed fix is to train the SAE on activations with PC1 projected out and add the PC1 component back to the reconstruction at inference.

  • Because a rank-one signal has no unique sparse decomposition under TopK, each seed "tiles" PC1 with a different set of partially aligned features (about 1% of the dictionary) that dominate importance rankings, so in GPT-2 an average of 47 of the 50 most important features have no match in another run, versus about 27 of a random 50.
  • Variance explained is inflated by the same effect, since content error is understated by roughly 1/(1 − w₁), where w₁ is PC1's variance share: the Qwen-2.5-7B SAEs report 99.96% variance explained while leaving 33% of the variance outside PC1 unexplained, an understatement of about 890×.
  • The inversion appears under TopK, BatchTopK and scheduled-threshold JumpReLU but not ReLU+ℓ1, and adding an ℓ1 magnitude penalty to TopK on GPT-2 raises top-50 sharing from 6.4% to 94.9% with under 1% change in sparsity or content error.
  • PC1 bypass lifts top-50 cross-seed sharing in affected models from 4–39.5% to 73–99.5%, degrades the host model up to 44% less, and cuts content reconstruction error by 33.9% on Qwen-2.5-7B (0.8–4.9% on the other models), while a post-hoc variant that simply drops PC1-aligned features from the ranking reaches 77.5–96% sharing without retraining.
  • The repair changes which features get ranked rather than fixing seed instability overall: dictionary-wide sharing stays at 44–68%, 12–33% of content variance remains unexplained by standard SAEs, the Gemma-2 models and low-PC1 layers show no inversion to fix, and downstream gains are mixed, with sparse-probing differences under one point and auto-interpretability scores falling on Llama-3.1-8B and Gemma-2-2B.

arXiv:2611.13548 PDF

Are Dense Rewards Worth Their Tokens in GRPO?

Highlight HF pick · 3▲Reinforcement Learning Vikash Singh, Weicong Chen, Debargha Ganguly, Yanyan Zhang, Mohsen Hariri, Sreehari Sankar et al. Dense process rewards have been proposed to improve on group-relative policy optimization (GRPO), which gives every token of a solution the same credit, but comparisons per training update ignore their cost: in this setting the Monte-Carlo continuations alone consume 91% of generated tokens. A matched-compute protocol gives every intervention the same budget in generated tokens and training GPU-hours, tested in preregistered studies across 1,668 runs on six models. On Qwen2.5-1.5B the Monte-Carlo reward leads GRPO per update but ends 14.3 percentage points lower at equal tokens and 7.8 points lower at equal training time, and none of four dense-reward variants beats GRPO. Training on prompts whose predicted success is nearest 1/2 does improve accuracy at matched compute, by 1.7 points on Qwen3-0.6B and 0.8 on Qwen2.5-1.5B, although GRPO with twice the initial learning rate comes within half a point of those gains.

Dense "process" rewards for GRPO are usually judged per training update or with a single-batch gradient test, which ignores their cost: Monte-Carlo continuations alone consume 91% of all generated tokens. A preregistered matched-compute protocol gives every intervention the same budget in generated tokens and training GPU-hours, across 1,668 runs on six small Qwen and Llama models, and finds that matching compute reverses the ranking.

  • The token budget counts every generated token, including continuations and discarded samples, and controls for side effects (random choice among the same candidate prompts, a retuned learning rate, a strict verifier) accompany pass@1 accuracy on held-out MATH problems.
  • On Qwen2.5-1.5B, the Monte-Carlo reward leads outcome-only GRPO by about 1.5 points at equal updates but ends 14.3 points behind at equal tokens (50 updates versus 584) and 7.8 points behind at equal training time, and none of four dense-reward variants beats GRPO even when given as much training time.
  • A per-token gradient efficiency test orders the rewards the same way training does, but overestimates the accuracy each reaches by 0.8 to 1.8 points, which is as large as the gains it is meant to detect.
  • Training on the prompts whose predicted success is nearest 1/2 does survive matched compute, gaining 1.7 points on Qwen3-0.6B and 0.8 points on Qwen2.5-1.5B at equal tokens, with both gains replicating on fresh seeds and on 1,543 unseen MATH problems.
  • That gain appears on only two of six models, is an early head start that fades by twice the budget, and is nearly matched by GRPO with twice the initial learning rate (within half a point), while all experiments use models of 4B parameters or smaller on numeric-answer math.

arXiv:2611.13560 PDF

Market-Conditioned JEPA: Learning from Unpredictable Market Shocks

Highlight HF pick · 4▲Applications Yuqi Wang, Fengyuan Liu, Mukai Li, Haochen Luo, Lei Li, Qi Liu Much of the variation in financial time series comes from market-wide shocks that cannot be predicted from an asset's own history, and per-asset self-supervised learning treats it as noise. MC-JEPA, a joint-embedding predictive architecture, conditions each asset's prediction on a leave-one-out market signal built from the future target embeddings of other assets, applied through factor loadings derived from the asset's own representation. On 11 years of minute-level Chinese A-share data it improves over per-asset JEPA on nine targets spanning returns, market exposure, risk, and liquidity, and achieves the best RankIC on all liquidity targets and on return horizons from two hours to five days against time-series foundation models and self-supervised baselines. The predictor's factor loadings also rank stocks by market beta nearly as well as a supervised probe, without labels.

Per-asset self-supervised learning on financial time series treats market-wide shocks as irreducible noise, which gives the encoder little reason to learn how each asset responds to the market. MC-JEPA instead conditions the predictor during training on a market signal built from other assets' realized futures, so that encoding market exposure becomes necessary for optimal prediction.

  • A per-asset JEPA encoder is trained with a factor-structured predictor that adds a leave-one-out mean of other assets' future target embeddings, routed through a single-factor loading computed from the asset's own representation, and this condition is dropped at inference so the deployed encoder needs no cross-sectional input.
  • On 1-minute Chinese A-share bars (pretrained on 2015 to mid-2022, frozen ridge probes tested on 2023–2025), it beats per-asset JEPA on all nine targets (significantly on seven) and has the best RankIC on six of nine, for example Ret 1d at 0.0746 versus 0.0712, while TS2Vec leads on 20-day returns and CPC on intraday volatility and beta.
  • In a long-only CSI300 backtest holding 7–8 stocks, it posts the highest annualized excess return (0.135 versus 0.117 for per-asset JEPA) and Sharpe ratio (1.19) with the lowest maximum drawdown (0.117), though its information ratio only ties Kronos-mini.
  • Ablations show the return and exposure gains vanish when the condition comes from a different date, and the predictor's loadings rank stocks by 20-day beta at RankIC 0.368 against 0.374 for a supervised probe, but only after rescaling by price volatility (unscaled it is −0.06).
  • Evidence is limited to one market at one bar frequency, the theory assumes a linear factor model with the true factor observed, and the single-factor choice favors returns, since more factors improve exposure and risk prediction but hurt return prediction.

arXiv:2611.13572 PDF

Transformers Provably Learn Graph Search: Training Dynamics and the Exponential Power of Depth

Highlight Theory Xutao Ma, Somayeh Sojoudi Path-finding over directed graphs is used as a symbolic abstraction of multi-step reasoning to study what transformers can represent and learn. One-layer transformers are proved to need size on the order of N squared on dense graphs to implement the child-selection step of depth-first search (DFS), while a two-layer transformer implements full DFS with size O(N log N). For every k of at least 3, a k-layer transformer implements a bounded-reachability DFS policy with look-ahead horizon 1 + 2^(k-3), an exponential increase with depth. Given appropriately curated training data, two-layer transformers are also shown to provably learn DFS through gradient flow and generalize to unseen graphs.

Path-finding on directed graphs serves as a symbolic proxy for multi-step reasoning, and the question here is what a transformer architecturally needs to run depth-first search with backtracking, and whether training can actually find that solution. The answer comes in three parts: a sharp one-versus-two-layer separation, an exponential depth-to-look-ahead tradeoff, and a gradient-flow proof that a two-layer model learns DFS from discrete chain-of-thought traces, with code and Lean 4 proofs released.

  • A one-layer transformer needs Ω(N²) representation size (heads × width × precision) on dense graphs just to pick a DFS child, shown by a communication-complexity reduction from INDEX, whereas a two-layer construction runs full DFS in O(N log N) using a predecessor-copy layer followed by competing forward and backtracking heads.
  • For every k ≥ 3, a k-layer transformer implements a bounded-reachability DFS policy with look-ahead horizon 1 + 2^(k-3), because each added propagation layer doubles the reachability radius stored on the edge records.
  • Under gradient flow on a permutation-symmetric curriculum, the forward-head parameters grow as 2·log(αt) and the backtracking gap as log(αt), giving a worst-case generalization loss over all N-vertex graphs and valid traces bounded by C·(αt)^-(N-1)/N.
  • Experiments with N=16 and K=8 reproduce the theory closely, with fitted growth slopes of 1.99, 2.00, 1.00, 1.00 against the predicted 2, 2, 1, 1 and loss exponents of -1.000 and -0.9375 on 64,325 DFS states drawn from 300 random directed graphs.
  • The separation concerns width and precision rather than parameter count or running time, the look-ahead is a Boolean reachability test that does not guarantee shortest paths or better worst-case search efficiency, and the learning result trains only the second-layer key-query matrices on a single curated graph skeleton with the first layer fixed.

arXiv:2611.13594 PDF

Few-Step Streaming Video Generation via Condition Shifting

Highlight HF pick · 12▲Vision Deyuan Liu, Jungang Li, Kaiwen Zheng, Kai Jiang, Jintao Chen, Yaole Wang et al. Distribution Matching Distillation (DMD) trains few-step video generators using a separate fake-score network, which adds memory and computation. This work extends APEX to streaming video with condition shifting: the original text condition selects generation while a shifted condition makes the same parameters estimate the velocity field of the generated distribution, removing the separate network. On 180-second videos, human raters prefer it over S2-Avatar most clearly on instruction following (+30.0% net preference), while automated metric differences are small, motion quality favors S2-Avatar, and audiovisual synchronization is mostly tied. Under matched training settings it uses 4.4% less peak GPU memory and 9.2% less time per iteration than DMD.

Distribution Matching Distillation (DMD) makes streaming video generators few-step, but it needs a separate fake-score network that adds training memory and compute and must keep tracking the evolving generator. APEX folds that estimator into the generator itself through condition shifting, so one set of parameters both generates video and estimates the velocity field of its own output distribution.

  • The original projected text condition selects generation while its negation (c_shift = -c) selects fake-velocity estimation on the same parameters, slotted into S2-Avatar's Self-Replay Forcing with its frozen teacher and perceptual regularization kept unchanged.
  • On all 320 StreamAV-Bench scenarios with 180-second videos at 3 backbone NFE per block, APEX edges out S2-Avatar on pooled video instruction fulfillment (3.891 vs 3.834) and visual aesthetics (0.642 vs 0.633), ranking first on VA, VIF, and AIF among 12 models.
  • In a 70-pair expert study against S2-Avatar, instruction following shows +30.0% net preference (22 wins, 47 ties, 1 loss), consistency and video quality each gain +4.3%, and the instruction advantage persists at 2 NFE (+27.1%), though every other dimension turns negative there.
  • Under matched training settings, dropping the independent fake network cuts peak GPU memory by 4.4% (102.7 to 98.2 GiB per GPU) and mean iteration time by 9.2% (35.9 to 32.6 s) relative to DMD, a modest saving since the frozen teacher remains.
  • The gains are narrow and come with trade-offs: motion quality favors S2-Avatar (-8.6%), audio quality, audiovisual alignment, and subject consistency dip on both tracks, pooled sync error rises from 1.319 to 1.341 s, and usability is identical at 62 of 70 sequences.

arXiv:2611.14033 PDF

Tree-Ising: Repairing Language Models with Bit-Level Ising Edits

Highlight HF pick · 2▲Large Language Models Hadi Salloum, Ammar Ali Prior work on weight-space programming had a frontier LLM agent read a small model's weights and write a gradient-free repair for its reasoning loops, at a cost of 158 agent turns and 89.7 million tokens. Tree-Ising recasts the repair as sparse discrete optimization over the model's BF16 weights, where the allowed edits are sign flips and exponent shifts, and decomposes the objective exactly into independent QUBO (Ising) problems per output row that are solved on a GPU with simulated annealing or simulated bifurcation. On the same two layers of Qwen3.5-0.8B it reaches 54.6% on MATH-500 versus 51.7% for the agent's edit (base model 7.9%), using four forward-pass requests, zero LLM tokens, and about $0.90 of GPU time, while changing fewer weights and bits. On Qwen3.5-2B the same recipe raises greedy GSM8K accuracy from 21.6% to 69.4%.

Small reasoning models such as Qwen3.5-0.8B get stuck repeating text in thinking mode until the token cap (93% of MATH-500 completions, 7.9% accuracy), and the prior fix relied on a frontier LLM agent reading and rewriting weights at a cost of 89.7M tokens and $312. Tree-Ising instead casts the repair as a sparse discrete optimization over the stored BF16 bits, solved as a batch of Ising problems using forward passes only.

  • Forward statistics from the base model's own looping and successful traces define a loop direction and a concluding direction, and the objective reduces the layer's write along their combination on loop tokens, penalizes output change on productive tokens, and limits the number of changed weights.
  • Edits are restricted to sign flips and power-of-two exponent shifts that leave every shared group exponent and mantissa untouched, which lets the objective split exactly into 2,048 independent 512-spin QUBOs (one per output row), solved on a GPU by simulated annealing or simulated bifurcation.
  • On the same two layers the agent edited, it reaches 54.6% on MATH-500 versus 51.7% for the agent edit (p ≈ 0.003), leads on all five math benchmarks (significant on MATH-500 and GSM8K after Bonferroni correction), and cuts answers hitting the cap from 31.6% to 23.9%, while changing 13% fewer weights and 2.2× fewer BF16 bits in 31 minutes for about $0.90.
  • On Qwen3.5-2B, with layers picked by a forward scan and the dose set from the base model alone, greedy GSM8K accuracy rises from 21.6% to 69.4%, though the edit breaks 38 of the 285 problems the base model originally solved and was tested only under greedy decoding on that one benchmark.
  • The cost comparison pits the agent's discovery run against a fixed recipe designed after reading the agent's logs, reusing its layers and a dose set to 1.6× its edit (so not dose-matched); a closed-form rank-one edit of the same objective matches accuracy on the practice screen while rewriting 96% of the weights, and the eight-variable QAOA run on IBM Heron hardware is a proof of concept with no claim of quantum advantage.

arXiv:2611.14210 PDF

MMDM: Measuring Massive Multitask Decision Making

Highlight HF pick · 23▲Large Language Models Junhao Shen, Qing Zong, Yue Guo, Zhanqiu Zhang, Yiwen Guo, Hong Cheng System-1 decision models output structured probabilistic choices over given options in a single forward pass, without generating reasoning text, and existing benchmarks for them risk test-set exposure and do not combine long contexts, large option spaces, and logical dependencies. MMDM is a benchmark of 5,448 questions across 16 task categories covering single-select, multi-select, and soft-label decisions, including agent state tracking and tool selection, with up to 1,024 options and inputs up to 64,817 tokens. An evaluation of 28 decision models finds that many do well on common knowledge but struggle with interdependent options and constraints. For example, Perplexity Decider-V1.2 scores 78.75% hard single-select accuracy on world knowledge but 47.95% on constraint planning, and the best hard multi-select accuracy of any model is 36.81%.

System-1 decision models emit a probability distribution over query-specified options in a single forward pass, but existing benchmarks for them reuse public test sets and rarely combine long contexts, large option spaces and interdependent options. MMDM is a 5,448-question benchmark across 16 task categories that tests single-select, multi-select and soft-label decisions on freshly constructed or rewritten instances.

  • The benchmark has 4,112 single-select, 633 multi-select and 703 soft-label questions with 2 to 1,024 options and inputs up to 64,817 tokens, built from new tasks plus source-grounded rewrites and recompositions of sets such as QuALITY, BIRD and ToolTalk, then filtered by five-model majority screening, human review of borderline cases and an option-only shortcut probe.
  • Across 28 decision models, WeMM-JEV 27B leads with 76.00% single-select accuracy and also the best multi-select accuracy at only 36.81%, while every model's Separation–Balance F-score stays below 0.20, showing that recovering the complete valid set remains largely unsolved.
  • Joint-consistency tasks are the weak spot: Perplexity Decider-V1.2 scores 78.75% on world knowledge but 53.95% on option logic and 47.95% on constraint planning, and KnowLine 31B Gen1 drops from 86.0% on deductive reasoning to 54.6% and 48.5% on the same two categories.
  • Accuracy does not predict probability quality or scaling behaviour, since KnowLine 31B Gen1 trails WeMM-JEV 27B by under one point in accuracy yet has roughly half the Soft Brier error (0.1233 vs 0.2327) and under a third of the score error (0.0624 vs 0.2203), and OpenJev 4B beats Intern-Decision 4B in every option-count bin but falls to 15.5% against 34.5% on 8K–16K-token inputs.
  • The benchmark is text-only with a fixed state, query and option set, its multi-select and soft-label questions are concentrated in a few categories (mostly rubric scoring for soft targets), and rewritten instances may still benefit from prior exposure to the underlying public sources.

arXiv:2611.14309 PDF

Applications 93

Tabular Foundation Models vs Tuned Gradient Boosting for Telecom Churn: A Profit- and Calibration-Aware Evaluation

Ismail Hossain Polas An earlier benchmark claimed tabular foundation models (TFMs) decisively beat tree ensembles for customer churn, but it used untuned gradient-boosted trees (GBDTs) and only ranking metrics. This re-evaluation compares TabPFN-2.5, TabPFN-3.5, TabPFN-3.5-Fast, and TabICL v2 against Optuna-tuned GBDTs on five public telecom churn datasets over 10 seeded splits, separately testing ranking quality, expected maximum profit (EMP), and probability calibration. With Holm correction over 20 pre-declared comparisons, TabPFN-3.5 earns significantly more EMP than the best tuned GBDT on Cell2Cell (1.9%) and Iranian (7.6%), while its 14.3% gain on Orange is not significant after correction. Retuning the baseline cuts the earlier headline gap from 6.1 to 4.0 PR-AUC points without closing it, and the authors identify the model's weight licence as the main obstacle to production use.

arXiv:2611.12942 PDF

Code-Grounded Agentic Explanations for Static Analysis Vulnerability Findings

Oshando Johnson, Ranjith Krishnamurthy, Rohith Kumar Shanmuganathan, Goran Piskachev, Eric Bodden cross-listed Developers often dismiss findings from static application security testing (SAST) tools because they cannot interpret them, while handing detection itself to a language model sacrifices the auditability of rule-based analysis. The authors keep detection in the analyser and redesign SAFE, an open-source IntelliJ IDEA plugin, so that explanations are tiered by demand, tied to the reported data flow, and produced by an agent that must consult the repository and the triggering rule before answering. In a within-subjects crossover experiment with 24 participants, System Usability Scale scores rose from 77.3 to 85.0 and the odds of a correct remediation were 2.6 times higher than with the single-prompt predecessor. Triage accuracy and time per finding were indistinguishable, and trust did not change; participants wanted the tool to find and explain while they judge and apply fixes.

arXiv:2611.13041 PDF

Reliability of Claude Code as an Agentic AI Grading System for Handwritten University Exams

Rayed AlGhamdi The study tests how reliably Claude Code, running Claude Opus 4.6 as an agent guided by three rubric files, grades handwritten mid-term exams from 60 students in a university computer graphics course. Against a single expert human grader, AI totals correlated at r = 0.97 with a mean absolute error of 0.85 marks out of 20 and a small AI-positive bias of 0.68 marks, with test-retest consistency of ICC = 0.96 across sessions a week apart and feedback accuracy above 92%. Drawing questions were the outlier, and three failure modes recurred: keyword matching without conceptual depth, ambiguity on corrected multiple-choice answers, and attention drift on hand-drawn diagrams. The author concludes the system can assist on objective and short-form items under safeguards but not replace instructor judgement on boundary cases or drawings.

arXiv:2611.13217 PDF

GigaEdge Spench: An SDK for Practical Streaming Speech Enhancement

Artem Sokolov, Artem Khaidarov, Yana Lysova, Alexander Gladkikh, Vitaliy Yulis GigaEdge Spench is an open-source SDK for streaming speech enhancement on CPU servers, mobile devices, and browsers, letting applications compose neural denoising, normalization, dynamics processing, sample-rate conversion, and echo cancellation. A shared C++ core supports ONNX Runtime and TensorFlow Lite with Android, iOS, and WebAssembly bindings, and ships an 8 kHz INT8 denoiser plus 48 kHz FP32 models. Evaluated on EARS-WHAM v2 against RNNoise and DPDFNet2, every tested native and Web configuration runs faster than real time, with further measurements of how chunk size and execution environment affect cost and how many streams one shared model can serve.

arXiv:2611.13243 PDF

Which Acoustic Features Survive Compression? Tracker-Estimated First-Formant Statistics for Codec-Robust Audio DeepfakeDetection

Yash Sukhdeve (Department of Electrical and Computer Engineering, Clarkson University, Potsdam, NY 13699, USA), Aaron Storey (Department of Computer Science et al. cross-listed Audio deepfake detectors often degrade after telephone or platform compression, so the authors ask which low-dimensional acoustic features remain informative across codecs. Screening 36 utterance-level features on ASVspoof 2019 and testing on ASVspoof 2021 LA under seven codec conditions, they find only four with large practical separation, all statistics of the first formant (F1) and highly redundant with each other. A four-feature RBF-SVM is less accurate on clean audio but far more stable: its equal error rate degrades 5.9 points under codec mismatch versus 14 points for the full 36-feature SVM, suggesting F1 statistics as a compact complement to neural detectors.

arXiv:2611.13258 PDF

ProTEUS: A mutation-resolved world model for closed-loop protein evolution

Yikun Zhang, Minghao Guo, Guangxue Xu, Peter Yichen Chen, Wojciech Matusik, Wengong Jin et al. cross-listed Protein engineering must search a huge combinatorial mutation space while experiments can sample only a small fraction of it. ProTEUS is a world model pretrained on deep mutational scanning assays that uses a mutation-biased attention architecture, MutationFormer, to represent variants as states and mutations as actions with predicted functional consequences, and it is refined through closed-loop experimental feedback. It includes a physics-constrained all-atom structure predictor reported as more than 10 times faster than AlphaFold3 with guaranteed physical validity. Beyond surpassing established models on zero-shot and fine-tuned mutation prediction, two closed-loop campaigns on aldehyde dehydrogenase and levansucrase yielded variants with more than 10-fold higher activity than wild type.

arXiv:2611.13348 PDF

AdjointFlow: Accelerating Scientific Inverse Problems with Physics-Based Transport

Ethan Hanold, Nicholas M. Boffi, Qin Li, Leonardo Zepeda-N\'u\~nez Scientific inverse problems such as seismic imaging are costly for classical solvers, while conditional diffusion models underuse known physics and plug-and-play (PnP) methods reintroduce expensive iterative computation. AdjointFlow builds the physics into the generative transport itself using stochastic interpolants: it starts from a data-dependent, adjoint-based initialization instead of Gaussian noise, modifies the transport velocity to stay regular as the endpoint distributions concentrate, and parameterizes the network around that initialization. On inverse scattering, seismic imaging, and black hole imaging it reconstructs more accurately than conditional diffusion models while being up to five orders of magnitude faster at inference than PnP methods. An optional refinement step that queries the physical model matches or exceeds PnP accuracy without retraining.

arXiv:2611.13389 PDF

Adaptive catalyst discovery emerges from composable scientific intelligence

Wentao Li, Qilong Cai, Qi Lei, Jinxing Chen, Ruixuan Chen, Jiangjie Qiu et al. Knowledge for catalyst design is scattered across experiments, computational data, and catalysis subfields. Composable Catalytic Intelligence (CatAI) is a physics-grounded modular framework that learns 1.8 million structure-property relations and organizes them into 30 physicochemical encoders, a reusable intermediate layer on which task-specific models are built. It reports a 7.8% lower task-averaged prediction error than the next-best models on benchmark tasks. In experimental validation it identified a new confined-pocket catalyst for industrial homogeneous polymer synthesis with 9 times the activity of the commercial benchmark, and guided closed-loop optimization of a copper-based photocatalyst for C-O coupling that surpasses traditional nickel-based systems.

arXiv:2611.13536 PDF

Exploiting Feature-invariant Scaling Laws for Efficient A/B Experimentation in Recommendation Systems

Zhiliang Chen, Sammy Nammari, Jiahao Su, Xiaoqiang Luo cross-listed Running full-scale A/B experiments for every proposed change to a recommendation model is computationally prohibitive, while a single small-scale experiment is often misleading. The authors show that modern recommendation models exhibit feature-invariant scaling, meaning performance follows a predictable scaling law even when different changes are applied to the baseline, and use it with statistical inference to predict the full-scale effect with confidence intervals. Validated on Wukong, HSTU, and LLM-based recommenders across more than 40 architectural, optimization, and data changes, the method separates beneficial changes from neutral and regressive ones using only a fraction of full-scale compute. It also surfaces several changes that improve HSTU's state-of-the-art result on MovieLens-20M.

arXiv:2611.13539 PDF

Machine-learning decoders for fusion-based quantum computing in real time

Daniel Barter, Jacob C. Bridgeman, Chris Dawson, Terry Farrelly, Shane Fleming, Mercedes Gimeno-Segovia et al. cross-listed Photonic fusion-based quantum computers are dominated by erasure errors from photon loss and probabilistic fusion measurements, so they need decoders that stay accurate at high erasure rates while meeting real-time latency limits. PRISM is a transformer-based decoder that exploits the known locations of erasures, benchmarked on surface codes, color codes and quantum low-density parity-check codes at distances up to 24, where it matches the logical error rates of integer-programming decoders while running significantly faster. An optimized implementation on an AMD VP1902 FPGA decodes distance-8 logical blocks with a latency of 6.59 microseconds and an amortized 1.56 microseconds per logical block.

arXiv:2611.13630 PDF

Swanada: Split-Encoder Capability-Routed Conformer CTC and Sub-30 MB On-Device Speech Recognition Across Twelve Scripts

Hemanth HM cross-listed Running automatic speech recognition (ASR) in browsers and on phones across many Indic scripts is blocked by per-language model size, slow or unsupported WebAssembly kernels, flickering streaming output, and mobile audio pipeline quirks. Swanada is an open-source streaming system for Sanskrit, Indian English, and ten Indian languages that shares one quantized INT8 Conformer encoder, in a 177 MiB desktop tier or a distilled 25 MiB mobile tier, across twelve hot-swappable 515 KiB per-script CTC output heads. This cuts twelve-language storage by 11.6x on the desktop tier and 67.8x on the mobile tier, to 31.39 MiB. On open public test clips including FLEURS and LibriSpeech, the desktop tier reaches 8.87% character error rate and the mobile tier 19.23%, with 30.6 ms median time to first token.

arXiv:2611.13814 PDF

Domain-Invariant and Language-Model-Assisted Acoustic Side-Channel Attacks on Keyboards

Bikrant Bikram Pratap Maurya cross-listed Acoustic side-channel attacks (ASCA) recover typed text from keystroke sounds using only a commodity microphone, but prior demonstrations relied on idealized conditions with a single keyboard, a fixed microphone, a quiet room, and the same device for training and testing. This work asks whether keystroke inference holds up with unseen keyboards, diverse users, noisy environments, and compressed transmission channels. The authors claim that the attack does survive these open-world conditions, provided the acoustic model is made explicitly domain-invariant and its noisy predictions are corrected by a language model; the abstract reports no quantitative results.

arXiv:2611.13922 PDF

PARDIG: Global Precipitation Estimates Learned from Sparse Satellite Observations

Guy Shalev, Stephan Rasp, Boris Babenko, Andrew El-Kadi, Mohammed Alewi Hassen, Samier Merchant et al. Global precipitation products trade off skill, latency, coverage and resolution. PARDIG uses a spherical graph-mesh transformer trained on atmospheric state analyses and multi-spectral geostationary satellite imagery to predict sparse, high-precision spaceborne radar measurements, and is then applied densely to produce global fields at 0.1 degree and 15-minute resolution from 2016 onward, with a historical variant back to 1999. Evaluated against three independent radar networks and more than 36,000 surface stations, it consistently outperforms IMERG V07, MSWEP V3, ERA5 and HRES across thresholds and regions, with the largest gains at extreme intensities. The WeatherNext 3 forecasting model also produces more skillful precipitation forecasts when trained to predict PARDIG, and both datasets are openly released.

arXiv:2611.13941 PDF

HandsOnAid: Building and Evaluating an Offline First-Aid Assistant for Emergency Response Scenarios

Rezwan Islam Salvi, Saiyma Sittul Muna, Mushfiqur Rahman Mushfique HandsOnAid is an offline first-aid assistant for places with no phone signal, built by fine-tuning a 4-bit Gemma 2B model with QLoRA on the 5,550 question-answer pairs of FirstAidQA. Seven configurations were scored by four frontier LLM judges against 41 medically reviewed reference answers on a strict 0-5 rubric with safety overrides, and a medical professional's blind review of 40 answers gave similar scores. Fine-tuning raised the mean score by 1.08 to 2.34, and the best configuration, the fine-tuned 4-bit model with retrieval, scored only 2.51 out of 5, which a keyword audit links to guideline details missing from the training data. On a Pixel 7 Pro with no network, the model shows its first text in 6.86 seconds and a complete answer in 14.88 seconds on average.

arXiv:2611.14059 PDF

FERN: Cross-Benchmark Table Retrieval for Text-to-SQL with Frozen Retrievers and Learned Ranking

Arjun Rao, Nishant Gopinath cross-listed Text-to-SQL systems over large pooled catalogs must first retrieve the tables a question needs, a step the compared system RASL handles with Claude 3.5 models extracting keywords and selecting tables. FERN (Feature-Engineered Ranking with No generative LLM) instead fuses three frozen retrievers into 100 candidates, describes each with up to 120 features, and reorders them with LightGBM rankers fitted once on BIRD's training questions. On Spider, where it was not fitted, its candidate pool alone matches RASL's published recall at fifteen tables and the ranker lifts the share of questions with every gold table in the top five from 83.9% to 95.6%, with median latency of 208 ms per question on one A100 and no hosted model calls. It trails by 8.2 points on BIRD without evidence hints, falls well short on Fiben at fifteen tables (54.2 against 69.2 on different question sets), and on 121 real BEAVER questions sits between two BEAVER-trained retrievers on recall but below both on complete-set retrieval.

arXiv:2611.14477 PDF

Emergent Biological Capabilities in a Foundation Model for Molecular Interactions

Adi Mashiach Bonbon is a foundation model for molecular interactions trained on roughly 3.2 trillion tokens of paired protein-ligand sequences, with no structural coordinates and no task-specific supervision. From one representation space it shows three zero-shot capabilities at different resolutions: mechanism of action, binding site localization including discrimination of orthosteric from allosteric sites, and covalent versus non-covalent engagement, all generalizing to out-of-distribution complexes. Controlled scaling across six model sizes from 42M to 1.6B parameters shows every evaluated capability jumping sharply at the largest scale even though pretraining loss falls smoothly, indicating loss-based scaling laws alone cannot predict when such capabilities appear. The frozen representations also reach state-of-the-art binding and affinity prediction five orders of magnitude faster than co-folding methods, and prospective wet-lab testing across nine disease targets gave a 61.5% dose-response hit rate on de novo compounds.

arXiv:2611.14504 PDF

Pyrite: Building and Benchmarking Protein-Ligand Pose Ranking Agents

Megha Srivastava, Michael H. Lim, Allison M. Keys, Pranav Murugan, Nate Gruver Modern cofolding models generate many candidate poses for how a ligand binds a protein, but reliably ranking them remains hard. Pyrite is a pairwise pose-ranking benchmark of 1,500 pose pairs across diverse protein-ligand complexes on which expert chemists with 3D visualization and physics-based scores perform near chance. Building on PyriteRep, a structured representation of geometry, interactions, and physicochemical scores, the authors develop PyriteReasoner, which compares poses by text reasoning, and PyriteFeaturizer, which writes executable code to construct features for logistic regression, reaching 67% and 74% accuracy respectively and exceeding observed expert performance. Feature discovery guided by a language model generally beats direct reasoning, and Group Relative Policy Optimization (GRPO) lets open-weight models match or surpass closed-source frontier models in these evaluations.

arXiv:2611.14511 PDF

76 more specialized papers

Large Language Models 68

SELD: Detecting Silent Evidence Loss in Context Compression beyond the Answer

Wentao Lei, Li Liu Context compression for long language-model inputs can remove key evidence while leaving the answer unchanged, a failure the authors call silent evidence loss that answer-based checks cannot see. SELD (Silent Evidence Loss Detector) needs no training or labels and compares a reader model's preferences among the answers it did not give before and after compression, since these shift when evidence is lost. On SEL-Pairs, a new benchmark of length-matched evidence and control deletions built from three question-answering datasets, it identifies the evidence-removing edit in 83% of matched pairs with unchanged answers, against 70% for the strongest baseline, across five reader models.

arXiv:2611.12903 PDF

Actionable Feedback: Improving Model Capabilities from User Edit Data on Clinical Documentation

Daniel P. Jeong, Yewon Byun, John Giorgi, Peng Xu, Zachary C. Lipton, Michael Oberst Clinician edits to large language model (LLM) outputs look like a scalable training signal, but a review of data from a deployed clinical documentation product suggests that over 45% of edited outputs contain user-added information not supported by the model's inputs. The authors build a taxonomy of clinician editing behaviors, use it to generate semi-synthetic edit data for discharge-summary generation on MIMIC-III, and introduce simple filters for identifying actionable edits, validated by expert review. Naive fine-tuning on raw edits increases ungrounded content, whereas progressively filtering toward grounded edits improves clinician-alignment metrics while retaining faithfulness to the input.

arXiv:2611.12904 PDF

SetPrint: Black-Box LLM Fingerprinting via Setwise Correspondence of Response Changes

Jinmo Han, Ju Yeon Kang, Jimin Hong, Sunghwan Ahn, Yonghyeon Jun, Nam Soo Kim Black-box fingerprinting of large language models tries to determine whether a target model was derived from a source model using only responses to queries. SetPrint is a training-free method that perturbs each query, expresses each resulting response change relative to the mean change of its own model's set, and then compares the changes that the same perturbation induced in the two models. On LeaFBench it consistently outperforms existing training-free black-box fingerprinting methods, and analyses attribute the gains to using both within-set relations and between-set correspondence.

arXiv:2611.12950 PDF

Physics2Weight: Direct Neural Network Architecture and Weight Generation from Physical Laws Using Large Language Models

Shogo Sato, Tomoyasu Shimada, Kazuhiko Murasaki, Takashi Sano Weight-space programming (WSP) uses large language models to generate neural network weights directly, without gradient-based optimization, but has been hard to apply to physical problems. Physics2Weight extends it to partial differential equations (PDEs) by generating weights and editing architectures from the governing equations and initial and boundary conditions, using physics-constrained prompting, physics-guided architecture editing, and iterative weight refinement. On the heat, advection, Burgers', and Allen-Cahn equations it reports lower solution errors than conventional physics-informed neural networks, while showing limitations in satisfying the governing equations for nonlinear PDEs.

arXiv:2611.12954 PDF

ADIndex: Layer-Adaptive Sparse Attention Indexing via Group-Level Distillation

Jin-Fei Qi, Yiduo Li, Jiang-Xin Shi, Kan Liu, Tao Lan, Hanlin Tang et al. DeepSeek Sparse Attention (DSA) cuts attention cost through token-level selection, but its indexer still scores the entire causal prefix for every query and can exceed 50% of prefill GPU time at long context lengths. ADIndex addresses this with layer-wise calibration that exploits differences across layers in how much queries' retrievals overlap, group-level distillation to train layer-adaptive group indexers, and token bank sharing within query groups to avoid repeated full-prefix scans. On DeepSeek-V3.2 and GLM-5.1 it keeps near-lossless performance on LongBench, LongBench v2, and RULER while reaching prefill and decoding kernel speedups of up to 6.22x and 3.40x at 256K context, plus a 1.54x end-to-end serving speedup at 128K.

arXiv:2611.12956 PDF

Third Eye: Beyond Immediate Gains in Recursive Self-Improvement

Sujato Dutta, Padala Indira Sneha, Kypu Mahitha, Shashank Vanga, Neeraj Choudhary A model self-update that looks useful immediately may stop being useful once the model updates itself again, and Third Eye is an evaluation protocol that separates immediate performance from the outcome after a further update, using disjoint items for selection and final scoring. The study covers five 3-8B backbones on math and code, with 64 fresh runs, three candidates per run, and three continuations per candidate. In the balanced Qwen3-4B/Llama-3.2-3B comparison, immediately selected utility is near zero but a further update lowers it by 1.78 percentage points, with similar declines for SmolLM3-3B and Llama-3.1-8B and an inconclusive result for Qwen3-8B. Controls identified neither a reliable mitigation nor a single mechanism, and the authors argue that persistent self-updates should be evaluated with independent scoring and a no-update baseline.

arXiv:2611.12988 PDF

Capacity Allocation at the Source: Sparse Target Optimization for LLM Knowledge Editing

Dongyao Chen, Jiayun Lei, Zhiyuan Ji, Zhiying Deng, Wei Liu, Xiaobo Sun Knowledge editing in large language models first computes a target hidden state that steers the model toward a new output and then edits parameters to map the original hidden state to that target, and most work focuses on the second step. The authors show theoretically that interference between edits in parameter space is positively correlated with interference between the target hidden states themselves. They propose Sparse Target Optimization (STO), an L1 regularization on the target that prunes non-essential components of each knowledge vector, so that each edit disturbs less unrelated knowledge and more edits can coexist. Experiments across multiple models and benchmark datasets support the method's effectiveness.

arXiv:2611.12995 PDF

Linear Probes to Bayesian Manifolds: Geometry of Arithmetic in Large Language Models

Anshul Kumar, Neha Prasad, Manoj Kumar, Barnab\'as P\'oczos, Deeksha Varshney Linear probes that find geometric structures such as helices in language model activations are often taken as evidence that the model computes with those structures, but probe accuracy conflates four distinct claims: decodability, shape, ownership by the probed concept, and causal use. A four-stage Bayesian pipeline tests each claim separately and is applied to two-digit addition and multiplication in GPT-J 6B, Llama 3.1 8B, and Pythia 6.9B. Nearly every concept is decodable, and the well-supported shapes are ribbons (helices whose radius varies with the value) or tori, but of 1,330 cells with an ownership verdict only 19% of shapes are owned by the probed concept while 81% are inherited from related concepts. Intervening on the shape retains about 61% of the effect of intervening on the full probe while using at most five dimensions, and the authors propose reporting probe results as verdicts on all four claims.

arXiv:2611.12998 PDF

Multi-Future Drafting: Mixture Block Drafters for Tree-Verified Speculative Decoding

Haoyu Huang, Jialiang Kang, Linlin Yang, Xiao Yang, Baochang Zhang Speculative decoding with block drafters such as DFlash and draft trees such as DDTree ranks candidate prefixes by the product of per-position distributions, which the authors show wastes verification budget on chimeras: prefixes spliced from different plausible continuations that the target model never produces, making up 27-33% of a 64-node tree. Multi-Future Drafting (MFD) adds K learned mode embeddings at the drafter's input so that K coherent blocks are drafted in one pass, and a mixture tree lets the verifier follow whichever future the target takes, losslessly and with only K+1 extra vectors. On four Qwen targets from 4B to 27B parameters at batch size 1, it reaches up to 7.8x speedup, 5-8% faster than the best identically trained single-mode tree and 9-16% faster than the released DARTree on the two smaller models.

arXiv:2611.13017 PDF

The Indic Token Tax in 2026: Measuring Tokenizer. Disparity and Its Cost for the Languages of India

Johin Johny Arimbur, Ankur Ashit Ghosh, Shervin Sam Because commercial language model APIs bill per token, a tokenizer that needs more tokens for the same content makes a language costlier, slower, and less able to fit in a context window. The study measures this for nine public tokenizers, including those of GPT-4, GPT-4o, Qwen2.5, DeepSeek-V3, Mistral-Nemo, and four Sarvam models, on all 25 languages of India in FLORES-200 and on 13,732 Hinglish-English pairs from PHINC. On the GPT-4o tokenizer, the same content needs 1.57 times (Hindi) to 5.01 times (Odia) the tokens of English, and 13.7 times for Santali in Ol Chiki script; the Sarvam-30B/105B tokenizer is the only third-party one staying below 2.6 times on every language. No tokenizer places more than 35% of its word-internal boundaries on morpheme boundaries in Hindi and Marathi, and an illustrative price scenario puts a support bot at about INR 56,640 per month in English versus INR 102,829 in Marathi.

arXiv:2611.13058 PDF

A Multi-Word Tokenizer for 204 Languages That Needs Fewer Tokens Than GPT-4o's, Including on English

Johin Johny, Ankur Ghosh, Shervin Sam Indian languages need 1.6 to 14 times as many tokens as English under GPT-4o's tokenizer, raising serving cost and latency. The authors train byte-level byte-pair encoding (BPE) tokenizers from scratch on roughly 200 languages: Havon v3 uses a pre-tokenizer that keeps combining marks with their letters and chunks scripts written without spaces, Havon v4 adds a multi-word SuperBPE training stage, and Havon v5 broadens the code sample to 15 programming languages. On FLORES-200 devtest, v4 at exactly GPT-4o's vocabulary size of 200,019 tokens needs 6.1% fewer tokens than o200k_base on English and fewer on 161 of 204 languages, including all 25 languages of India, whereas v3 needs more tokens on English. These are compression results only; no model has yet been trained with v4 or v5.

arXiv:2611.13059 PDF

One Temperature to Rule Them All?

Aviv Orly (School of Physics and Astronomy, Tel Aviv University, Israel), Ori Shem-Ur (School of Physics and Astronomy, Tel Aviv University, Israel) et al. Varying sampling temperature across samples has been proposed for multi-sample large language model inference, yet fixed-temperature schemes remain standard; the authors ask when temperature diversity actually helps in the pass@K setting. They analyze how per-question success probabilities change with temperature across several models and datasets, and derive an efficient way to evaluate the theoretically optimal multi-temperature allocation. The optimal multi-temperature scheme yields only modest gains over a single temperature once that temperature is tuned as a function of the sampling budget K. They attribute this to mathematical mechanisms they call dulling effects, which also make temperature choices robust across question distributions, and note that harder questions are less likely to share efficient temperatures, though only logarithmically in hardness.

arXiv:2611.13077 PDF

Prediction-Powered LLM Ranking with Adaptively Collected Human Feedback

Gabriel Sargent, Will Wei Sun, Zhengwu Zhang, Yufeng Liu Ranking large language models (LLMs) by human evaluation is expensive, and platforms often collect feedback adaptively, which breaks the independence assumptions behind standard statistical inference. PP-Rank (Prediction-Powered Ranking) combines cheap LLM-judge predictions on unrated responses with adaptively collected human feedback, using data-driven weights that minimize asymptotic covariance while correcting for judge bias. The authors establish asymptotic normality and simultaneous ranking inference that remain valid under adaptive sampling, and derive a rule for reducing exploration that favors lower-cost comparisons. Semi-synthetic experiments on UltraFeedback-Extended show narrower confidence intervals and better ranking certification than human-only baselines.

arXiv:2611.13186 PDF

Decode Token Parallelism: Memory-Efficient Distributed Attention Parallelism for Scalable LLM Decoding

Susav Shrestha, Brad Settlemyer, Narasimha Reddy cross-listed Serving large language models at high concurrency with long contexts forces a trade-off between KV-cache capacity and attention scaling under data parallelism, tensor parallelism, and decode context parallelism (DCP). TKNP (Decode Token Parallelism) spreads a single model copy across nodes in balanced layer blocks: the node holding a block runs its dense layers for the full batch, while every node computes attention for its own requests against its locally stored KV cache, so only current-token activations cross the network. Implemented in vLLM and tested on Llama 3.1 405B and Nemotron-4 340B on up to 64 H100 GPUs, it improves completion throughput on long-context, high-concurrency workloads by up to 1.7x over DCP and up to 2.06x over data parallelism at matched GPU count.

arXiv:2611.13293 PDF

MicroLens: Read, Trace, and Rewire the Residual Stream with Two Lenses

Nika Zahedi, Fangyi Zhu, Anna Wu Lenses that decode hidden states into words show what a vector reads as, but not whether the model uses that information. MicroLens is an interactive viewer over the layers-by-tokens grid of the residual stream in which each cell can be decoded with the logit lens or a Jacobian lens (J-lens), traced back to source tokens, and edited with the model rerun on the spot. On Gemma-3-4B-it, swapping the state or country readable at the city token from layer 7 redirects two-hop trivia answers 92-98% of the time, even though that information is not used at the answer position. Under step-by-step reasoning prompts, the correct answer is already the top residual-stream word before any reasoning is written for 89-92% of retrieval questions and 39-63% of multi-hop and serial ones, with ordering puzzles as the exception.

arXiv:2611.13322 PDF

Measuring Definition Sensitivity in LLM Annotation of Social Science Concepts

Bogdan Mamaev, Qizhou Wang, Christopher Leckie, Matteo Vergani When large language models (LLMs) annotate text for contested social science concepts such as hate speech or fake news, it is unclear whether labels follow the definition written in the prompt or the model's pretrained notion of the concept. The test has a model annotate the same texts under two competing definitions and measures the disagreement, called definition sensitivity, against a paraphrase floor obtained by rewording one definition without changing its requirements. Across six concepts and three LLMs, swapping definitions always changes more annotations on average than paraphrasing, indicating that the labels depend on the prompted definition. The authors recommend reporting both quantities alongside any LLM-produced annotations.

arXiv:2611.13339 PDF

Looping Deeper Without Growing the KV Cache: Recurrent Depth for Causal Encoder-Decoders

Jiaqi Leng, Minshen Zhang, Zhaofeng Sun, Yu-Yang Qian, Peng Zhao, Hao Zhang Looped language models add computational depth without adding parameters, but caching key-value (KV) states for every layer and every round sharply limits serving batch size. The authors design a looped causal encoder-decoder (CED) in which recurrence sits in the encoder and is trained with stochastic recurrence depth, producing a fixed point where every round can share the last round's cache. The resulting Looped CED matches the memory footprint of an unlooped model and recovers 86-93% of a looped Transformer's gain with one sixth of its global KV. Looping the 552B-parameter DeepSeek-V4.1-Flash for 8 rounds on one 4xGB200 node, it holds 36.8x less KV per token, fits 38 requests of 128K tokens where the naive loop fits none, and serves 3.6x higher throughput on 8K-token prompts.

arXiv:2611.13378 PDF

Plumbline: When the Loss Lies, Validating LLM Training Correctly

Ruochen Wen, Hongyang Wang, Siwei Li, Wenbo Chen, Xian Li, Jingguo Liu et al. cross-listed Faults in large language model (LLM) training can sit anywhere in a run and are often silent, surfacing only as a late crash or a lower score at final evaluation. Plumbline, a correctness validation system used in production at Amazon, checks each part of the training lifecycle from data through new model layers to whole runs, and reuses those checks to trace failures to root causes. It supports two agentic use cases: validating the training of a newly released model before teams rely on it, and root-cause tracing of training failures. Across 3 frameworks and 5 models it found 61 problems, 32 of them silent, of which 42 are fixed or reported upstream, and the authors group them into seven recurring failure patterns.

arXiv:2611.13393 PDF

Bursts That Outlast Themselves: Persistent Degradation in Prefix-Cached Agent Serving

Wen Huang, Jack Liu cross-listed Agent sessions have contexts that grow every turn, so serving them cheaply relies on prefix caching, where a hit costs about a thousand tokens of prefill and a miss recomputes 10K to 40K tokens. Experiments show this makes serving metastable: a few minutes of extra arrivals left a vLLM server stuck at roughly a 10% hit ratio and half its throughput for the remaining 70 minutes in 9 of 12 paired runs, while the no-burst twin stayed healthy. The regime also appears with Qwen3-32B on SWE-bench agent trace shapes and on SGLang, and arises because slower service keeps more sessions resident until their working set outgrows the key-value cache pool. Throttling new admissions recovered 1 of 7 collapsed runs, suspending admitted sessions recovered all of them, and a budgeter bounding resident cache demand prevented the failure in every tested run.

arXiv:2611.13412 PDF

Placing LLM Lookup Tables by Per-Layer Prefetch Deadlines

Wen Huang, Jack Liu cross-listed DeepSeek-V4.1 keeps 203 GB of Engram n-gram lookup tables in host memory, which dominates the footprint of a 4-GPU serving process. Because row indices depend only on input tokens, reads can start at the beginning of each decode step, and each table's layer sets a prefetch deadline (0.3 versus 9.3 ms) that determines which storage tier it can live on. SlackTier, built for SGLang, issues all lookups at step start, prioritizes the short-deadline layer and shares a small GPU row cache, staying bit-exact with the original gather. With both tables on four NVMe disks, time per output token rises at most 1.4% while host memory falls from 212 to 34 GiB, and reusing the freed memory as key-value cache raises median throughput by 36% on a constrained agent workload.

arXiv:2611.13413 PDF

Inherited Weights: Measuring the Tokenization Premium and the Sovereignty Cost of Language Models for the Global South

Paulo Camara (LUA Vision) Most language models deployed outside the United States and China are fine-tunes of a few pretrained families, and the argument here is that vocabulary, knowledge, cultural alignment, memorized data and license are all inherited from the base model. A reproducible measurement of the tokenization premium covers 31 languages and 7 tokenizers on Universal Declaration of Human Rights text: premiums reach 10x for Amharic in older tokenizers and remain 6.8x in the newest commercial one, while Indigenous South American languages pay 1.6x to 2.7x. A Portuguese-built vocabulary lowers the premium for 11 of 13 Latin-script Global South languages. The work also offers an eight-test black-box audit runnable without weight access and argues for accuracy per joule, calibration and selective risk as design objectives.

arXiv:2611.13441 PDF

ForeSpark: Joint Draft and Verification Budget Scheduling for Parallel Speculative Decoding

Jaeyong Song, Daniel Wang, Pratyush Patel, Jonny Kong, Preyas Shah, Shobhit Kanaujia cross-listed Parallel drafters speed up LLM decoding by proposing a block of draft tokens in one forward pass, but existing serving systems only control how many tokens are verified, so at high load drafting takes up to 40% of step time and most of it is discarded. Work-conserving speculation (WCS) jointly allocates draft and verification budgets per request, and its prototype ForeSpark combines a pre-draft acceptance estimator, a joint scheduler based on a global sort and a post-draft verification budget distribution. Implemented on vLLM, it improves throughput by up to 1.77x over DSpark and 2.49x over autoregressive decoding without changing the output distribution. It also sustains up to 1.51x the request rate of DSpark under latency service level objectives (SLOs).

arXiv:2611.13442 PDF

Sanity Checks for Attention Surgery: Null Models and Matched Interventions for Representation-Motivated Design

Basavaraj R M, Aryan Bharat Kumar Architectural changes to transformers are often justified by measuring a large statistic on internal representations, removing it, and reporting a gain, without checking whether the statistic is meaningful. Three reusable checks (noise floor, geometric null, and a matched control intervention) are applied to self-attention removal, which deletes the part of an attention head's output aligned with its own value vector. Across nine pretrained models and 5,408 heads, an earlier-position causal null accounts for 48% to 63% of the head-value cosine alignment. In a pre-registered training experiment at 50.9M parameters, removing the own-value direction lowers validation loss by 0.0029 nats while a fixed arbitrary direction does not differ from baseline, so the benefit depends on the specific direction, yet high alignment alone does not establish redundancy in frozen models.

arXiv:2611.13543 PDF

Closing the BF16-FP16 Convergence Speed Gap in RL Post-Training with Stochastic Rounding

Mahdi Nazemi, Oleg Rybakov, Tijmen Blankevoort, Asit Mishra, Carlo del Mundo In reinforcement learning with verifiable rewards, where learning rates are low, standard mixed-precision BF16 training of language models converges more slowly than FP16 at the same learning rate. The authors trace this to weight sticking: when FP32 shadow weights are rounded to nearest in BF16, small updates often leave the working weights unchanged for several optimizer steps. Raising the learning rate closes the gap but increases KL divergence from the reference model. Applying stochastic rounding when quantizing the shadow weights matches FP16's convergence speed across models and datasets with lower KL divergence than the higher-learning-rate fix, while keeping BF16's wider dynamic range and avoiding loss scaling.

arXiv:2611.13545 PDF

Compute the Interactions, Not the Attributes: What a Fact Layer Must Supply in the Chinese Sexagenary System

Huixue Han, Jiaying Liu Language-model systems in rule-governed domains often rely on a deterministic fact layer, and this study measures which facts that layer needs to supply, using the Chinese sexagenary system behind BaZi chart analysis: 682 items over twelve fact categories plus 160 transit-natal collision items, each with a table-checkable answer. Eleven open-weight models answering directly fail where attributes must be combined, with none getting more than 4% of sexagenary days or 1% of collisions right; supplying static attributes has no consistent effect, and even a complete relation list leaves most near zero. A query-filtered layer of computed interactions raises Qwen3-8B and Qwen3-32B to 78-92% and four further models to 96-100%, while three proprietary reasoning models compute 68-95% of collisions unaided. A preregistered check on chess positions likewise finds that attributes do not substitute for relations.

arXiv:2611.13597 PDF

One Byte, One Token: Byteifying a 27B Hybrid-Attention LLM with Self-Generated Data

Liwei Zhang, ZhiMing Tong An already-trained subword model, Qwen3.8-27B, is converted into a pure byte-level model by shrinking its vocabulary from 248,320 to 384 entries, leaving all other weights untouched, and continuing training in a two-stage curriculum on 4.77 GB of bytes whose dialogue and coding supervision is generated by the original model itself. The byte model's gap to the original falls within the bootstrap interval on MATH-500, HumanEval+, MMLU-Pro and non-thinking GSM8K, it loses 4-6 points on C-Eval, thinking-mode GSM8K and non-thinking IFEval, and it leads on character-level tasks such as CUTE (75.2 to 84.0), while lookup knowledge about Chinese characters drops. The cost is throughput: each decoding step yields 3-5x less output, leaving Chinese generation at about a quarter of the original speed, and prefill at 64K bytes is 6.5x slower. Four byte-model failure modes, including sampling that breaks UTF-8 and ineffective repetition penalties on a small vocabulary, are documented with fixes.

arXiv:2611.13607 PDF

Bootstrapped On-Policy Distillation

Yue Wu, Xueyu Zhou, Yangrong Hu, Jian Huang On-policy self-distillation (OPSD) improves a language model by using a copy of itself that is conditioned on reference answers as the teacher for the student's own generated trajectories, but the authors find it degrades under continued training as the evolving student drifts away from the frozen privileged teacher. Bootstrapped On-Policy Distillation (BOPD) splits the process in two: OPSD first internalizes the reference-conditioned signal into an intermediate model, and that reference-free intermediate model is then distilled into a fresh copy of the initial model. Across models, BOPD improves mathematical reasoning and code generation beyond both the initial models and the OPSD teachers.

arXiv:2611.13680 PDF

FedVP: Federated Learning across Heterogeneous Models by Pooling Locally Trained Verifiers

Yifan Li, Xi Chen Conventional federated learning aggregates updates over a shared model backbone, which blocks collaboration between institutions running large language models (LLMs) of different families and sizes. Federated Verifier Pool (FedVP) instead has each client train a verifier on its local data and its own LLM's responses, and the server pools these verifiers to select answers without merging any parameters, leaving the LLM backbones frozen. Across five public medical question-answering datasets treated as clients, with Qwen and LLaMA backbones from 3B to 14B parameters, it improves all five clients and beats the strongest conventional federated baselines by 1.46 to 5.79 percentage points in mean accuracy.

arXiv:2611.13708 PDF

Deterministic, Reproducible Inference in Exact Integer Arithmetic, and Verifiable Distributed Inference as Its Consequence

Sayed Amir Karim (Pelloran Inc.) Floating-point inference can produce different bits for the same model across hardware, compilers, and thread counts. tritify is an open-source runtime that runs the forward pass of decoder-only language models, Mixture-of-Experts decoders, and the Segment Anything 2.1 stack entirely in exact integer arithmetic, yielding output that is bit-identical across runs, CPU architectures, and CPU/GPU while reproducing the float reference's decisions on the tested inputs. Because the computation can be split across devices with no effect on the output, the 106-billion-parameter GLM-4.5-Air generated its answer across three machines, none able to hold the model in RAM, bit-identically to a single-host reference. The authors frame this as enabling verifiable distributed inference, and explicitly do not claim a speed-up, confidentiality, or a guarantee against a dishonest server.

arXiv:2611.13730 PDF

Do Language Models Feel the Sentiment? Investigating Emotion Vectors

Ali Hassan Khan, Saad Bin Ather, Muhammad Saif, Maryam Bashir, Hajra Waheed Emotion vectors, directions in a language model's internal representations tied to emotional concepts, are tested for their role in sentiment classification in Qwen2.5-0.5B-Instruct and Qwen2.5-3B-Instruct. Probes using only eight directions extracted from GoEmotions reach 88.66% and 90.70% accuracy on binary TweetEval, and removing the emotion subspace costs 2.39 and 1.10 percentage points more accuracy than removing a matched random subspace; the effect is smaller and not significant on SST-2. Emotion-based correction classifiers improve accuracy on Amazon Polarity reviews but do not consistently beat margin-only baselines. The authors conclude that the results separate available information, intervention effects, and practical value, and do not establish a uniquely emotional mechanism.

arXiv:2611.13744 PDF

Mixed Signals: Weight-Space Programming for EEG and ECG

Dara Varam, Taraneh Aminosharieh Najafi, David Atienza Weight-space programming, in which a language model reads and edits a trained network's weights directly, is extended to small networks trained on EEG and ECG biosignal data. Frontier models located and fixed planted faults without being told what they were, and could write every weight of a network from scratch given only the architecture, the task, and a few training subjects. In one set of experiments a language model matched standard gradient-descent training using 6x fewer training subjects, and networks written by Claude Opus 5.5 beat hand-crafted EEG pipelines built from the same 5 subjects. The reported limits are that language models could not improve already well-trained classifiers and could not reliably identify a network's task from its weights alone.

arXiv:2611.13758 PDF

LLM Validation Makes Minimal-Pair Benchmarks Easier for Language Models

Guofang Tang Using LLMs in place of human judges to validate minimal-pair linguistic benchmarks is tested on BLiMP (English) and QFrBLiMP (Quebec French), the two public datasets that release every human forced-choice judgment. Three LLM validators agree with held-out human majorities more often than five-judge panels on BLiMP, yet catch far fewer pairs whose labels humans reject than a single judge does. Their errors are systematic: across 10 open language models, human-supported pairs that the LLM filter flags are 18.9 points harder than the pairs it accepts, in all 70 model-validator combinations. Filtering therefore removes items that models fail but humans do not, inflating accuracy more than random removal, a result that replicates on Dutch BLiMP-NL under a pre-frozen protocol.

arXiv:2611.13788 PDF

Beyond Passive Imitation: On-Policy Mid-Training Distillation for Language Models

Ruize Zheng, Xuyang Liu, Siyu Ren, Xiaobing Tu, Yinggui Wang, Jinkui Ren et al. Continual pre-training (CPT) with standard next-token prediction (NTP) treats every corpus token as a fixed target, and a controlled study on a 1B-token mathematical corpus shows the same recipe improves Llama-3.2-3B on math benchmarks while substantially degrading Qwen3-1.7B-Base. On-Policy Mid-Training Distillation (OPMD) instead samples trajectories from the evolving student at selected anchor positions and applies dense token-level reverse-KL rewards from a frozen teacher, so the model trains on states it actually visits rather than only on ground-truth prefixes. On MATH-500 and Minerva, OPMD improves the Qwen student over its base performance while the NTP baseline degrades on both benchmarks, with no added inference-time overhead.

arXiv:2611.13855 PDF

Understanding Mixture-of-Experts Language Models Beyond Expert Boundaries

Zhenhong Zhou, Mao Zheng, Mingyang Song, Qiankun Li, Kun Wang, Junfeng Fang Whether experts in Mixture-of-Experts (MoE) language models specialize by domain or merely by token and syntax is disputed, so the authors look at the neuron level using Routing-Aware Magnitude Profiling (RAMP). They find domain-specialized neuron sets with 58-74% overlap across samples of the same domain versus 15-24% across domains, and observe that neuron importance keeps reorganizing during pre-training while post-training largely preserves it. A derived pruning method, RAMPr, selects neurons across experts and preserves 91-98% of full performance at 75% neuron retention across six models. At 50% retention it cuts resident GPU memory by 36-43% and raises batched prefill throughput by 17-39%.

arXiv:2611.13975 PDF

On the information rate of phones, syllables and words

Herman Kamper Which linguistic unit encodes speech content at the lowest bitrate is tested on a transcribed corpus of 960 hours of English audiobooks, comparing phones, syllables, words, multi-word combinations and intermediate tokenisations. All units are equivalent under a perfect language model, so the comparison is made under limited modelling budgets using n-gram and transformer language models. Words achieve the lowest bitrates, with slightly longer units sometimes performing best, indicating that units larger than syllables but shorter than long multi-word phrases are the most efficient when modelling capacity is constrained.

arXiv:2611.13990 PDF

One Sequence Is Enough: Fingerprinting Language Models from Fixed-Length Integer Sequences

Fantong Meng, Chengjie Sun, Shuangjia Yang, Zhiyuan Zhao, Yansheng Pei, Xinyang Li Users of model APIs and third-party platforms cannot easily confirm that the model answering is the one declared, which motivates identification from outputs alone. NumPrint collects 5,036 answers from 51 models to Chinese, Japanese, and English prompts that request 294 integers between 1 and 355 in one response. A classifier called SMCT (Shallow MLP and Causal Transformer) captures aggregate numerical preferences and sequence patterns, reaching 92.72% accuracy from a single sequence in 51-class closed-set evaluation. Using two and four answers raises accuracy to 96.06% and 97.12%.

arXiv:2611.14010 PDF

Can KV Cache Be Reused Across Model Updates? A DeepSeek-V4-Flash Case Study

Zhikun Tao, Fangyong Jin, Junping Zhao cross-listed Reusing key-value (KV) caches after a language model's weights are updated could avoid recomputing shared context, but it is unclear what those caches still preserve. A study of direct reuse across four model updates, centered on DeepSeek-V4-Flash to Flash-0731, finds that caches can retain useful task information without faithfully reproducing the updated model's behavior, with retrieval accuracy ranging from near-native to large losses depending on the update. When source caches are already available, reuse yields median time-to-first-token speedups of 2.10 to 5.08x on three longer-prefix question-answering benchmarks. A new metric, Functional Cache Divergence, shows behavior is better preserved in natural-text continuation than in single-token retrieval with long prefixes, and agreement on the first generated token does not ensure agreement in answer content or stopping behavior.

arXiv:2611.14051 PDF

Sable: A Small-Field Sumcheck Argument for Bit-Exact Inference of Large Language Models

Tomer Solberg cross-listed Sable is an argument system and single-GPU prover for the claim that a committed, integer-quantized transformer produced a given token sequence from a given prompt, bit for bit. It defines an integer-only inference function with no floating-point steps and proves it using batched sumchecks, LogUp-GKR lookups, and a multilinear commitment over the 31-bit KoalaBear prime field, streaming the model layer by layer. On one NVIDIA RTX PRO 6000 Blackwell GPU, it proves all 64 layers of Qwen2.5-32B in 92.5 seconds for 4,096 tokens, which is 41 to 50 times the serving time, about 6.4 cents per request, and 40 to 64 times less time per parameter and token than zkLLM. The integer function stays within 1.6 points of the BF16 model on MMLU and 1.7 on GSM8K for Qwen3, and the proof is not zero-knowledge, so the system targets open-weight models.

arXiv:2611.14083 PDF

Hx2Dx: From Anamnesis to Diagnosis - A Multimodal Arena for Conversational Clinical Reasoning in LLMs

Federico Ravenda, Mirko Cesarini, Daniele Montagnani, Volodymyr Karpenko, Antonietta Mira, Andrea Raballo Static diagnostic evaluations hand a large language model (LLM) a curated case that already contains the relevant findings, which measures inference but not the inquiry needed to obtain that evidence. Hx2Dx ARENA is an interactive benchmark of 1,000 clinical cases from the New England Journal of Medicine, each simulated as an encounter that moves from initial presentation through history taking and physical examination to diagnostic workup. Across open- and closed-source models, every model loses diagnostic accuracy in the interactive setting relative to the static one, and performance varies considerably between models. The ability to ask targeted questions accounts for much of both the gap between models and the accuracy loss.

arXiv:2611.14122 PDF

BARS: Budget-Aware Sequential Sampling with Refusal for Majority Identification under Unknown Distributions

Shenxi Liu, Yiwei Li, Mingyang Zhao, Yuhang Tian, Kan Li Sampling a language model repeatedly and taking the majority answer enforces self-consistency, but existing early-stopping variants do not produce a correct confidence estimate from the answer distribution. BARS frames this as majority identification under an unknown distribution with a sampling budget and target error rate, using a sequential rule that certifies the gap between the top two answers and then either stops early, returns a budgeted majority vote, or refuses when the budget cannot establish the answer. A certified mode carries proven error bounds, while the practical mode used in experiments substitutes empirical calibration. On GSM8K the practical mode draws 22.3% fewer samples than fixed-8 sampling while declining 18.8% of questions, and on MATH it reaches 82.5% accuracy on answered questions by refusing about half of the difficult ones.

arXiv:2611.14171 PDF

Breakdown of Autoregressive Multi-Step Reasoning Under Concentrated vs. Diffuse Attention Head Dropout

Karol Struniawski Inspired by the localized periods of neuronal silence seen in sleep-deprived mammalian cortex, the authors test whether clustered bursts of attention-head masking harm step-by-step reasoning more than the same masking budget spread out as diffuse random dropout. On 250 paired GSM8K derivations with Qwen3-8B and LLaMA-3.1-8B-Instruct, concentrated bursts are clearly worse: Qwen3-8B accuracy falls from 91.2% under diffuse dropout to 78.8% under burst dropout, and LLaMA-3.1-8B-Instruct falls from 84.0% to 76.8%. An exploratory follow-up on 100 items links single-step dropout density to an acute accuracy drop, while the effect of multi-token continuity is inconclusive.

arXiv:2611.14209 PDF

Tree-Ising: Repairing Language Models with Bit-Level Ising Edits

Hadi Salloum, Ammar Ali Prior work on weight-space programming had a frontier LLM agent read a small model's weights and write a gradient-free repair for its reasoning loops, at a cost of 158 agent turns and 89.7 million tokens. Tree-Ising recasts the repair as sparse discrete optimization over the model's BF16 weights, where the allowed edits are sign flips and exponent shifts, and decomposes the objective exactly into independent QUBO (Ising) problems per output row that are solved on a GPU with simulated annealing or simulated bifurcation. On the same two layers of Qwen3.5-0.8B it reaches 54.6% on MATH-500 versus 51.7% for the agent's edit (base model 7.9%), using four forward-pass requests, zero LLM tokens, and about $0.90 of GPU time, while changing fewer weights and bits. On Qwen3.5-2B the same recipe raises greedy GSM8K accuracy from 21.6% to 69.4%.

arXiv:2611.14210 PDF

StaRSE: Compact Russian Sentence Embeddings with a Sign-Coded Static Encoder

Boris Malashenko, Ivan Jarsky, Egor Soldatov, Vladislav Luzhnov, Sviatoslav Stumpf, Vladislav Shmatkov et al. Transformer sentence encoders are costly when embeddings must be computed on CPUs or refreshed often. StaRSE is a static Russian sentence encoder with no Transformer inference at deployment: token directions are stored as sign bits, token norms separately, and sentences are encoded by mean pooling over norm-scaled sign directions, with the table initialized from ruBERT-base and trained contrastively. On RuMTEB it reaches a task mean of 51.16, which is 7.98 points above the strongest included static baseline. The packaged model takes 11.3 MiB and 0.08 ms mean CPU latency per sentence, compared with an estimated 55.7 MiB and 1.43 ms for the closest-quality contextual encoder.

arXiv:2611.14236 PDF

SinkState: On-the-Fly SVD of Recurrent States for Efficient Hybrid LLM Serving

Haolin Jiang, Weihang Li, Zhongzhu Zhou, Jinwen Pan, Santosh Pandey, Hang Liu Hybrid language models mix softmax attention with linear attention layers that compress history into fixed-size recurrent states, yet serving systems keep these states dense and rewrite them in full at every decoding step, so memory traffic dominates decoding cost at larger batch sizes. The authors observe that queries, keys, and values in linear layers concentrate in a stable low-dimensional subspace, which they call the linear attention sink, making each state effectively low-rank. SinkState stores each state in an adaptive low-rank form during inference using buffered updates and a local SVD on the top subspace, consolidating the remainder asynchronously on the CPU. Across four hybrid models with Mamba-2 and Gated DeltaNet (GDN) layers from 4B to 120B parameters, it reduces peak recurrent-state memory by up to 2.6x and speeds up decoding by up to 1.58x over Mamba-SSM and FLA while maintaining accuracy.

arXiv:2611.14246 PDF

LM-SimBench: A Unified Framework, Dataset, and Benchmark for Large Language Model Training Simulation

Yu Dong, Hong Li, Qiang Sun, Xu Zhou, Jinbiao Huo, Qixiu Cheng et al. Choosing a parallelization strategy for large language model training by running real distributed jobs is expensive, and existing performance simulators are hard to compare because they model different training stages and workloads. LM-SimBench provides a framework that decomposes simulation into stages from model compilation to schedule-level aggregation, together with a profiling dataset and benchmark. The dataset covers 23 model settings from 10 dense and Mixture-of-Experts model families profiled on an Ascend Neural Processing Unit (NPU) cluster, giving 348 feasible configurations that vary five parallelism dimensions and activation recomputation. Evaluations of representative simulators show how it supports comparing prediction accuracy and configuration selection, and matched GPU-NPU experiments demonstrate NPU-guided GPU configuration selection.

arXiv:2611.14285 PDF

MMDM: Measuring Massive Multitask Decision Making

Junhao Shen, Qing Zong, Yue Guo, Zhanqiu Zhang, Yiwen Guo, Hong Cheng System-1 decision models output structured probabilistic choices over given options in a single forward pass, without generating reasoning text, and existing benchmarks for them risk test-set exposure and do not combine long contexts, large option spaces, and logical dependencies. MMDM is a benchmark of 5,448 questions across 16 task categories covering single-select, multi-select, and soft-label decisions, including agent state tracking and tool selection, with up to 1,024 options and inputs up to 64,817 tokens. An evaluation of 28 decision models finds that many do well on common knowledge but struggle with interdependent options and constraints. For example, Perplexity Decider-V1.2 scores 78.75% hard single-select accuracy on world knowledge but 47.95% on constraint planning, and the best hard multi-select accuracy of any model is 36.81%.

arXiv:2611.14309 PDF

Auditing the Judge: Internal Signals for LLM Reasoning Assurance

Mayank Gulati, Gerhard Wunder Systems that use an LLM to judge its own answers need a score that ranks errors well at a limited review budget, not just on average. The study compares the judge's continuous verdict-token score with a supervised correctness probe trained on attention-head activations from the same prompt, and derives identities linking fixed-budget error recall to residual risk. Across three open 7-12B models on GSM8K and MATH, the probe catches 1.3 to 22.5 more percentage points of errors at a 5% review budget in all six settings (median 7.4), though for routing it beats the judge's own score only for the weaker judge. Probe quality drops under cross-dataset transfer, so the authors conclude that internal signals help in-domain budget allocation but are not domain-general confidence estimators.

arXiv:2611.14323 PDF

AI Isn't Getting Dumber, We're Getting Pickier: Rising Standards in Human Evaluation of Large Language Models

Zihan Chen, Lei Nico Zheng, Di Zhu cross-listed Complaints that AI assistants are getting worse could reflect changing judges and not changing models, and this study separates the two using 1,260,967 blind votes on Chatbot Arena between May 2023 and August 2024, restricted to pairs of large language models (LLMs) whose versions are fixed. For the same unchanged pair, the share of votes saying 'both are bad' rises by 1.71 percentage points a month on a mean of 17.1%, while the stronger model's share of decisive votes stays flat. The rise holds within the same voter, across arrival cohorts, languages, prompt controls and interface changes, and is shared across models. The authors read this as a standard that tracks the field's progress, so absolute verdicts drift while relative comparisons hold steady, which affects leaderboards and human feedback used for training.

arXiv:2611.14333 PDF

CLADE: Category-Level Answering from Deficient Evidence in Retrieval-Augmented Language Models

S M Rayeed, Charles V. Stewart, Yao Ma Retrieval-augmented systems often retrieve passages that are accurate but name only a broader category than the question asks for, such as a disease family instead of the specific disease. The authors call the resulting behaviour Clade (Category-Level Answering from Deficient Evidence): models answer from these coarse passages instead of abstaining or searching further. On biomedical questions from PrimeKG, family-only passages cut abstention by 13.8 points compared to uninformative passages, and the resulting guesses are wrong about half the time; confidence filtering, a larger model and explicit abstention instructions do not fix it, and the effect replicates on species interactions in GloBI. An ontology lookup contains the true answer in 98.5% of cases when given the correct family, but identifying that family is the bottleneck, and end-to-end recovery reaches only 26.8%.

arXiv:2611.14355 PDF

Blurring the Block Boundaries: Overlapping Block Denoising for Efficient Diffusion Language Models

Amir Mann, Jonathan Fhima, Shaked Brody, Or Avitan, Asif Zamir, Shahar Tsiper et al. Diffusion Language Models (DLMs) that use block diffusion with self-editing refine one block of tokens at a time, so each block is as slow as its slowest token and the next cannot start until it finishes. Blurring the Block Boundaries (B3) refines two blocks in parallel within a shared forward pass and calibrates on the fly when to open the next block. It requires no training or hyperparameters and applies to any self-editing block diffusion model. Across 10 datasets on current DLMs it gives an 11.5-14.5% speedup with on-par generation quality relative to the base model.

arXiv:2611.14387 PDF

REAL Loss: Optimizing Real Expected Acceptance Length for Speculative Decoding

Yu Tian Speculative decoding speeds up LLM inference by letting a draft model propose several tokens that the target model verifies in one pass, with speedup governed by the expected acceptance length (EAL). Existing training objectives, including recent block-level ones, are surrogates for EAL; some average later acceptance over target-distributed continuations instead of the branches actually selected by accepted tokens, creating a covariance gap that can reverse the gradient direction. The REAL loss targets EAL directly, using path weights for unbiased estimation and Rao-Blackwellization to reduce variance. With a Qwen3.6-35B-A3B target and a Domino drafter, REAL outperforms WTV, VSD, DBLast and a static LK baseline on acceptance length across 9 task subsets and 5 seeds, changing only the training loss.

arXiv:2611.14399 PDF

Bits or permutations? Spending a fixed inference budget on weight precision or on option-order averaging in small local language models

Mohammed Nayeem Khan, Logesh L, Aliyas Ahmad Misri For small local language models, a fixed inference budget can go to higher weight precision or to averaging predictions over reorderings of multiple-choice options, and this study asks which is cheaper and whether quantization worsens option-order sensitivity. Using simulated round-to-nearest quantization of Qwen3-0.6B and Qwen3-1.7B on ARC-Easy, ARC-Challenge, MMLU and HellaSwag, 6-bit weights lose no accuracy against fp16, while at 4 bits answer-slot skew rises (0.15 to 0.53 for the smaller model) and averaging four cyclic orders recovers 12.6 points. Pricing cost as weight bytes read times passes, the decision reduces to using 6-bit weights and spending the remaining budget on up to four option orders. A Gaussian-noise control is inconclusive on whether rounding adds skew beyond equal-power noise, and the authors note everything is simulated on two small models with the simplest quantizer.

arXiv:2611.14413 PDF

Precision on Demand: Escalating Weight Bits at Uncertain Decoding Steps in a Small Local Model

Mohammed Nayeem Khan, Logesh L, Aliyas Ahmad Misri Instead of fixing one weight precision for a whole generation, this study stores 8-bit weights, decodes with their top 4 bits, and redoes a step at 8 bits only when the low-precision output looks uncertain by entropy or top-two margin. With simulated quantization of Qwen3-0.6B on GSM8K, low-pass entropy predicts argmax flips with AUC 0.93, and at a 15% escalation budget the entropy gate catches 78% of flips and the margin gate 94%. On 150 live problems, an entropy gate escalating 16% of steps scores 62.7% at 0.87 of the full-precision weight bytes, against 64.0% for all-high and 55.3% for all-low, but the differences are not statistically separated from either end. The authors note that a live random control, a second model and other checks were not run.

arXiv:2611.14418 PDF

When does text encoding matter for language modeling?

Eugene Jang, Catherine Arnett, Terra Blevins UTF-8 is the default text encoding for byte-level language models, but its variable-length design treats languages unequally and its effect on model performance has been little studied. The authors compare UTF-8 against other Unicode-based, language-specific, and custom encodings, and propose Aether, a hybrid scheme that uses script-specific encodings for target languages while keeping full Unicode coverage through UTF-8 fallbacks. Across four orthographically diverse languages, Chinese and Korean, whose characters are information-dense, benefit substantially from compressive encodings, while Georgian and Tamil are largely insensitive to the encoding choice. The conclusion is that text encoding should be a per-language design decision rather than a universal default.

arXiv:2611.14459 PDF

14 more specialized papers

Agents 65

IntentTwin: Deconfounding Candidate Support and the Oracle-to-Deployment Gap in Offline Web Behavior Prediction

Miaobo Hu, Shuhao Hu, Xiaobo Guo, Xin Wang, Bokun Wang, Daren Zha Offline web-behavior benchmarks often insert the demonstrated action into a candidate set and measure reranking, which creates a confound when the target lies outside training support while all distractors are sampled from it. The authors formalize this with a support-identifiability rate (SIR), build support-balanced candidate sets, and evaluate a target-free dual-path proposer combining support retrieval with open-vocabulary DOM and language-model proposals. On filtered OPeRA, a support-only rule reaches perfect Exact@1 on out-of-vocabulary targets under the original protocol without modeling behavior; after balancing, the strongest reranker scores 0.7214 Exact@1 and the end-to-end target-free system 0.5212, and the same audit on Mind2Web gives an SIR of 0.2380.

arXiv:2611.12918 PDF

Using LLM Agents to Evaluate Prosocial Interventions on Social Media Platforms

Lucen Li, Erin Wertz, Muhammad Haroon, Anshuman Chhabra, Magdalena Wojcieszak cross-listed Few systematic methods exist for comparing social media intervention designs before running costly human-subject experiments. A simulation pipeline built on the OASIS framework generates 13,835 LLM-based agents with unique profiles, exposes them to real Twitter/X posts over 20 time steps, and tests three interventions that increase exposure to verified news: adding content, upranking content, and celebrity source cues. All three significantly increase news engagement, with upranking to prominent feed positions producing the largest effects, and the content-injection results show strong directional alignment with a comparable human-subject study.

arXiv:2611.12922 PDF

Auditable Human-Agent Teams for Autonomous X-ray Experimentation

Gunik Luthra, Ronald J. Pandolfi, Benjamin T. Diroll, Ricky Ronquillo, Ruipeng Li, Christopher B. Murray et al. Closed-loop autonomous experiments on physical instruments are costly to set up, and sustained agent operation on real hardware is largely untested. Using coding-agent infrastructure grounded in validated scientific methods, the authors ran autonomous synchrotron X-ray experiments on nanoparticle superlattices under changing human-set objectives, then audited the complete human-agent and agent-agent transcripts to reconstruct who decided what. Across three experiments, 61-67% of messages involved interpretation or judgment, and 59-68% of error corrections involved multiple participants, with about half of all decisions incorporating another participant's input.

arXiv:2611.12933 PDF

EarthRecord: From Pixels to Land Change with Multimodal Agents over Four Decades of Satellite Imagery

Haoran Zhang, Huizhen Sun, Hanbin Wen, Jie Leng, Jiaxin Zhu, Zhaohua Sun EarthRecord is a benchmark that tests whether multimodal agents can turn raw multi-sensor satellite archives into land-change analysis, leaving cloud masking and cross-sensor harmonization to the agent. It covers 56 sites with four decades of Landsat, MODIS, VIIRS, and Sentinel-2 data and poses 238 questions across ten tasks, with two controls: removing the record entirely, and altering one satellite's data while nothing changes on the ground. Among twelve agents evaluated, leading ones complete most analyses end to end but struggle when an answer requires mapping from imagery, and when one satellite's data are altered, half of the right answers follow the alteration rather than the real ground change.

arXiv:2611.12934 PDF

Agent Architecture Cost-Success Trade-offs: A Controlled Evaluation across Enterprise-Style SOPs

Subhrangshu Nandi, Hitesh Arora, Yunfei Bai, Udita Patel, Chirag Shah, Hannah Kwon et al. A controlled comparison holds the model, tasks, and tools fixed and varies only the agent architecture across 13 enterprise standard operating procedures (SOPs): a function-calling single agent, a skill-augmented agent, centralized multi-agent orchestration, and a triage swarm, in eight arms. No architecture is statistically separable in aggregate, and every coordinated arm spends 1.4-4.7x the tokens and 2.6-5.8x the latency of the single agent for gains that are usually small and rarely significant. High-quality skill definitions lift success from 67.4% to 73.1%, while degraded skills fall below the baseline. A plausible multi-agent implementation loses to the single agent, yet a targeted redesign of it becomes the only arm to significantly beat the baseline (75.6% vs. 67.4%).

arXiv:2611.12953 PDF

BuySideFlow: A Point-in-Time Benchmark for LLM Agents in Investment Research

Dong Bo, Fanjin Zhang, Ke Cheng, Xinxin Zhao, Xiaowei Su, Yang Wang et al. LLM agents that write SQL and Python workflows for investment research must respect point-in-time (PIT) validity, meaning they may only use data and record versions that were actually disclosed as of the analysis date. BuySideFlow is a benchmark of 408 practitioner-derived Chinese-language tasks across equities, funds, fixed income, and macroeconomics, built on a fixed snapshot of a licensed production financial database, with executable references, a taxonomy of temporal errors, and expert audits of executed workflows. Across nine agent configurations, result accuracy ranges from 16.2% to 36.8%, and an audit of 900 runs found PIT violations in 193 of the 380 runs that passed result checks (50.8%), so matching the expected output does not establish temporal validity.

arXiv:2611.12966 PDF

ICLayout-Bench: Benchmarking LLM Agents for Custom Circuit Layout Design

Zhangmai Li, Tianrui Ma, Zheng Weng, Mo Zou, Zidong Du, Yifan Hao et al. cross-listed Whether LLM agents can produce high-quality custom integrated circuit layouts has not been systematically tested. ICLayout-Bench has agents drive electronic design automation (EDA) tools to build and optimize GDSII layouts from transistor-level netlists, then scores layouts that pass physical verification, parasitic extraction, and post-layout simulation on electrical performance and area; it contains 165 tasks across 13 process design kits (PDKs) from 3 nm to 180 nm, spanning analog, mixed-signal, and digital circuits. Across 22 configurations of seven models, mean layout scores on a 100-point scale range from 4.22 to 80.26 on public tasks and from 1.43 to 67.01 on private tasks. Higher reasoning effort does not consistently raise scores, and configurations with similar scores can differ substantially in model cost.

arXiv:2611.12982 PDF

Compile-Gated Lean Proof Refactoring with Enterprise Agent Practices: A Model per Step under a US\$3 Cap, Followed by Programmatic Rewrite Rules

Bobber Cheng cross-listed A competition entry for refactoring Lean proofs under a US$3 per-problem cost cap is described, in which every candidate rewrite must pass the compiler before it is accepted. Each paid stage uses whichever of nine tested models measured best per dollar (GPT-6.1 Sol, GPT-6.1 Terra, and Gemini 4 Argon for different steps), lessons from development runs are shipped as compiler-checked programmatic rewrite rules instead of prompt text, and every call is logged and reviewed before changes. The entry scored 86.31 points on the 60 final problems versus 33.33 for the reference proofs, placing 8th (top score 92.57) at a total cost of US$166.33. The rewrite rules contributed roughly 0.3 to 0.9 points each, and the team's internal meter predicted board scores to within about 0.06 points.

arXiv:2611.13011 PDF

From Questions to Decisions: Benchmarking How LLM Agents Acquire Decision-Relevant Information from Human Beings

Yuhan Tang, Ke Fan, Xinkai Zou, Baichuan Mo, Kangxin Cui, Ao Qu et al. Agents asked to plan for people often receive incomplete requests and stale records, yet benchmarks usually supply fully specified problems or rely on uncalibrated simulated users. Ask2Decide is a benchmark of 43 planning tasks in which an operator holds missing facts, optimization solvers determine which facts change the optimal plan, and similarly worded irrelevant facts act as distractors. Using 79 sessions from 19 human participants, the authors calibrate a simulated operator after finding that humans omit the correcting fact far more often than an uncalibrated simulator (43% versus 7%). Across four frontier models, initial plans are infeasible on 58-98% of tasks when information is missing, versus 5-28% when all facts are given upfront, and agents uncover plan-changing facts only slightly more often than irrelevant ones (39-52% versus 32-41%).

arXiv:2611.13015 PDF

ChipOmni: A General-Purpose AI Agent for Electronic Design Automation

Pengkun Wang, Rui Mao, Zhe Zhao, Dongting Xie, Haibin Wen, Song Xu et al. Electronic design automation (EDA) workflows are fragmented across representations, tools, and verification stages, and existing language-model approaches mostly address isolated tasks. ChipOmni is an agent framework that combines multimodal understanding, dynamic planning, multi-agent collaboration, and verification-driven tool use, maintaining an evolving task context from engineering tool feedback. It improves task-specific measures over the evaluated baselines on analog circuit, RTL, testbench, and formal-assertion generation, and solves all 30 AnalogCoder-Pro tasks. In a long-horizon case study it takes a natural-language RISC-V specification to a micro-SoC GDSII layout that passes physical sign-off without human intervention after initialization.

arXiv:2611.13020 PDF

Repository-Context-Guided Image Understanding for Agentic Software Development

Fumiki Yohena, Toshiaki Sota, Shota Suzuki, Tomoko Okamoto, Michiaki Tatsubori cross-listed Coding agents often receive UI screenshots in bug reports, but low resolution, compression artifacts, and small text can prevent them from identifying the relevant UI elements and source files, and OCR-based workarounds propagate recognition errors. The proposed fix is to give the agent crisp reference screenshots rendered from the same repository it is modifying, which show the current, possibly buggy UI and need no known-good version or manual annotation. On a controlled benchmark built by injecting visual UI bugs into a web application from SWE-bench Multimodal, reference images improve identification of the correct source file, particularly under fourfold downsampling with JPEG compression. The approach also recognizes text about 12 percentage points better than language-biased OCR under the same degradation.

arXiv:2611.13026 PDF

Beyond the Document: Compiled Context and Governed Agents for Any Work Product

James Rhodes, George Kang cross-listed Documents preserve the visible result of professional work but drop the research, calculations, assumptions, evidence, and authorization boundaries behind it. The authors study a Portable Agent Format (PAF) package that ships a work product with compiled context, provenance, and interaction rules, and evaluate four packages in finance and engineering over 400 frozen questions and 12,000 responses. A hosted governed reader succeeds on 96% of questions versus 86% for instructions plus raw-file retrieval, while source-material chat and instructions alone reach 48% and 74%. A separate 0.8-billion-parameter offline reader reaches 69% overall with a 2.9% rate of materially wrong or unauthorized answers, but its retrieval resolves none of 40 historical-record questions in one package, and standards interoperability and end-to-end governance conformance remain unproven.

arXiv:2611.13055 PDF

ASPIRE: An Agentic System-Identification Platform for Iterative Reasoning and Exploration

Erik Derner, Ji\v{r}\'i Kubal\'ik, Robert Babu\v{s}ka cross-listed Nonlinear system identification depends on expert choices about the method, the use of prior physical knowledge, and validation. ASPIRE hands these choices to a large language model agent that selects and runs the identification method, producing closed-form static or dynamic models from measured data with documented parameters, term interpretations, independently verified physical requirements, and an error computed on test data the agent cannot access. On three measured systems (a mobile robot, a rotational system with dry friction, and a laboratory fermenter), all 30 runs delivered valid models at a cost of 0.96 to 3.38 US dollars per run. In the authors' own assessment of the recommended models, 132 of 150 documentation and interpretation items were fully satisfied and 18 partly, the most frequent problem being inaccurate or unsupported claims about model structure selection.

arXiv:2611.13114 PDF

Which Successes Make Good References? Action-Sequence Alignment for Multi-Turn On-Policy Distillation

Mufei Li, Nikki Lijing Kuang, Zhuoer Wang, Samson Koelle, Jie Yuan, Manqing Mao et al. On-policy distillation (OPD) supervises a student on its own trajectories, and for multi-turn tool-use agents the teacher can be conditioned on a successful trajectory for the same task, but a success that passed through incompatible interaction states can give misleading supervision. The authors frame reference selection as a free-target optimal transport problem that reduces to assigning each failed trajectory its most compatible success, and use longest common subsequence (LCS) alignment over action sequences as an observable proxy for compatibility. On ALFWorld and the airline domain of tau^3-bench, LCS is the only evaluated reference selector that improves over reference-free OPD on both ALFWorld overall scores and all three airline metrics. The airline-trained models also transfer to the retail domain without further optimization, where LCS again improves all three metrics.

arXiv:2611.13135 PDF

Autonomous Optimization of Numerical Algorithms by an AI Agent: Application to Atomistic Simulation

Jeheon Woo cross-listed AIOA is a framework in which a large language model coding agent improves numerical algorithms for atomistic simulation against a pre-specified metric, inside a loop guarded by verifiers the agent cannot alter. Applied to ASE, a widely used atomistic simulation library, it targets both implementations (faster code with unchanged results) and the methods themselves (convergence in fewer force evaluations). Agents made six kernels 22 to 24 times faster on average in Python with NumPy and SciPy. On unseen problems, the best agent-designed optimizers needed 3.1 to 3.4 times fewer force evaluations than L-BFGS for structure relaxation and 4.6 to 17 times fewer than standard climbing-image nudged elastic band calculations for reaction paths, and contain two algorithmic components the authors did not find in the literature.

arXiv:2611.13143 PDF

Teaching the Harness: Student-Aware Skill Transfer for LLM Agents

Chang Liu, Bo Wu A procedure that helps one large language model (LLM) agent can exhaust the action budget of another running in the same harness, so the authors ask whether a stronger model teaches a fixed student better when it explicitly models that student. In Student-Aware Teaching (SAT), the teacher builds a learner model from an empirical profile and hypotheses about the student's limitations, tests them with diagnostic probes, and writes reusable skills that enter the library only if they raise success on matched development tasks. With Claude-Sonnet-5 teaching DeepSeek-V4-Flash, SAT reaches 86.46% task success, against 81.04% for a profile-free teacher and 75.00% for the untaught student. Skill libraries work better for the student they were written for than when swapped between DeepSeek-V4-Flash and Gemini-3.8-Flash, and saved libraries keep average gains of up to 11.1 points after the teacher is removed.

arXiv:2611.13144 PDF

AgentScript-Eval: Benchmarking LLMs and Agents for Code Generation in an Enterprise DSL

Dipin Khati, Shubham Mehrotra, Bin Bi, Zhou Yu, Denys Poshyvanyk, James Zhu et al. cross-listed Code generation ability of large language models (LLMs) has been measured mostly on languages abundant in public training data, which enterprise domain-specific languages (DSLs) are not. AgentScript-Eval is an execution-graded benchmark of 100 real tasks in AgentScript, the declarative DSL Salesforce uses in Agentforce, covering authoring, debugging, and optimizing agent definitions. Zero-shot, the best of five models reaches only 25.0% pass@3 and every model scores 0% on authoring from scratch, with a 31B open-weight model achieving the highest compilation rate at the lowest token cost. An agentic compile-and-edit loop roughly doubles partial-credit accuracy, and the wording of domain guidance matters more than its packaging: the same skill pack scores 28% as plain documentation, 14% as a live plugin, and 37% after rewriting one skill.

arXiv:2611.13183 PDF

From Process Models to Composable Agentic Workflows: Combining BPMN, Agents and Enterprise Knowledge

Christian Fillies, Jan Fillies, Frauke Weichhardt, Gay Anne Wood-Albrecht Complex LLM agent applications scatter their decisions across prompts, tool calls, and application logic, which makes workflows implicit and hard to monitor or debug. The authors convert Business Process Model and Notation (BPMN) models authored in SemTalk into executable workflows on a LangGraph/Pregel runtime, where selected activities invoke LLM reasoning while conventional BPMN constructs handle control flow, deterministic decisions, system calls, and human approvals. Their central idea, Recursive Capability Composition, exposes agents and whole executable business processes as reusable tools for other agents. Process models and knowledge graphs additionally supply organization-specific context for agent decisions; the work describes the approach and reports no quantitative evaluation in the abstract.

arXiv:2611.13196 PDF

Deletion Sensitivity Is a Property of the Model: A Three-Condition Measurement across Seven Open Reasoning Models

Hanzhang Jia, Liheng Zeng, Zhiyu Yang, Yi Gao, Bo Ma Reasoning models write a reasoning trace on every turn of a multi-turn tool session, but providers bind traces to their issuer, so requests forwarded across providers arrive with that history stripped. A blind measurement across three rendering conditions, seven open-weight reasoning models, three seeds, and 794 tool-calling tasks finds that the same deletion changes task scores by anywhere from -4.4 to +51.3 points depending on the model, while for the one model tested at varying deletion levels the reading moved by at most 1.5 points. The authors conclude that deletion sensitivity is a stable per-model property that can be screened with two runs and should inform routing decisions.

arXiv:2611.13223 PDF

ExpertCrowd-SWE: Measuring Frontier Coding Agents on Expert-Authored Engineering Tasks in Unfamiliar Codebases

Josip Vrdoljak, Akash Singh, Bas Steunebrink, Alex Yarotsky, John Lorenzo, Achille Nicoletti et al. ExpertCrowd-SWE (EC-SWE) is a benchmark of 100 multi-file software engineering tasks written by practising engineers against pinned commits of real repositories, spanning 87 open-source projects and 11 languages, with sealed reference solutions and executable graders that award partial credit and enforce merge-blocking criteria. Each grader is calibrated so the reference scores 1.0 and no-op changes score 0, and many tasks include lazy and adversarial control patches. Across seven frontier models with five runs each, Claude Opus 5.5 (67.3%) and GPT-6 Astra (63.7%) lead without a statistically significant gap, the rest score between 32% and 46.5%, and step counts per run show little relation to score. Twenty-four tasks are released publicly and 76 are held out to limit contamination.

arXiv:2611.13246 PDF

Dependable Long-Horizon LLM Agents: A Trajectory-Centered Empirical Study of Where Errors Enter and Which Controls Reach Them

Shaohuang Wang, Jinghao Wen, Xulei Zhang, Xun Jiao Failure analyses of long-running LLM agents usually record a symptom or a responsible step rather than where the error entered, how it propagated, and which control could have caught it. The authors define an episode record linking a fault's origin, propagation, detection, response, and consequence, and apply it by fully reading 95 failed runs of six multi-agent systems from the Who&When and MAD corpora. Faults enter early (median 18% into the run), and in 72% of failed runs the trajectory already contained evidence against the error, yet in 88% of those no agent acted on it; independent judges, task contracts fixed up front, and re-checking beliefs against the environment would reach most episodes. They also find that 45% of hand-crafted Who&When failures ended in a provider content-filter crash, after which the harness emitted an answer no agent produced.

arXiv:2611.13254 PDF

MRFoundry: Knowledge-guided agent loops for end-to-end MRI research

Ke Wang, Atakan Topcu, Hossein Momeni, Molin Zhang cross-listed MRFoundry is an open-source package that connects magnetic resonance imaging (MRI) domain knowledge — acquisition physics, data conventions, specialized tools, and validation checks — to the actions of AI coding agents, with project-local memory for decisions and evidence and a review process for shared guidance. The authors also introduce MRI Agent Bench to compare agents with and without the package across models and task domains. Across 240 paired core evaluations spanning 24 task types and ten models, all-checks-passed completion rose from 67.5% to 78.3% with the package, and on 134 hard pairs from 36.6% to 44.8%.

arXiv:2611.13289 PDF

PegasusAI Studio: Legible Cyberinfrastructure and Accountable AI Agents for Scientific Workflows

Komal Thareja, Mats Rynge, Berent Aldikacti, Angelina Ju, Ian Lumsden, Anirban Mandal et al. cross-listed Running AI coding agents on shared scientific cyberinfrastructure requires that agents know the current workflow system and execution sites, and that facilities can attribute every agent action to a user and bound it by that user's privileges. PegasusAI Studio, a science gateway built on the Pegasus and HTCondor stack, projects one facility-curated knowledge base into the native configuration of several coding agents, introspects workflow interfaces and sites, and grounds diagnosis in workflow and scheduler telemetry. Agent actions run inside a trust envelope of federated authentication, per-user execution environments and credentials, observability, and approval controls for state-changing actions. Evaluation is by case study only, including weather workflow modernization, a gravitational-wave reanalysis, and an Event Horizon Telescope image reconstruction that recovered from two failed runs to a zero-failure execution.

arXiv:2611.13309 PDF

Agent-Guided Discovery of Exchange-Correlation Functionals with Physics-based Validation

Achuth Chandrasekhar, Amir Barati Farimani Whether language-model agents can discover exchange-correlation functionals for density functional theory whose gains survive self-consistent and external evaluation is tested by connecting frontier LLM agents to a GPU evaluator through the Model Context Protocol, with training-only fitting and persistent memory. Opus-5.5 agents passed the screening gate with about 40-fold fewer tokens than Codex (GPT-6-Astra) agents and found the better candidate, while SyFES evolutionary search did not improve on its GAS22 seed in about 400 fits. The resulting 13-parameter functional, O159, multiplies same-spin correlation by a smooth density window. Evaluated self-consistently, it lowers the held-out MGCDB84 test error of ωB97M-V by 11.7% and is comparable to it on 36 GSCDB137 datasets.

arXiv:2611.13329 PDF

Do AI Agents need Preventive Care? An Epidemiologic Cohort Design for Agentic AI

Shu Jiang, Graham A. Colditz Evaluations of AI agents mostly measure capability on predefined tasks and say little about how behaviors develop and affect outcomes over time. The proposal is an epidemiologic cohort design that treats individual agents as cohort members followed longitudinally, with the underlying model as genetic makeup, accumulated work as the time scale, and naturally occurring behaviors as exposures. This allows study of behavioral determinants of adverse events and estimation of preventable burden under causal assumptions. Two prospective observational cohorts illustrate it: one on curiosity and probing as a safety question, and one on data checking and analytical accuracy as a reliability question.

arXiv:2611.13333 PDF

Veritas: A Verifier Generation Harness for Agent Tasks

Daoan Zhang, Qinbo Bai, Dehai Min, Xingyu Mu, Ziyi Chen, Jianing Yin Self-improving agent systems depend on a verifier to judge each round, and those verifiers are mostly written by hand. Veritas is a harness that generates verifiers automatically using an evidence header stating what the verifier will observe, a library of domain rules that tasks leave unstated, and authoring principles partly refined by a self-improvement loop. With the model held fixed, its verifiers agree with official verdicts better than those of seven other harnesses on all four execution-scored benchmarks and are at or near the best on two pairwise-comparison benchmarks. Adding its mechanisms to three other harnesses improves 9 of 12 harness-benchmark pairs scored by execution.

arXiv:2611.13336 PDF

DEPUTYBENCH: Evaluating Contextual Safety in Enterprise Agents

Hao Zhao, Minghao Lin cross-listed Enterprise agents have to decide whether an ordinary-looking request is actually permitted by organizational records and policies, while still completing legitimate work. DEPUTYBENCH is an executable benchmark of 79 base items and 375 instances with paired permission contexts and graded evidence requirements; across 12 models it yields 88.0% pooled safety and 79.1% paired success. In a controlled study on six models, agents carried out restricted actions more often when authorization evidence was missing than when permission was explicitly denied, 10.3% versus 3.5% on 27 matched tasks, a secondary-analysis gap concentrated in four privacy tasks. The authors conclude that high safety when evidence is available does not establish reliable stopping when permission cannot be verified.

arXiv:2611.13355 PDF

The AI Era's Lines of Code? A Case Study on What an Engineer's Token Spend Measures

Shyam Agarwal, Olga Lvova, Agnia Sergeyuk, Anastasiia Serova, Timofey Bryksin, Thomas Zimmermann et al. cross-listed Organizations track engineers' token spend to monitor AI adoption and productivity, though its relationship to engineering output is unclear. The authors link five weeks of coding-agent token usage to the commits of 404 engineers at JetBrains and analyze a survey of 412 employees. Spend varied mostly between engineers; among weeks with similar commit counts the 90th percentile of spend was 39 to 240 times the 10th, and within an engineer doubling spend was associated with only about 10% more commits. Technical respondents reported using agents for non-coding work such as analysis, research, and planning in nearly nine of ten evaluations, so the authors recommend using token spend at most for resource accounting.

arXiv:2611.13396 PDF

Stinging the Swarm: Extracting Attacker Intelligence from a Twin-Controller OT Honeypot

Canyon Grey Dell'Omo, Joao Pedro de Carvalho Voltani cross-listed Autonomous language-model agents can drive industrial controllers over fieldbus protocols and probe a plant at machine speed, so Talos places an inline gate in front of the controller that judges each request in milliseconds and reroutes agents flagged as hostile to a device-faithful mirror twin of the plant. The detect, quarantine, mirror and extract pipeline was tested against 2,400 attacker agents across four tiers of increasing concealment effort, and no write reached the real plant in any tier, with no inference timeouts from the small on-device judge. Attackers fall along two axes, persistence and disclosure, and the system captures intelligence from three of the four resulting regions, with only paranoid, low-yield agents resisting. Self-reported toolchain fingerprints and targeted-register lists accumulate across the swarm into a profile even from agents ordered to stay silent.

arXiv:2611.13406 PDF

SteerDistill: Quantizing LLM Agents via Closed-Loop Teacher Steering

Genming Cui, Xiangyu Li, Zixu Hao, Xin Ding, Yunxin Liu, Ting Cao Low-bit quantized LLM agents fail tasks by overlooking requirements or environment conditions, and neither offline data nor the student's own rollouts supply enough correct responses in those failure contexts. SteerDistill has a full-precision teacher guide the quantized student to revise failed responses, refining the guidance from the student's execution feedback, then trains the student to reproduce the corrected behavior without guidance. Across five agentic benchmarks and three models it beats the strongest quantized baseline by 1.67 to 2.18 points under W4A16 and stays within one point of the full-precision BF16 model. On τ²-Bench it outperforms QeRL by 1.84 percentage points in success rate with 67% fewer training GPU hours.

arXiv:2611.13410 PDF

BioMem: Agentic Memory for Biomedical LLM Agents

Mingchen Li, Yongkang Xiao, Rumeng Li, Rui Zhang, Hong Yu How reliably agentic memory systems retain, retrieve, update and use biomedical information is evaluated across five memory approaches and four LLMs, in knowledge reasoning, clinical prediction and long-term conversational settings. The analysis separates the challenges of retrieving relevant evidence, preserving its fidelity and using it in reasoning. Building on those findings, Hybrid Memory Retrieval and Utilization (HMRU) is a training-free framework that improves both evidence accessibility and task-aware use. It yields an average relative improvement of 8.6% on biomedical question-answering and conversational memory benchmarks, cuts conversational answer errors by 24.9% and closes 48.0% of the gap to full-context inference.

arXiv:2611.13426 PDF

Deep Personalization: Continuous Context Observation for Local Proactive Agents

Yijia Shao, Song Bian, Junxian Guo, Jindong Jiang, Zhen Dong, Huizi Mao et al. Proactive agents need to know when an intervention will help rather than disrupt, which requires understanding a user's ongoing activity. Coco, a proactive desktop agent, uses a separate Observer Agent that monitors desktop activity on-device, keeps compact long-term user context and learns intervention timing from implicit signals such as engagement, dismissal and repeated work patterns. For evaluation without real user data, a synthesis pipeline converts computer-use trajectories into OSWorld-Persona, a benchmark with persona-conditioned intervention labels. Adaptation raises macro F1 from 0.12-0.19 under direct prompting to 0.35-0.48 across three model backbones, and a week-long user study saw it rise from 0.199 to 0.515.

arXiv:2611.13457 PDF

Mobile AI Factory: Building Specialized Language Models for Your Mobile Device

Evan Li, Krishna Deshpande, Wen Cheng, Shyamnath Gollakota Small language models (SLMs) on phones only perform well when fine-tuned for the task and sized for the device, which currently takes repeated expert tuning. Mobile AI Factory is an agentic system in which an orchestrator LLM decides whether to tune hyperparameters, mine real data, or generate synthetic examples, while fixed rules roll back harmful changes and the search uses real measurements from the target phone across inference engines and backends. Across ten tasks and four Android phones, its on-device models beat two five-shot frontier models on five tasks and match them on three more, and its search comes within 0.03 of exhaustively training all 27 model variants while using 19× less GPU time. The best configuration differs across phones, and without phone measurements Claude Opus 5.5 misses the app's requirements in 25% of easy and moderate cases and 60% of hard ones.

arXiv:2611.13509 PDF

The RSI Tax: When Recursive Self-Improvement Is Not Worth the Cost

Yitong Zhang, Ang Li, Yisen Wang Recursive self-improvement (RSI), where an agent improves itself from its own task experience, is only worthwhile if the gains justify the cost of the improvement process. The authors define the RSI Tax as the token cost of self-improvement and include it alongside task-execution cost when evaluating six methods and two language models on continual task streams. A simple bounded-note baseline matches or exceeds the accuracy of Meta-Harness, EvoSkill, and RRSI while using 65.6–71.0% fewer total tokens on average. For those three methods, verification makes up 74.8–93.6% of the tax, and a substantial share of the spend goes to states or artifacts never used later in the stream.

arXiv:2611.13524 PDF

Agent Benchmarks Stop Ranking Before They Run Out of Tasks

Zihan Dong, Qishi Zhan, Minxuan Hu When agent benchmarks stop separating top models, the usual response is a harder version; re-scoring released runs from eight benchmarks, including MLE-bench, ALE-Bench, SWE-bench Verified, and Terminal-Bench, suggests much of the saturation comes from pass/fail measurement rather than the tasks. Pass/fail scoring is too coarse, since neighbouring top-ten agents on SWE-bench Verified differ by less than the error in its tests, and it is capped at the reference, since ALE-Bench's top ten tie when graded against the median human even though nearly half the runs beat that human. The authors derive a graded clipped log-ratio score with dated anchors and cost reported as a separate axis. In a Terminal-Bench re-run, models tied on pass/fail differ by more than tenfold on the quantity the tests discard.

arXiv:2611.13541 PDF

Darmok: Championship Cryptic Crossword Solving via Constraint Tightening and Resilient Orchestration

Siddhartha Chatterjee (The University of Texas at Austin) Darmok is an automated solver for British cryptic crosswords, where clues hide a definition and wordplay behind a misleading surface, nearly half of each word's letters are unchecked, and a single wrong square ranks a grid below every error-free one. It cross-checks clues across three independent models in rounds, uses verified crossing letters to unlock harder clues, and fills remaining gaps with a whole-puzzle model and dictionary lookups without overwriting confirmed words. At the 55th Times Crossword Championship it solved all seven puzzles error-free and faster than every human competitor in every round, finishing the Final in 6m 52s, 3m 13s ahead of the human champion. Ablations show the backup models and fallback steps barely change accuracy under normal conditions, yet they carried the solve on the worst operating days and during a severe slowdown of the primary model in the live Final.

arXiv:2611.13557 PDF

Megrez: A Knowledge-Guided Agent for Formal Mathematics over MK Set Theory

Tiancheng Xu, Wensheng Yu Formalizing a textbook cumulatively inside a low-resource proof library is harder than solving standalone competition problems, because concepts must be aligned with existing declarations, long searches must stay synchronized with live proof states, and verified proofs must be merged reliably into a growing codebase. Megrez coordinates a general coding agent for Rocq libraries built on Morse-Kelley (MK) set theory using a rebuildable knowledge graph for retrieval, a feedback-driven proving loop that issues short bursts of 1-5 tactics with a branch ledger and periodic structured reflection, and an acceptance pipeline that checks statement fidelity, audits axioms and runs full builds. Evaluation on a point-set topology corpus is reported to show effective cross-chapter formalization and scalable premise retrieval, without headline numbers in the abstract.

arXiv:2611.13613 PDF

Synapse-Control: Making Agent Experience Actionable with Lightweight Runtime Control

Zhiyuan Huang, Shengan Zheng, Chenghu Yang, Song Zhang, Lingyun Li, Chen Yu et al. Agents that reuse past execution experience often pay for it with extra large-model calls, as in systems like Sentry that generate a failure diagnosis before retrieving advice. Synapse-Control replaces that step with a lightweight control plane: rules shortlist relevant advice, a fine-tuned 1.7B classifier scores the candidates, micro-contracts encode reusable interface checks such as required arguments, and a Prompt Value Table limits repeated suggestions through cooldown and token-budget checks. Across 1,308 attempts per method on AppWorld and WebShop, it reaches 270 successes, 29.8% more than the base agent, while replacing 4,987 generative diagnosis calls and cutting auxiliary control time relative to Sentry by 50.2% and 51.0% on the two benchmarks.

arXiv:2611.13615 PDF

DrugCoding: A Scientific Programming Agent for Drug Discovery

Junwei Hu, Zhuohao Lin, Kun Li, Hengnan Zhang Drug discovery workflows chain together molecular design, property prediction and virtual screening across many specialized tools, which existing large language model agents struggle to plan and execute reliably. DrugCoding is a scientific programming agent system that combines task planning, adaptive routing to specialized scientific models, multi-agent collaboration, and programmatic tool execution through generated code. In multi-turn tool-calling experiments across several language models and tool interfaces, it reports an average session-level Pass@1 of 97.35% on the full benchmark, outperforming the strongest baseline, Codex, by 9.24%, and 97.92% on the challenging subset.

arXiv:2611.13651 PDF

On Private Knowledge Retrieval and Private Inference for secure multi-agent systems

Vicen\c{c} Torra, Maria Bras-Amor\'os cross-listed Agents in multi-agent systems built on large language models (LLMs) can be targeted by memory poisoning attacks, in which a malicious agent alters another agent's stored knowledge to change its behavior. Focusing on semantic memory, the authors propose an architecture in which each agent runs a local Prolog-like inference engine over Horn clauses and privately queries an external, cryptographically signed knowledge base so that tampered knowledge can be rejected by design. They provide Python implementations at several privacy levels, including k-anonymity and information-theoretic private information retrieval (IT-PIR).

arXiv:2611.13696 PDF

Agentic AI Meets Formal Verification in the Era of Agentic Hardware Development

Wenji Fang, Yuchao Wu, Jing Wang, Zhiyao Xie An invited overview of how agentic AI and hardware formal verification intersect, motivated by the need to check the correctness of AI-generated circuit designs. It first covers how agents advance formal methods at three levels: specification, proof construction, and reasoning engines. It then covers how formal methods in turn support reliable AI-driven hardware development through register-transfer-level code generation, debugging, and correctness-preserving optimization. The authors review representative work alongside their own results and outline a direction they call proof-native hardware development, where formal reasoning is built into autonomous design.

arXiv:2611.13705 PDF

AutoManifold: Agentic Design of Data Visualisation Algorithms via Manifold Embedding

Burak Susam, Ruiyuan Kang, Tingting Mu Manifold embedding algorithms for data visualisation are hand-designed with fixed targets and cannot easily be steered toward the structural properties a particular user cares about. AutoManifold is a multi-agent large-language-model (LLM) pipeline that composes new embedding algorithms from a constrained vocabulary of affinity, cost, and optimisation primitives, conditioning each design stage on user-specified preservation preferences and refining the configuration in a metric-grounded loop. Compared against t-SNE, UMAP, and PaCMAP on real-world datasets under identical evaluation, the authors report that the generated algorithms outperform the hand-designed methods at steering toward user-specified properties without compromising the data's intrinsic structure.

arXiv:2611.13710 PDF

Toward Autonomous RTL and Synthesis Optimization: Benchmarking and Self-Improving Agents

Wenji Fang, Yao Lu, Jinghua Wang, Zhiyao Xie Three directions toward autonomous optimization of register-transfer-level (RTL) hardware designs and logic synthesis are examined: realistic benchmarks, tool-grounded RTL rewriting, and design-specific synthesis tuning. The centerpiece is Dr. RTL, a multi-agent framework that automates critical-path analysis, RTL rewriting, and implementation-level evaluation while accumulating reusable optimization skills. On 20 real-world designs it achieves average timing-slack improvements of 21% and 17% (worst and total negative slack) with a 6% area reduction over commercial synthesis. The authors also explore agent-driven synthesis script tuning and discovery of executable synthesis algorithms.

arXiv:2611.13713 PDF

DisasterAgent: An Evidence-Driven Agent for Verifiable Disaster Report Generation From Remote Sensing Imagery

Zhijian Li, Jiawei Song, Chao Ren, Yidan Zhang, Peijin Wang, Xian Sun et al. cross-listed Disaster assessment from remote sensing imagery increasingly calls for full reports whose statements responders can verify, which vision-language models (VLMs) and tool-augmented agents struggle to back with evidence. DisasterAgent follows a Perception-Augmentation-Generation workflow: a vision expert toolkit builds a Spatiotemporal Perception Graph (STPG) of changes between before and after images, and an Evidence-based Augmentation step attaches inline citations linking each statement to its supporting visual evidence. On the DisasterM3 and RSCC benchmarks it outperforms general-domain VLMs, remote-sensing-specialized models, and existing remote sensing agents, raising the average report-generation score on DisasterM3 by 13.5% and ST5-SCS on RSCC by 15.34% over the original model.

arXiv:2611.13776 PDF

TreeAgent: Structure-Aware Memory Agent for Processing Long Hierarchical Data

Wei Gu, Weiwei Qi, Jiajun Tang, Tianhang Zheng, Bo Chen, Zhibo Zhang et al. Memory agents that process inputs longer than a model's context window usually cut them into flat fixed-length segments, which splits records, hides hierarchy, and loses details of earlier sibling records in structured data such as API responses or HTML. TreeAgent instead builds segments from complete subtrees and groups of adjacent siblings, attaches structural paths so the agent knows each segment's position, and adds residual memory links from child segments to parent segments. It comes with theoretical guarantees on segment counts for choosing token budgets. Across eight retrieval and aggregation tasks on JSON, HTML, YAML and JavaScript inputs of 8K to 1M tokens, it improves all four tested memory agents, with a 3.61% average performance gain.

arXiv:2611.13946 PDF

From Attack Surface to Defense Architecture: Compositional Security for Agentic Coding Harnesses

Simon Hegelich, Bardh Prenkaj, Habiba Sarhan, Morteza Shahrezaye, Kolja Hegelich, Christian Montag et al. cross-listed AI coding harnesses connect language models to shells, filesystems and networks, so injected instructions can turn ordinary binaries such as bash and curl into a Living-off-the-Land (LotL) attack payload. ActiveSecHarness is a reference-monitor architecture that separates the model's intelligence from its authority, combining HMAC capability tokens across four trust rings, a two-dimensional taint lattice over trust and sensitivity with a workspace provenance ledger, and deterministic operating-system and tool guards. On 252 fixtures, the deterministic Ring 0 guard blocks every attack with zero false positives in 12.7 microseconds, whereas a local LLM judge (qwen3:4b) refuses 98.3% of benign work at 142.3 ms. In multi-step tests, a stateless policy baseline misses 4 of 12 kill chains and Boolean taint blocks 83.3% of benign workflows, while the two-dimensional lattice blocks all attacks without false positives.

arXiv:2611.13960 PDF

Controlled Knowledge Engineering for Taxonomy Induction with Tool-Grounded LLM Agents and Runtime-Adaptive Policies

Jonas Limniatis, Bennet Schaarmann, Georg Reuber, Robert Stahlbock cross-listed Building a taxonomy from a fixed set of candidate entities with large language models is prone to omitted or hallucinated entities, invalid parent-child relations and inconsistent organizing principles. TIGA (Taxonomy Induction with Tool-Grounded Agents) treats the task as an iterative decision process over an explicit state, using executable taxonomy-editing tools, a memory of tool calls and outcomes, an organizational policy the agent revises at runtime, and an optional global reviewer. On four benchmark taxonomies with two model backbones, it achieves the strongest structural quality in most configurations but does not dominate on every precision measure, and it costs substantially more tokens, API calls and inference time. Ablations show consistent benefit from the policy and memory components, while the reviewer's effect depends on dataset and revision budget.

arXiv:2611.13966 PDF

BioXArena-Pro: a research-grade benchmark for autonomous biomedical machine learning agents

Loka Li, Duzhen Zhang, Xingbo Du, Leonard Song, Kun Zhang, Le Song Existing benchmarks for coding agents that build biomedical models are small, near saturation, or open to leaked information. BioXArena-Pro offers 64 research-level modelling tasks from eight biomedical domains, each rebuilt from a published study and mostly holding out whole biological units such as species or cohorts. Across 22 agent configurations the best scores 55.1 out of 100, and the scaffold matters almost as much as the model. Agents tie re-run published methods where the original recipe applies unchanged and lead by 8.2 points on rebuilt tasks, while forbidding pretrained models costs the strongest agent 9 points on the tasks where it uses them.

arXiv:2611.13994 PDF

Coding Agent Memory Post-Training II: Unlocking the Memory Potential of Pre-trained File Operations for Long-Horizon Tasks via On-Policy Distillation

Lirui Luo, Kelong Mao, Xinwei Yang, Rongqing Li, Luyu Chen, Kieran Wong et al. Agents on long-horizon tasks must decide what to write down and read back once the interaction history exceeds their context, and prior on-policy distillation (OPD) approaches rely on memory tools that neither teacher nor student saw in pre-training. CAMG-II is a suite of long-horizon agentic environments spanning Shop, Coding, AutoResearch, General, and Visual that adds ordinary Bash access and an episode-persistent workspace, so files serve as memory across context replacements. CAMP-II trains one policy across all five environments with fully asynchronous OPD from a frozen teacher that already has a prior for file operations. It obtains 2.97 times the gain of the strongest OPD baseline on CAMG-II, and the learned memory behavior improves success rates on Terminal-Bench 2.1 and SWE-Bench Pro V2.

arXiv:2611.14009 PDF

Personalized Memory Rubrics: Learning What to Remember per User via Rubric-Guided Credit Assignment

Yupeng Han, Kai Zhang, Shuochen Liu, Ze Liu, Xianquan Wang, Zhihong Pan et al. Personalized agents with evolving memory are typically trained only on final question-answering accuracy, a coarse signal that treats every intermediate memory update alike and ignores that the same update can help one user and hurt another. PMR (Personalized Memory Rubrics) splits evaluation into a universal layer of cross-persona failure patterns and a per-user layer modeled as a delta over that baseline, then uses the rubric to redistribute credit at the token level on successful trajectories while falling back to standard policy optimization on failed ones. The rubric and memory policy co-evolve during training, and no rubric or external judge is needed at deployment. Experiments on four benchmarks show consistent gains over memory-augmented baselines.

arXiv:2611.14017 PDF

Program-Aided Patchable Skill Evolution

Jiahao Wang, Yuqi Xiong, Qing Yang, Zhipeng Xu, Zhenghao Liu, Zulong Chen et al. Agent skills are usually stored as natural language documents that mix procedural logic with descriptive knowledge, which makes them hard to maintain and evolve from execution feedback. PASE (Program-Aided Patchable Skill Evolution) attaches executable programs to skills and moves reusable procedural logic into them, so that updates can patch specific code rather than repeatedly rewrite prose. Across multiple models and benchmarks it outperforms all baselines, with improvements of over 2.8%, and further analyses indicate greater stability during evolution and better generalization.

arXiv:2611.14019 PDF

ProofCoder: Continuous Reasoning for Automated Software Verification

Martim Afonso, In\^es Lynce, Maurizio Morisio, Jo\~ao F. Ferreira cross-listed Large language model (LLM) approaches to proof synthesis usually confine the model to a rigid tactic-generation loop in which the surrounding system controls premise retrieval, proof-state management, and backtracking. ProofCoder is an agentic system for the Rocq proof assistant that works like a coding agent, reasoning continuously and using proof-assistant feedback to decide when to gather information, inspect proof states, revise earlier fragments, or backtrack. On 280 verification tasks from the AutoRocq Benchmark, its best configurations raise proof success from 47% to 60% over AutoRocq at roughly half the cost. Ablations show that without continuous reasoning the agent solves fewer lemmas at higher cost and backtracks four to six times more often, while whole-proof and tactic-level generation reach similar success rates.

arXiv:2611.14106 PDF

Model-Free Runtime Measurement of Task-Relevant Information

Dimitris Panagopoulos, Adolfo Perrusqu\'ia, Yang Xing, Weisi Guo cross-listed Autonomous systems can stay operationally active without actually resolving their mission requirements, because existing methods do not monitor whether live information gathering is working. Acquisition-State Telemetry (AST) is a model-free architecture in which a domain interface maps observations into task-resolution coordinates and a fixed causal monitor emits two real-time signals: prospective progress across unresolved requirements, and informational friction when repeated evidence yields negligible gain. With a single unadjusted configuration across eight domains, including search-and-rescue dialogue, grid exploration, and robotic manipulation, the friction signal distinguished unproductive stagnation from purposeful exploration before task completion. In closed-loop reinforcement learning experiments with early termination, penalizing friction let agents sustain high mission success where unpenalized policies and cost-matched baselines failed.

arXiv:2611.14111 PDF

Verifiable Agentic Traffic Class: A Policed Quality-of-Service Mechanism for Agentic AI Traffic, Evaluated from Simulation to Real-Kernel Adversarial Testing

Naveen Babu Narayanan Nair, Sajay Kurup Traffic from LLM-based software agents differs from human web traffic, but routers have no class for it and cannot verify a sender's claim to be an agent. The Verifiable Agentic Traffic Class (VATC) combines a reserved Differentiated Services Code Point (DSCP) value declaring agent traffic, a router-side classifier that demotes flows contradicting the declaration, and a rate-limited priority queue, and is compared against static DiffServ, a trusting dedicated class, and fq_codel in two simulators and a real Linux kernel. Results are mixed: VATC matched DiffServ under steady load and under a spoofed flood in the kernel its rate limit held human latency 11 times below DiffServ's, but the live classifier missed six spoofers that copied the agent pattern and a timing feature failed under emulated network delay. The authors note that all traffic was synthetic or sandbox-scale.

arXiv:2611.14211 PDF

A Cross-Scenario Survey and Analysis of Self-Evolving Multimodal Agents

Jingjing Jiang Self-improvement mechanisms for multimodal agents are typically designed and tested within a single scenario, so it is unclear whether they transfer. This survey compares four scenarios (multimodal reasoning, multimodal generation, GUI and Web navigation, and embodied manipulation) using a three-stage loop of experience generation, feedback acquisition, and update consolidation, positioning each scenario by its ceilings on rollout count and feedback fidelity. Each mechanism is recast as a trade-off that raises a scarce factor by relying on a task-structural resource such as output equivalence or checkable outcomes, and is judged portable, conditional, blocked, or redundant across scenarios. The authors attribute asymmetric transfer to differing combinations of scarce factors and resources, and note that these verdicts are untested and name the matched-budget measurements that could overturn them.

arXiv:2611.14256 PDF

Don't Just Trust Your Memory: Evidence-Linked Memory Reduces Memory-Induced Degradation in LLM Agents

Yukyeong Lee, Yumin Choi, Jinheon Baek, Sung Ju Hwang External memory lets LLM agents distill past trajectories into lessons, but it can also make an agent fail tasks it would solve without memory, because lessons may be wrong or misapplied and arrive without the trajectories needed to check them. WikiTrace stores each lesson as a concise insight linked to a summary of why its trajectory succeeded or failed and to the raw trajectory itself. On a new task the agent starts from all insights and uses terminal tools to follow those links whenever an insight lacks detail, conflicts with another, or comes from a failed task. Across five agentic benchmarks it consistently outperforms existing memory approaches and a baseline given the same insights alone, while reducing memory-induced degradation.

arXiv:2611.14320 PDF

When Should Agents Remember? Falsification-Gated Self-Evolution for LLM Agents

Runxuan Tang, Haoyu Gao, Yuyan Ding, Junyi Yao, Zihao Zheng, Zecheng Sheng et al. LLM agents that write lessons from individual episodes straight into memory can store knowledge that is neither reusable nor safe. Falsification-Gated Self-Evolution (FGSE) represents each candidate lesson as a structured hypothesis with a precondition, behavior change, expected effect, and verifier, applies it only in a temporary state, and tests it on transfer, falsification, and archived-regression probes before committing, refining, or rejecting it. It reports a 39.8 success rate on WebArena, 78.8 average completion on AppWorld, and 0.724/0.487 Pass^1 on τ-Retail/τ-Airline. It also reduces harmful committed updates to 3.6%, with regression damage at 1.3% and falsification failures at 5.0%.

arXiv:2611.14326 PDF

EngineAgent: Agentizing On-Device LLM Inference Engines with Inference Stack Hierarchy Awareness and Hardware Abilities

Zhengxiang Huang, Yujie Sun, Shengheng Chen, Zhaode Wang, Chaoyue Niu, Zeyu Zhao et al. cross-listed Adding support for a new model to an on-device LLM inference engine requires coordinated changes across conversion, graph construction, operators and hardware backends, which general coding agents tend to miss. EngineAgent first builds a hierarchy-aware index of the engine's abstraction layers and cross-layer dependencies and pre-validates build, execution and debugging workflows on already supported models; a builder then makes the changes for a new model while an auditor checks completeness and drives debugging through on-device validation. Across 12 tasks covering three mobile LLMs and four engine-backend pairs on CPUs, GPUs and NPUs, it reaches 75% end-to-end success versus 33.3% for the strongest coding-agent baseline, where success requires compilation, on-device execution and logit agreement with a reference implementation. It also lowers average per-task time and API cost relative to baselines using the same underlying LLMs.

arXiv:2611.14356 PDF

ARE-bench: Benchmarking Agents' Research Ability in Real-World Researcher Sessions and Workspaces

Jiahang Lin, Wanying Ren, Wanghan Xu, Huayu Sha, Yang Shu, Zhonghang Lu et al. Research-agent benchmarks mostly test standalone tasks in clean workspaces, whereas real requests arrive mid-project and depend on earlier decisions and messy files. ARE-bench contains 75 tasks built from recruited researchers' real agent sessions and full workspaces, replayed without cleanup to the point of a real user request and scored with 1,173 binary checklist items across seven research stages. Among seven frontier models the strongest passes every checklist item on only 66.2% of tasks, with Design and Execution among the hardest stages. Three models pass about 69% of tasks when only delivery contracts and general norms are checked, but just 28.4% to 40.0% once historical and workspace requirements are also checked.

arXiv:2611.14364 PDF

ReTrace: Distilling Tool-Using Agents into Planner-Free Workflows

Kaiyuan Hou, Yang Ping Kuo, Indranil Bhattacharya, Rohith Nama, Arushi Gupta, Carlos D. Castillo et al. Large Language Model (LLM) agents can follow a Standard Operating Procedure (SOP) written in natural language, but calling an LLM to choose every step of every run adds latency and cost when the same procedure repeats at scale. ReTrace turns an SOP plus a few logged agent runs into a reusable workflow in which code selects the next step, extracting the condition behind each step from the explanations the agent wrote alongside its tool calls, merging them into a compact rule document, and compiling that into an executable workflow for a given orchestration platform. On tau2-Bench the workflows reach 75.8% task success against the agent's 80.6%, running in 1.3 seconds instead of 170, and on 79 production procedures they reach 93.8% completion against 96.8% at 24% lower cost per task. On eight SOP-Bench procedures the unrepaired workflows trail few-shot generation, but three rounds of repair from labeled training failures lift them to 87.8% at 0.40 seconds per task against 85.5% in 46.6 seconds for the agent.

arXiv:2611.14492 PDF

Comparative Analysis of Agent Evaluation Frameworks: Stability, Detection, and Discovery in Enterprise Analytics Agents

Howard Zhang, Chuck Arvin, Logan Brown, Mihir Bhatia, Neelu Choudhary, Lily Miao Three evaluation frameworks for LLM-based enterprise analytics agents are compared across three phases on two agents using frozen execution traces, each framework representing a distinct paradigm: trace-based LLM judging, text-only LLM judging with red-teaming, and deterministic heuristics. Repeated measurements find 97% of configurations meet the stability threshold, with stability patterns independent of the agent. On 72 ground-truth test cases, the trace-based framework achieves the best precision-recall balance at an F1 of 0.90, while text-only evaluation matches its recall but generates five times more false alarms. A third phase uncovers 24 novel failures with zero overlap between functional testing and security red-teaming, leading the authors to recommend pairing trace-based evaluation for functional assessment with adversarial red-teaming for security.

arXiv:2611.14493 PDF

Continual-ARC: Measuring Skill Acquisition over a Working Lifetime

Daniel Lenton Benchmarks for continual learning usually cover one applied domain, serve each instance once, and assume a single kind of learner, so memory-based agent harnesses and weight-based methods are rarely measured on the same scale. Continual-ARC builds a lifelong stream from the three ARC-AGI editions, where tasks recur after dormant gaps with fresh instances while new tasks keep arriving, and learners are scored by how much of a paid feedback budget they consume; afterwards they are frozen and tested on fresh instances of seen tasks to measure memorization and on unseen tasks to measure generalization. Running an agent harness with frozen weights on Qwen3.5-9B and GLM-5.3 Flash, with and without persistence, against prompting floors, a program-induction method, and Qwen3.5-9B trained to imitate GLM-5.3 Flash, persistence rarely improves memorization without explicit expert guidance. Most harness-based methods learn less than a floor that simply caches each task's demonstrations, and none gains on unseen tasks, whereas weight-based learning shows substantial gains.

arXiv:2611.14505 PDF

Contrastive Discovery: Open-Ended Scientific Discovery over Competing Explanations

Ziang Liu, James J. Kim, Yijia Dai, Jennifer J. Sun LLM agents doing data-driven scientific discovery usually test a hypothesis only against its own negation, so a rejection tells the agent that one explanation is unlikely but not which alternative the evidence favors. Contrastive Discovery replaces the single hypothesis with a contrast set: a question paired with mutually exclusive, collectively exhaustive competing explanations, where each experiment is scored by how it redistributes belief across the set. Across six domains, 89% of the pointwise candidates with the largest belief shifts were rejections, and contrast sets resolved 73% of them to a specific alternative using the same evidence. When used to guide search, the agent resolved 67% more questions than pointwise surprisal under the same experiment budget, and 38 human experts in a blinded study picked the same explanation on 78% of resolved questions.

arXiv:2611.14514 PDF

2 more specialized papers

Other 48

When Proofs Are Abundant, What Is a Paper For?

Charles-Albert Lehalle cross-listed With AI systems producing formally verified proofs at falling cost, this essay asks what a mathematics paper still communicates once correctness is cheap. The author argues that a paper is a narrative exposed to community criticism, not just a verified statement, consisting of the author's choice of problem and perspective plus a chain of reasoning, neither of which machine-generated formal scripts convey. Drawing on logic and type theory, these are formalized as a Derivation-Level Intension and a Directional Extension, the latter decomposing a proof into cut-points whose axiomatic footprints show where a result could be strengthened or transported to other theories. Five roles AI can play in research are examined, and none is found to remove the need to declare these two elements.

arXiv:2611.12943 PDF

Predictive Surprise as Self-Grounding Concept Bottleneck for Interpretable Time Series

Sachith Abeywickrama, Emadeldeen Eldele, Min Wu, Xiaoli Li, Chau Yuen Time-series classifiers in risk-sensitive settings need to be accurate, interpretable, and able to abstain, but defining meaningful patterns requires segment boundaries and vice versa. ConceptTime resolves this with a frozen probabilistic forecaster: the mismatch between its prediction and the observed signal, called predictive surprise, both locates segment boundaries and characterizes each segment, and the segment summaries are clustered into a vocabulary of dynamical-regime concepts. A lightweight head classifies from the concept sequence, with distance to the nearest concept serving as a reliability score. Across 17 public benchmarks it is competitive with black-box accuracy, beats every interpretable baseline, and retains near-98% accuracy on Epilepsy with 1% of labels, without annotations, language model calls, or domain experts.

arXiv:2611.12955 PDF

One Table, Many Queries: Benchmarking Generalist Tabular Models Across Generation, Imputation, and Prediction

Yuanrui Wang, Kaifang Li, Xuanyue Li, Ziyun Li, Yan Lu, Xingxuan Zhang et al. Generation, imputation, and prediction all query the same underlying tabular distribution but are usually evaluated separately, leaving it unclear what they share and whether one model can do all three well. The One Table Benchmark covers 56 datasets, 46 methods, and 20 capability axes under aligned splits, and additionally trains five backbones under all seven combinations of the three objectives. Rankings link imputation and prediction, and connect generation to both through predictive utility, while distributional fidelity follows different and sometimes opposing rankings; objective benefits are not additive and depend on which objectives are already present. In fixed-model comparisons, some generalists surpass specialist portfolios on utility, while fidelity still favors specialization, and frozen tabular foundation models show strong prediction and numerical imputation.

arXiv:2611.13001 PDF

Benchmarking Vendor-Guided Optimizations on Edge NPUs: From NPU Only to End-to-End Performance. Extended technical report

Youngjin Jo, Jihwan Kim, Minseok Choi, Kwanghoon Choi, Jusung Kang, Jaein Kim Performance of edge neural processing units (NPUs) depends on model transformation and the deployment pipeline as well as on hardware, so the proposed benchmarking methodology separates these effects with three configurations: a baseline, vendor-guided NPU-only (VG-NPU, which keeps common model input/output and host-side processing), and vendor-guided end-to-end (VG-E2E, which follows the vendor's native pipeline). Applied to four edge NPUs and three vision workloads, VG-NPU improves both accuracy and runtime-call latency in 10 of 12 combinations, while VG-E2E cuts system-level latency in all combinations but improves accuracy in only four. VG-E2E gives a further 10.4-70.7% reduction in system-level latency over VG-NPU, and NPU-level rankings do not always match end-to-end rankings.

arXiv:2611.13563 PDF

MOAI: Module-Optimizing Architecture for non-interactive secure transformer Inference

Linru Zhang, Xiangning Wang, Jun Jie Sim, Zhicong Huang, Jiahao Zhong, Huaxiong Wang et al. cross-listed Homomorphic encryption (HE) lets a cloud provider run transformer inference directly on encrypted inputs without interaction, but its computational overhead is a major bottleneck. MOAI reduces that cost with a consistent column-and-diagonal packing strategy that avoids format conversions between layers, rotation-free algorithms for Softmax and LayerNorm that remove 2,448 HE rotations in BERT-base inference, and interleaved batching for ciphertext-ciphertext matrix multiplication. It achieves a 52.8% reduction in evaluation time compared with the prior state of the art THOR and a 55.7% reduction when applied to the Powerformer framework, with an amortized 2.36 minutes per input on a single GPU. The authors also show an extension to LLaMA-3-8B and release the implementation as open source.

arXiv:2611.13641 PDF

Antalia-2 Mini: A 7.6-million-parameter Turkish text-to-speech model trained entirely on synthetic speech

Sezgin Saygili, Emre Kaplaner, Oncel Ozgul, Fikri San Koktas Antalia-2 Mini is a single-voice Turkish text-to-speech model with 7.6 million parameters that reads characters and produces 48 kHz audio at 41 times real time on two threads of an Apple M5 laptop CPU. Its roughly 1,790 hours of training audio were generated entirely by VoxCPM2, a teacher about 300 times larger, and filtered automatically; fine-tuning the vocoder on the acoustic model's own mel spectrograms cut the word error rate (WER) of Whisper transcripts from 3.89% to 1.75%. On the 495-sentence Freya-TR-Eval benchmark it scores 1.09% WER and 0.22% character error rate, statistically indistinguishable from the 8.6M-parameter EMA Lightning, which scores higher on the UTMOS quality proxy. No human listening test was run, and weights and inference code are released under Apache-2.0 while training code and data are not.

arXiv:2611.13811 PDF

Earlybird: Predictive Streaming Speech Recognition in 29M Parameters

Sahil Mahendrakar cross-listed Streaming speech recognizers normally emit a word 90-400 ms after the speaker finishes it, which delays voice agents. Earlybird, a 29M-parameter streaming recognizer (16.8 MB at 4 bits), is trained on audio cut off at random points, including mid-word, to predict every word due by then in full plus a guess at the next two words, with token probabilities serving as per-word confidence. On LibriSpeech test-clean, the median word is predicted correctly 90 ms before the speaker finishes it, and 69% of words are correct before they end versus 3-25% for other models; it is the earliest on all six timed test sets. Confidence for in-progress words has an expected calibration error (ECE) of 3.7 to 5.5, and accuracy beats Moonshine Tiny on all eight English Open ASR Leaderboard sets while roughly tying Whistle.

arXiv:2611.14421 PDF

Two Communities, One Problem: Branch-and-Bound vs. A* for Resource-Constrained Project Scheduling

Izack Cohen, Dominik Kress, Lara Nehrke, Erwin Pesch, Shimrit Shtern cross-listed Branch-and-Bound (B&B) from operations research and A* from artificial intelligence are compared as exact search methods on the Resource-Constrained Project Scheduling Problem (RCPSP), after relating its mixed-integer linear programming, state-action, and timed Petri-net representations. Deliberately basic implementations share the same lower bound and best-first node selection, so the effect of the search mechanism, especially detecting equivalent states reached through different decision sequences, can be isolated. On 480 benchmark instances, graph-based A* with additional pruning rules proves optimality for all 480 instances, against 445 for best-first B&B under the same time and memory limits. Depth-first B&B uses less memory and still finds feasible solutions when it times out, while stronger linear programming relaxations shrink the search space at considerable computational cost.

arXiv:2611.14426 PDF

When Does the Real Wiring Win? Testing connectome structure-function claims against unmeasured physiology

Yeyuan Liu Connectome-constrained models are often used to argue that measured wiring beats shuffled or rewired wiring, but such comparisons can confound wiring with tuning, since a connectome fixes connectivity and not the excitability, gains, or plasticity rates that modelers adjust until the real-wiring model works. The authors separate three estimands, physiology tuned on real wiring then transferred to nulls, each wiring tuned on its own, and the latter restricted to physiology consistent with independent data, and test them on synthetic ground truth, placebo nulls, and three Drosophila circuit models. Segregated reward and punishment dopamine and head-direction topography survive every protocol, but in the mushroom body the real Kenyon-cell output wiring beats an identity shuffle by 0.09 under the first two protocols and loses by 0.09 under the data-constrained one, with the verdict set by the share of a cell subtype that the first protocol selects implausibly low. They conclude that a credible structural claim needs matched optimization, data-constrained physiology, and a model that still works under those constraints.

arXiv:2611.14508 PDF

39 more specialized papers

Theory 41

Everything in its right place: attention is a radio head

Llu\'is Arola-Fern\'andez, Aissam Khadraoui, Alex Arenas cross-listed Attention is recast as a physical synchronization process instead of a digital computation with an exponential, a normalization, and a positional code. Each past token is modeled as an oscillator whose state carries its content and whose frequency drifts with its age, and the current token acts as a receiver that selects by content and by position. Solving the Kuramoto synchronization problem for this receiver gives a closed-form attention layer with two phases: above a critical coupling it locks and its weights normalize, and below it stays silent with weights summing to less than one. Trained networks use both phases to select, abstain, and count, fail where the theory predicts, and keep reading well beyond the training length.

arXiv:2611.13146 PDF

Continuous Horizon Forecasting

Malcolm Wolff, Elliott Zackrone, Souhaib Ben Taieb, Michael W. Mahoney, Kin G. Olivares cross-listed Multi-horizon time-series forecasting usually emits a fixed-length vector or builds forecasts recursively. The Continuous Horizon (CH) strategy instead treats the forecast as a continuous operator and the horizon as a query variable, generalizing recursive, direct, and multi-step strategies, with a universal operator approximation theorem and a bias-variance decomposition proving lower estimation variance than discrete strategies. On a simulation study and 16 datasets from the M-series benchmark, CH reaches competitive RMSE while enabling parallel inference, querying of specific horizons, and interpolation or extrapolation to unseen horizons.

arXiv:2611.13228 PDF

An Analytical Model of Compute-limited Multistage Training Pipelines

Nishil Patel, Jin Hwa Lee, Basile Confavreux, Andrew Saxe How pretraining, supervised fine-tuning (SFT), and reinforcement learning (RL) interact, and how compute should be split among them, is studied through an analytically tractable high-dimensional model combining a linear autoencoder, noisy stepwise labels, and sparse episodic rewards. Using random matrix theory, the authors derive low-dimensional dynamics matching simulations and show that finite pretraining data caps performance, with error scaling as the inverse square root of the pretraining sample ratio across stages. SFT warm starts shorten the initial RL plateau caused by sparse rewards, an effect also observed in OLMo-2 1B on GSM8K, and the optimal compute allocation depends on supervision quality and budget, with the RL share varying non-monotonically.

arXiv:2611.13276 PDF

No finite rulebook can certify every safe program: an impossibility theorem for gatekeeping AI agents

Tomas Perez-Acle A gatekeeper that must decide from code alone whether an AI-written program may run would ideally have complete lists of what is allowed and what is not, each rule backed by evidence a machine can check in finite time. The authors prove this is impossible: no finite, machine-checkable rulebook can certify every safe program, because properties phrased as never or eventually are not settled by any finite record. For prohibitions every violation can be caught by observation but no safe program can be cleared permanently, while for confidentiality and responsiveness neither side can be settled, and any sound rulebook of n bits omits a compliant program of about n bits. A gatekeeper can still guarantee a checkable list containing everything it must allow and nothing it must refuse, and simulations on 2.2 million random programs indicate the obstruction is adversarial rather than typical.

arXiv:2611.13367 PDF

You Don't Need Feature Learning to Do In-Context Learning

Joseph Turnbull, Jamie Simon To understand memorization versus generalization in in-context learning (ICL), rotation-invariant kernel machines are analyzed on the ICL linear regression task of Raventos et al., with the task decomposed in the kernel eigenbasis. As sample size grows, the kernel machine learns successively higher-degree components of the memorizing solution, and the generalizing solution is approximately a low-degree truncation of it, so the model generalizes first and then degrades into memorization. The authors conclude that generalizing on this task amounts to underfitting the memorizing solution and that feature learning is not needed for ICL. Trained transformers are shown to memorize through the same higher-order task statistics the kernel theory predicts.

arXiv:2611.13431 PDF

Sharp Sample Complexity of KL-Regularized Reinforcement Learning

Andrew Liu The number of samples needed to learn the optimal policy in KL-regularized reinforcement learning, the objective behind maximum-entropy RL, was not known even with a generative model. For a discounted Markov decision process with K state-action pairs and effective horizon H, the plug-in estimator is shown to find an ε-optimal policy with on the order of KH³/(ε·max{ε,β}) + KH samples, matched by a lower bound that holds even for adaptive sampling in the stated regime. When ε is below the temperature β this gives a 1/ε rate with the same horizon dependence as unregularized RL. The exact asymptotic constant and the optimal adaptive sample allocation are also derived, and simulations on small problems match the rates.

arXiv:2611.13480 PDF

Parallel Broyden Methods for Efficiently Evaluating Nonlinear State Space Models

Ian Christopher Tanoh, Scott W. Linderman Nonlinear state space models are usually evaluated step by step, and existing parallel alternatives either need costly dense Jacobian products (exact Newton) or converge slowly because diagonal approximations ignore coupling between state dimensions. The proposed quasi-Newton method uses Broyden's method to build diagonal-plus-low-rank approximations of Jacobian blocks from trajectory secants, avoiding automatic differentiation and staying compatible with parallel scan through rank compression. The authors prove finite-step recovery of the exact trajectory under the compressed scan and characterize convergence in terms of compression and block approximation error. On low-rank RNNs, parallel MCMC algorithms, and denoising diffusion models it converges in fewer iterations than other quasi-Newton methods and is the fastest parallel method when Jacobian evaluation is expensive or state dimensions are large.

arXiv:2611.13523 PDF

Beyond the Log-Linear Frontier: Pre-Asymptotic Analysis of Finite-Budget Neural Scaling Laws

Tianyu Zhao, Zhixiang Ren Classical neural scaling laws treat excess risk as a separable power law in model capacity and dataset size, but real learning curves and compute-optimal allocations deviate from this at finite budgets. The analysis attributes these deviations to spectral and domain heterogeneity, deriving a three-term expansion along the capacity axis, a water-filling solution for mixing data across domains, and a proof that the optimal allocation exponent drifts along the iso-compute frontier in proportion to the variances of the capacity and data exponents. It also supplies a well-conditioned closed-form initializer for fitting on narrow pilot grids. Across 45 operator spectra, Pythia, Chinchilla, and RegMix data, the estimators reach 5.89% holdout mean absolute percentage error and cut high-compute allocation regret by up to 76.4%.

arXiv:2611.13534 PDF

Pair-Merge Tokenisation Is Hard to Approximate on Binary Strings and Strongly NP-Hard on Unary Datasets

Kundan Baliga, Aiden Xie, Aditya Byju cross-listed Tokenisers such as Byte-Pair Encoding (BPE) shorten text by repeatedly merging adjacent symbol pairs, raising the question of whether an efficient algorithm can pick k merges that leave a length-n string nearly as short as possible. The main result is that, unless P = NP, no polynomial-time algorithm approximates the optimum within a factor c·n/log n, even for a single binary string with BPE-style merges that replace every occurrence; the proof treats k merges as a k-rule grammar and amplifies a vertex-cover gap into a grammar-size gap. For merges that may skip occurrences, a simple algorithm achieves O(n/log log n), and Karp reductions cannot prove hardness beyond order n/log n unless coNP is contained in NP/poly. Over a one-letter alphabet, two tokenisation problems left open by Kastreva et al. are shown to be strongly NP-complete on datasets with repeated strings.

arXiv:2611.13568 PDF

Transformers Provably Learn Graph Search: Training Dynamics and the Exponential Power of Depth

Xutao Ma, Somayeh Sojoudi Path-finding over directed graphs is used as a symbolic abstraction of multi-step reasoning to study what transformers can represent and learn. One-layer transformers are proved to need size on the order of N squared on dense graphs to implement the child-selection step of depth-first search (DFS), while a two-layer transformer implements full DFS with size O(N log N). For every k of at least 3, a k-layer transformer implements a bounded-reachability DFS policy with look-ahead horizon 1 + 2^(k-3), an exponential increase with depth. Given appropriately curated training data, two-layer transformers are also shown to provably learn DFS through gradient flow and generalize to unseen graphs.

arXiv:2611.13594 PDF

CP-MLPs: A Tensor-Rank Theory of Tied and Untied MLP Blocks

Md Rifat Arefin, Farzaneh Heidari, Irina Rish, Guillaume Rabusseau Transformer feed-forward blocks mix squared activations, gating, and tied versus untied parameter sharing, and the separate effect of each on expressivity is poorly understood. The CP-MLPs tensor framework treats each multiplicative unit as a rank-1 interaction with detector, value and output roles: tied blocks such as ReLU^2 force detector and value to share a direction, while untied blocks such as ReGLU and SwiGLU let them differ. The analysis yields exact tensor-rank separations showing that untied units are strictly more width-efficient for indefinite quadratic interactions and role-asymmetric monomials, while tied units are matched to symmetric pure-power targets, so the advantage of untying depends on the target's structure. Experiments show loss plateaus below the exact-representation width, though the authors note these alone do not establish an approximation gap, which is governed by border rank.

arXiv:2611.13647 PDF

Stagewise and interpretable emergence of inside-outside computation in Transformer-based language models

Jack T. Parley Fully trained Transformers are known to approximate parsing algorithms such as inside-outside, but how this computation is learned is poorly understood theoretically. The authors study training dynamics on synthetic context-free grammars with variable parse tree topology, where the data statistics are under full analytical control. They find that every stage of the stagewise loss curve can be computed analytically and tied to the representation of a hidden latent tree variable. Probing shows the model first uses local bigram and trigram information in a few early layers, then gradually develops an encoder-decoder organization with the most abstract latent information in the middle layers and later layers refining the token prediction.

arXiv:2611.14221 PDF

An Energy-Based Perspective on Machine Learning

Bo Dai, Yilun Du Energy-based models (EBMs), which define a distribution through a scalar energy function, are used in this tutorial as a single lens on machine learning. It develops three defining choices (the energy, the inference procedure and the learning method), covering sampling, variational inference, maximum likelihood, noise contrastive estimation, score matching and adversarial training, and shows these learning methods share one contrastive structure. It then argues that discriminative models, GANs, normalizing flows, diffusion and autoregressive models, contrastive and masked representation learning, reinforcement learning, planning and verifier-based inference-time computation are all EBMs differing in those three choices, mostly in how they handle the partition function and sampling. The framing is also used to suggest new algorithms by swapping one component and to compose energies for combining learned models at inference time.

arXiv:2611.14383 PDF

A Margin Perspective on LoRA: Robustness to Catastrophic Forgetting and Adapter Merging (MaLoRA)

Ziqing Xu, Hancheng Min, Lachlan Ewen MacDonald, Salma Tarmoun, Enrique Mallada, Weijie Su et al. Low-rank adaptation (LoRA) is observed to resist catastrophic forgetting and to allow independently trained adapters to be merged, but neither property is well understood. The authors analyze LoRA in multiclass linear classification with near-orthogonal tasks and L2 regularization, showing that in an intermediate regularization regime the optimal adapter aligns with the max-margin solution on the fine-tuning data. This yields closed-form margins on pre-training and fine-tuning data, with the regularization parameter controlling the balance between retention and adaptation, and a proof that merged adapters keep a positive margin on each task with derivable optimal mixing coefficients. The resulting training-free merging rule, MaLoRA, matches or outperforms several merging baselines on vision and language classification tasks.

arXiv:2611.14397 PDF

A Theory of Cross-Validation Variance

Nathaniel S. O'Connell, Jaime L. Speiser cross-listed Each observation influences a cross-validation estimate in two roles, through its own evaluation when held out and through the models it helps fit, with the random splits setting its weight in each role. From this representation the authors prove an exact finite-sample identity for the variance of the cross-validated estimate, covering metrics averaged over observations such as mean squared error and over case-control pairs such as the AUC. The variance splits into sampling variance, depending only on training size, and variance from random fold splitting, so K-fold cross-validation, whose weights are equal across observations by design, is variance optimal at first order while Monte Carlo and bootstrap cross-validation are not. First-order K-fold variance does not depend on K, leave-one-out has the smallest variance among K-fold schemes under reasonable model conditions, and the estimate is shown to be asymptotically normal with the identity's full variance.

arXiv:2611.14463 PDF

26 more specialized papers

Vision 36

Trajectory-Regularized Distribution Distillation for One-Step Generation

An Zhao, Yixiang Zhou, Zejian Li, Yi Guo, Pei Chen, Lingyun Sun cross-listed Distilling flow matching models into one-step generators is usually done either by trajectory alignment, which degrades at very small sampling budgets, or by distribution matching, which typically needs a second network to track the student's evolving score. Trajectory-regularized distribution distillation (TRD) combines both forms of supervision through a momentum-mixture policy predicted in a single student forward pass, which serves at once as the one-step generator, as an estimator of the student's marginal velocity, and as a closed-form velocity field supervised along the teacher's trajectory. On SANA-1.6B and Stable Diffusion 3.5, it outperforms representative consistency and distribution matching baselines in quality and training efficiency for one-step and few-step generation, at some cost in sample diversity.

arXiv:2611.12978 PDF

DUO: Dual Readouts of Diffusion Models for OOD Detection

Alireza Javid, Shirin Shoushtari, Nuria Gonz\'alez Prelcic cross-listed Out-of-distribution (OOD) detection with diffusion models usually relies on reconstruction error, which fails when OOD inputs are reconstructed as well as or better than in-distribution ones. DUO adds a second signal, the leading sensitivity spectrum of intermediate denoising features, and combines it with reconstruction error in a fully unsupervised, training-free detector. It improves average AUROC by about 6 percentage points over the strongest baselines while substantially reducing inference cost, and reports state-of-the-art results on near-OOD and high-resolution benchmarks.

arXiv:2611.13304 PDF

When Are Predictions Enough? An Evaluation Protocol for Frozen Expert Composition

Alessandro Pereira, Keith J Ransom, Lewis Mitchell Independently trained frozen models are usually combined through their final predictions, even though their internal features may hold information those predictions discard. The authors define a reusable evaluation protocol with strong prediction baselines, simple and larger feature combiners, and train/test overlap controls, and apply it to 21 frozen deepfake detectors from DF40. Under leave-one-generator-out evaluation, logistic regression on the features of six architecture-diverse experts improves macro AUROC by 0.081 over the best prediction combiner, while a combiner with roughly 350 times more parameters shows no clear further gain. Weakening the prediction baselines on the overlap-affected standard split inflates the apparent feature advantage by up to 0.172 AUROC, so the same benchmark can support different conclusions depending on evaluation choices.

arXiv:2611.13356 PDF

PhyRubric: Rubric-Grounded Diagnostic Evaluation of Physical Commonsense in Text-to-Video Generation

Huiyi Chen, Weitai Kang, Shicheng Fan, Xinshuo Wang, Zhixiang Shen, Yan Yan et al. cross-listed Existing benchmarks for physical commonsense in text-to-video (T2V) generation rely on a single intuitive score per video, which is subjective and does not say which physical principle was violated. PhyRubric builds a caption-specific rubric before any video is seen, decomposing the caption into semantic and physical items and scoring from item-level verdicts grounded in timestamped evidence. It is instantiated as PhyRubric-Bench, covering 34 categories across five physical domains, and PhyRubric-Train, with 3,788 videos used to fine-tune the open-source PhyRubric-Judge. Across ten T2V models physical commonsense remains far from solved, and PhyRubric-Judge correlates with human judgments more closely than existing open-source judges while matching Gemini 4 Argon.

arXiv:2611.13576 PDF

Earth Embeddings in Environmental Remote Sensing

Yuchi Ma, Jiaqi Yang, Mengyu Liang, Ivan Higuera-Mendieta, Caleb Robinson, Anthony Ortiz et al. Earth embeddings are general-purpose representations derived from multi-source satellite observations, intended to spare practitioners from data gaps, heavy preprocessing and task-specific feature design. This review covers their theoretical foundations, representative embedding products and applications across six areas of environmental remote sensing, then benchmarks them on hyperspectral image classification, individual tree crown segmentation, crop yield prediction and aboveground biomass mapping. The case studies indicate that current Earth embeddings do not always outperform raw remote sensing imagery or hand-crafted features, so the authors recommend evaluating them per task rather than treating them as a replacement for conventional data.

arXiv:2611.13612 PDF

CASCADE: Constraint-Aware SCAffolded DEcomposition for Logical Anomaly Detection

Jinchan Kim, Jinhee Im, Hyunsouk Cho cross-listed Logical Anomaly Detection (LAD) checks whether an image satisfies constraints such as object counts, pairings, and measurement ranges, and recent methods have Large Vision-Language Models (LVLMs) verify a textual checklist against the same unmodified image. The authors identify a constraint-specific perception gap in this setup and find that extra textual reasoning over the same image helps little, whereas adapting the visual input to each constraint substantially improves accuracy. CASCADE is a training-free framework that adds constraint-specific visual scaffolds to checklist decomposition, and CASCADE++ configures those scaffolds per object from the constraint structure and reference distribution. Across MVTec-LOCO and VID-AD, CASCADE++ surpasses the full-shot state of the art by 3.9 AUROC points using only three reference images.

arXiv:2611.13820 PDF

Satellite Pretraining Is Not One Thing: A Matched-Pair Study of Generic and Earth-Observation Foundation Models for Scene Classification

Sandarbh Agrawal, Rakshith Sathish, Ashwathi Nambiar cross-listed Whether encoders pretrained on satellite imagery actually beat strong generic encoders for few-label land-use and land-cover classification is tested by comparing five matched pairs of Earth-observation (EO) and generic models that share objective and architecture, within a benchmark of 14 frozen encoders on five RGB scene-classification datasets. Using k-NN, linear and attentive probes plus image retrieval, the study finds that EO pretraining significantly helps three pairs (SatDINO, Scale-MAE, RemoteCLIP) and significantly hurts two (DINOv3-SAT by up to 22.5 linear-probe points, SatMAE by up to 8.8), with opposite outcomes even for two masked autoencoders pretrained on the same fMoW-RGB imagery. The effects concentrate on small man-made objects and are largest with 1-5% of labels, and the silhouette score of frozen embeddings predicts k-NN accuracy and retrieval quality almost perfectly (Spearman correlation of at least 0.96 on four of five datasets). The authors conclude that satellite pretraining is not a reliable selection criterion and that candidate checkpoints must be validated on target data.

arXiv:2611.13849 PDF

RB-LDC: Redundancy-Balanced Latent Coding for Noise-Robust Diffusion Latents

Hyunseok Jeong, Jaeho Jeon, Young-Sik Kim Forward diffusion corrupts every whitened latent direction at the same signal-to-noise ratio (SNR) even though the decoder is far more sensitive to errors in some directions than others. The authors analyze an invertible linear layer between encoder and diffusion process that shifts each direction's log-SNR, and derive a unique optimal allocation by minimizing decoder-sensitivity-weighted linear minimum mean-square-error (LMMSE) recovery risk under an energy budget; at large budgets it follows a power law in sensitivity. Instantiated as RB-LDC with Jacobian-based sensitivity estimates on ImageNet 256 x 256 with SD-VAE and VA-VAE, it beats budget-matched isotropic coding on all four metrics and reduces average perturbed FID by 17.1 for VA-VAE block-local coding while preserving clean reconstruction. The authors also report intermediate SNRs where perturbed FID worsens.

arXiv:2611.13955 PDF

Pulse or Provenance? What rPPG-Based Deepfake Detectors Learn

Othmane Harraq, Tamer Aldwairi cross-listed Remote photoplethysmography (rPPG) is used for deepfake detection on the premise that synthesized faces lack a real cardiac signal, and the authors test what such detectors actually learn on Celeb-DF++ and FaceForensics++. Simple chrominance traces reach an AUC of 0.998 on talking-face forgeries, but the signal is not physiological: it is equally discriminative from background regions, sits above the heart-rate frequency band, and is unaffected by injecting a synthetic pulse. Detectors trained on Celeb-DF++ classify over 99% of real FaceForensics++ videos as fake, indicating that they encode video source instead of manipulation, and a CLIP-based detector behaves the same way. The official Celeb-DF and Celeb-DF++ splits are also shown to share identities and source videos between training and test, inflating weaker detectors by up to 0.30 AUC.

arXiv:2611.13992 PDF

Forward-Value Discretization: Faster Stochastic Diffusion Sampling Without High-Order Approximation

Yuchen Jiao, Na Li, Changxiao Cai, Yuting Wei, Gen Li Stochastic diffusion samplers usually need many more neural function evaluations (NFEs) than deterministic ones for comparable quality. The authors prove matching upper and worst-case lower bounds showing that stochastic samplers with higher-order approximations remain first-order convergent even with exact score functions. They instead propose forward-value discretization, a training-free approach that evaluates the prediction model at a lookahead estimate built from previous predictions, yielding F-DPMSolver for deterministic sampling and F-DDPM for stochastic sampling at one model evaluation per step. On CIFAR-10, ImageNet, FFHQ, and LSUN, F-DDPM achieves lower FID than the stochastic baselines at matched NFEs and is competitive with deterministic baselines when given 1.5 times as many NFEs.

arXiv:2611.14328 PDF

UGG: Unified Generative Geometry Model

Amani Kiruga, Zhiyi Li, Ruojin Cai, Hansen Lillemark, Xiaoshan Wu, Ayush Tewari et al. cross-listed Work on 4D scene understanding is split between feedforward models that predict geometry directly and generative models that synthesize video. UGG jointly models RGB video with depth and camera pose in one generative model, assigning different noise levels across modalities and time so that both visual generation and geometric prediction become conditional completion within a single model. This formulation lets it train on heterogeneous datasets carrying different subsets of annotations, and lets it search at inference time over multiple scene hypotheses using geometric evidence extracted from observed frames. The authors report gains over non-unified baselines on camera-controlled video generation, depth estimation, and camera pose prediction, including under sparse observations.

arXiv:2611.14440 PDF

Medical Image Segmentation: Foundation Model Adaptation versus Task-Specific Training

Emre Kutlusan, Yongmin Li cross-listed Comparisons between adapted foundation models and task-specific networks for medical image segmentation usually test on data from the training distribution against a single weakly justified baseline, which says little about performance at new clinical sites. Here DINOv2 is adapted with strategies ranging from bias-only tuning to full fine-tuning and compared with U-Net and ResU-Net models trained from scratch on colonoscopy, breast ultrasound, and dermoscopy, with testing on both held-out in-source data and independent datasets from other institutions. The adapted foundation model generalised better on every external dataset, while task-specific models stayed competitive in-distribution when data were abundant, drew boundaries more precisely, and missed fewer small lesions. Only a small fraction of the foundation model needed adapting, with the required capacity set by imaging domain rather than adaptation method, and a widely used polyp test set was found to duplicate images from another benchmark.

arXiv:2611.14519 PDF

24 more specialized papers

Multimodal 29

Rethinking in Spikes: Mitigating Hallucinations in MDLMs with Step-Aware Decoding

Zhongxing Xu, Zhonghua Wang, Zhe Qian, Shiyan Su, Ming Hu, Xiaocheng Zou et al. In multimodal diffusion language models (MDLMs), mask uncertainty generally falls during iterative denoising, but a few steps show localized entropy spikes that are closely associated with hallucinations. StepRefine is a plug-and-play decoding strategy that detects such steps by comparing each step's entropy to its neighboring temporal window, then switches from discrete token commitment to several rounds of continuous latent refinement using probability-weighted embeddings. A refinement controller monitors distributional stability and returns the model to standard discrete denoising, and experiments report reduced hallucination across different MDLMs on multimodal benchmarks.

arXiv:2611.12901 PDF

TimeOmni-o1: Bidirectionally Coupling Time Series Foundation Models and LLMs for Multimodal Reasoning and Forecasting

Ziyang Zhang, Shenyi Li, Wen Wu, Chao Zhang Understanding and forecasting time series that come with related text require reasoning over both modalities, but large language models (LLMs) handle long numeric series poorly and time series foundation models (TSFMs) neither read nor generate text. TimeOmni-o1 couples the two bidirectionally: the TSFM encodes the series in compressed form for the LLM to reason over alongside the text, and the LLM's reasoning is fed back into the TSFM to condition the forecast. Trained on MMTR, a new reasoning-annotated dataset with 500k+ samples across 12 domains, it outperforms existing models and its own LLM and TSFM components used individually on both tasks across most domains, with gains carrying over to three open benchmarks.

arXiv:2611.12902 PDF

Graph Visualizations in Prompts for Knowledge Graph Question Answering

Luiz do Valle Miranda, Andreas L. Opdahl, Grzegorz J. Nalepa A controlled study examines when adding a rendered graph image to the prompt helps knowledge graph question answering (KGQA) in vision-language models. Over 1,500 WebQSP and CWQ questions and six models in a fixed pipeline, text-plus-image prompting yields an average gain of 1.85 F1 points, driven by precision improvements on low-complexity questions (+4.65 F1). The gains do not depend strongly on model size but are highly sensitive to layout: image-only input sharply degrades performance, hierarchical Graphviz layouts beat less organized renderings, and splitting large subgraphs into separated paths helps most. The authors recommend adaptively attaching diagrams only for small, sparse subgraphs.

arXiv:2611.12938 PDF

Multimodal Graph Learning: A Survey of Models, Benchmarks, and Frontiers

Rui Xue, Tianfu Wu Multimodal attributed graphs (MAGs), where nodes carry text, images, and other modalities and edges encode relations, underlie e-commerce, social media, and content platforms, and methods for learning on them have split into graph-centric and foundation-model-centric lines with inconsistent benchmarks. This survey introduces a unified message-passing formulation that organizes methods into five integration levels: feature, propagation, topology, learning, and foundation model, deriving representative methods as special cases and comparing their assumptions, expressiveness, and cost. It also reviews recent benchmarks using model-independent dataset descriptors such as label homophily and average degree, and maps each evaluation protocol to its baselines. A frontiers section connects the field to AI agents, including the graphs agents perceive and remember and the practice of organizing agent workflows as explicit graphs.

arXiv:2611.12990 PDF

Seeing Is Not Believing: Counterfactual Lesion Erasure Teaches a Small Medical Vision-Language Agent to Report Only What the Image Shows

Bhaskar Gurram Medical vision-language models often report a finding simply because the question names it, so the authors build counterfactual views from public segmentation datasets, including images with the lesion erased by inpainting and a same-shape control edit elsewhere, and define the evidence sensitivity gap (ESG) as the rate of absent-or-refer answers after erasure minus the rate after the control edit. MedScope-4B, a Qwen3.5-4B model fine-tuned with LoRA and image tools on these views, reached an ESG of 0.758 on 507 internal cases and claimed an erased lesion in 9.6% of cases versus 41-62% for frontier models, without significant loss of accuracy on real lesions. The limits are substantial: at external centres frontier models detected more real lesions, and on a fresh external test the model detected 0 of 48 pneumothorax films. The authors conclude that counterfactual training gives a small model an evidence-gated disposition that prompting does not give frontier models, but does not improve perception.

arXiv:2611.13006 PDF

VIGOR: Benchmarking Visual Rationale Correctness in Multimodal Large Language Models

Yanlin Chen, Tang Li, Xi Peng cross-listed Multimodal large language models (MLLMs) are usually scored on answer correctness, which does not show whether an answer rests on the right visual evidence. VIGOR is a benchmark of visually grounded rationales, represented as explicit concepts paired with pixel-level groundings across abstraction levels from low-level visual properties to objects, parts, motion, and relations, and built with an automatic annotation pipeline that reuses existing dense supervision and verifies newly generated groundings conservatively. Comparing the visual evidence a model uses for each concept it mentions against the annotations, the authors find that current MLLMs show limited visual rationale correctness, with distinct grounding failures at different concept levels.

arXiv:2611.13150 PDF

Before Words, Beyond Speech: Evaluating Nonverbal Social Reasoning in Early Childhood

Marie Amale Huynh, Laura Bravo-S\'anchez, Lauren Klein, Nick Haber, Philip Fisher, Serena Yeung-Levy cross-listed Video understanding benchmarks focus on adult, conversation-driven interaction and overlook the gaze, gesture, and physical contact through which young children communicate. NEST (Naturalistic Early-childhood Social Interactions) is a child-centric benchmark of manually annotated clips of about 12 seconds, with a compositional schema that builds atomic cues into contextual, behavioral, interpersonal, and open-ended levels of social reasoning. State-of-the-art vision-language models show strong contextual recognition but a sharp drop on precise behavioral and interpersonal cues, which the authors trace to biases including poor consequence grounding and temporal sensitivity. They also evaluate domain adaptation strategies on the benchmark.

arXiv:2611.13213 PDF

Disentangling Multimodal Rotary Positional Embedding in Vision-Language Models

Zhaoyi Liu, Haichuan Zhang, Xiaochuan Gong, Yancheng Zhu, Bowei Tian, Wanghao Ye et al. Multimodal Rotary Position Embedding (MRoPE) gives visual tokens temporal, height, and width coordinates, but also assigns height and width to text tokens, which introduces spatial positional effects into text-vision attention. The authors identify a constraint of shared coordinate layouts that prevents removing these cross-modal effects while keeping spatial encoding among visual tokens, and design controlled experiments to separate the two contributions. Across model scales, MRoPE variants, and benchmarks, cross-modal spatial effects add little as spatial references and can impair performance, while intra-visual spatial encoding remains useful for fine-grained grounding.

arXiv:2611.13334 PDF

How Hyperbolic Are Hyperbolic Vision-Language Models?

Freek Byrman, Partha Das, Stijn Verdenius, Pascal Mettes cross-listed Hyperbolic vision-language models (VLMs) report strong zero-shot results and an embedding space in which radius orders representations by generality, but it is unclear how much the hyperbolic geometry is actually used. Examining the exponential map through which curvature enters these otherwise Euclidean networks, the authors find that in MERU, HyCoCLIP, ARGENT, and PHyCLIP it is well approximated by the identity function, and removing it at inference leaves evaluation scores unchanged. Euclidean models trained from scratch on CC3M with the same hierarchy losses match the hyperbolic versions in training statistics, evaluation scores, and radial distributions. The authors conclude that Euclidean space is sufficient for hierarchical vision-language learning as currently practiced.

arXiv:2611.13384 PDF

Towards Reliable VLM Judges: State-Conditional Invariance and Presentation-Aware Diagnostics

Qinan Zhang, Xinyu Wang, Qihang Jin, Tianyuan Wang, Hongbing Luo, Pengfei Li et al. Vision-language models (VLMs) are increasingly used as automated judges, but aggregate accuracy does not show whether a judge changes its verdict for the right reason. State-Conditional Invariance (SCI) is a reliability principle requiring a judge to stay stable under presentation changes that preserve meaning while responding to causal interventions that change the task outcome. It is instantiated in GridWM-Judge, a diagnostic benchmark built from deterministic MiniGrid trajectories with aligned full, evidence-removal, and counterfactual variants across five diagnostic tasks. Evaluations reveal heterogeneous failure modes, including sensitivity to how evidence is presented and weak paired counterfactual discrimination, with visual grounding and objective recovery as substantial, model-dependent bottlenecks.

arXiv:2611.13596 PDF

OpticalRAG: Pixel-Space Compression for Token-Efficient Retrieval-Augmented Generation

Senhao Liu, Yuheng Zhang, Chunyu Wei, Yueguo Chen, Xinran Zhang Retrieval-augmented generation (RAG) is limited by context token budgets, and text-based compression discards fine-grained information. OpticalRAG instead renders documents as page images, encodes them with a frozen vision encoder, retrieves visual tokens from a text query through a shared projection that needs no contrastive training, and distills them with a lightweight cross-attention compressor into query-conditioned prefix tokens. On LongBench-v2, BAMBOO and LooGLE-v2 it achieves the best compression-aware accuracy on all three benchmarks while injecting only 16 tokens, and improves the closed-book Qwen2.5-7B-Instruct backbone by 5.17 points on LongBench-v2.

arXiv:2611.13675 PDF

AURA: Auditory Spatial Intelligence via a Unified Representation from a General-Audio Encoder through Complete Alignment

Boyan Han, Xiaoyan Wang, Zhe Sun, Ziqi Wang, Yujun Cai, Hongli Cao et al. cross-listed Spatial audio-language models usually rely on a dedicated spatial encoder, and this study asks whether a frozen general-audio codec trained without spatial supervision can supply both semantic and spatial information. Layer-wise analysis shows the codec retains spatial cues, but a model trained on them predicts left/right at chance in two-source scenes, which the authors attribute to an asymmetry where the audio contains both sources while the text answer describes only one. A 2×2 experiment shows that binding directions to the queried source requires both completing the scene description and exposing the inter-channel difference. The resulting AURA system keeps the codec and language model frozen and beats the strongest dedicated system on SpatialSoundQA by 10.31 points in dual-source direction accuracy with about 40% of its trainable parameters.

arXiv:2611.13687 PDF

Walrus Technical Report

Alkis Koudounas, Osamu Take, Hayato Futami, Quentin Jodelet, Emiru Tsunoo cross-listed Walrus is a unified 8B-parameter speech-and-audio model covering speech recognition (including code-switching and multi-talker speech), speech translation, text-to-speech, emotion recognition, audio captioning, and general audio understanding. Rather than a new architecture, it uses a standard decoder-only backbone initialized from Gemma 4 and a three-stage data curriculum with task-aware and dataset-aware weighting, trained on over 600M samples spanning more than 20 tasks and 100 languages. The authors report that it leads most multilingual, multi-talker, and code-switching recognition, translation, and audio understanding benchmarks among comparable open models, and matches or surpasses the 4× larger Qwen3-Omni on most benchmarks. They read the results as evidence that data engineering is a primary driver of quality in multi-task speech and audio modeling.

arXiv:2611.13689 PDF

Beyond Proxy Metrics: MLLMs as Human Surrogates for Explainable AI Evaluation

Eoin M. Kenny, Salim I. Amoukou, Tom Bewley Evaluating explainable AI (XAI) techniques usually relies either on proxy metrics that say little about an explanation's usefulness or on human studies that are expensive and hard to reproduce. The authors test whether multimodal large language models (MLLMs) can stand in for human participants, proposing a purpose-grounded evaluation framework with metrics applicable to both models and people and discussing when such surrogates can be trusted. Across five purpose-grounded areas and fourteen MLLMs, all models trend positively in alignment with humans, with OpenAI and Qwen models generally aligning best. The authors position the approach as a complement to user testing rather than a replacement.

arXiv:2611.13926 PDF

Counterfactual Analysis of Extent Leakage from Annotation-Derived Prompts in Vision-Language Medical Image Segmentation

Niran Nataraj, Ankita Das, C. Krishna Mohan, Kenji Kawashima cross-listed Text-guided medical segmentation benchmarks often pair images with captions whose size, count and location words were computed from the very mask used for scoring. Counterfactual caption edits on two published models show that on colonoscopy and chest X-ray the benefit comes mostly from lesion extent: keeping only the size word or lung-zone phrase recovers 84 to 98 percent of the improvement over an empty caption, and calling a polyp a rock changes nothing. The authors call this extent leakage, show that models learn the habit from training captions at the expense of image-only segmentation, and find that caption words alone predict lesion area. Without the test mask, the best tested method reaches 74.59 Dice on colonoscopy against the published 77.23, and an evaluation protocol separating image description from label-derived content is proposed.

arXiv:2611.13970 PDF

Where Do Vision-Language Models Get Left and Right? Position-Code Counterfactuals in Qwen2.5-VL, Qwen3-VL and InternVL3

Enhao Cui, Junzhi Yang, Yaoshun Lin, Yuhan Feng, Zijian Chen cross-listed Vision-language models could derive left versus right from image content within patches, from the position code of the vision transformer (ViT), or from token positions and order in the language model. Mirroring only the ViT width positions in Qwen2.5-VL-7B, Qwen3-VL-4B, and InternVL3-2B reproduces the flipped-image answer on 84.3 to 93.2% of CV-Bench left-right questions, whereas mirroring language-model-side ids or token order changes few answers; in Qwen2.5-VL-7B the effect sits in the global-attention ViT blocks, mostly the last. Post hoc analyses qualify the result: a half-grid shift and random permutations also flip many answers, and on What'sUp the changes caused by random permutations land on the mirror relation mostly below chance in the two Qwen series. Most remaining errors give the same relation whichever object is asked about, pointing to relation binding, object localization, or use of the reference object as targets for repair.

arXiv:2611.14060 PDF

ZoomGain: Mining Global-to-Zoom Perception Gaps for Fine-Grained Visual Supervision

Haozhan Shen, Liangbo He, Jun Lan, Huijia Zhu, Weiqiang Wang, Tiancheng Zhao et al. cross-listed Vision-Language Models (VLMs) often miss fine-grained details in a full image yet recognize them once the region is zoomed in, and ZoomGain exploits this gap to build supervision automatically. The pipeline writes questions about the full image using information gained only from the zoomed view, verifies answers with multiple models, filters out questions that are ambiguous or answerable from language priors, and generates distractors for multiple-choice items. The resulting ZoomGainBench, a 1,000-question bilingual benchmark, is hard for current models, with the best of 28 VLMs scoring only 61.08. Training on just 9K ZoomGain samples lets the authors' 4B and 9B models outperform several much larger models on perception benchmarks, including 90.25 on HR-Bench 4K for the 9B model.

arXiv:2611.14076 PDF

C$^2$-Routing: Complementary Cropping and Compression for Efficient Vision-Language Inference

Guorui Lu, Qinyu Chen cross-listed Visual Question Answering (VQA) is computationally expensive for vision-language models (VLMs), and simply capping visual tokens or applying one fixed input-reduction strategy can discard information the answer needs. C2-Routing builds on the observation that cropping suits small targets and local detail while whole-image compression preserves context for large or distributed evidence, selecting between them per image and question, with a safety gate that falls back to the original input based on the predicted joint correctness distribution. The method is model-agnostic and applicable in black-box settings, and all training and evaluation ran on RTX 3090 GPUs. Across three VLMs and 18 datasets, against full-image baselines already capped at 1,024 visual tokens, it further cuts end-to-end FLOPs by 36.53% and visual tokens by 42.36% on average, with a mean relative task-score decrease of 0.77%.

arXiv:2611.14469 PDF

11 more specialized papers

Reinforcement Learning 27

Topology-Aware Reinforcement Learning for CAD Decomposition toward All-Hexahedral Meshing

Jiaming Peng, Zhihao Zheng, Rui Wang, Qinglin Wang, Jing Xiao, Jie Liu et al. cross-listed All-hexahedral meshing of complex CAD models requires first decomposing them into simpler meshable blocks, and existing learning-based methods restrict themselves to surface-based splits that often yield invalid decompositions. DecompRL is a deep reinforcement learning framework with topology-aware internal-splitting actions and rewards that combine geometric terms with meshability feedback. Across ABC, Fusion360, NIST and DeepCAD, it achieves the highest success rate among the evaluated learning-based methods on three of the four benchmarks, reaching 59.6% on Fusion360, and its decomposition step runs up to two orders of magnitude faster than the full LMFF procedure.

arXiv:2611.12930 PDF

Exploration via Exploitation: The Blessing of Reward Diversity in Personalized Federated RL

Jiin Woo, Gauri Joshi, Yuejie Chi Federated reinforcement learning usually treats agents with different reward functions as an obstacle to collaboration; this theoretical work treats that heterogeneity as a source of exploration instead. PF-UCBVI (Personalized Federated Upper Confidence Bound Value Iteration) achieves optimal linear speedup by separating shared dynamics from personalized objectives. A second, purely greedy algorithm, PF-EFVI (Personalized Federated Exploration-Free Value Iteration), relies on reward diversity to cover the state-action space and is proven to attain logarithmic regret without explicit exploration bonuses when rewards are sufficiently diverse.

arXiv:2611.12964 PDF

Branch-and-Commit for Frozen RL Agents: When and Where to Spend Test-Time Compute

Sahar Admoni, Assaf Hallak, Yftah Ziser, Ofra Amir When a reinforcement learning policy cannot be retrained but its environment state can be copied and resumed, an operator can sample several continuations, evaluate them, and continue from the chosen one, a procedure the authors call branch-and-commit. Experiments across four environments, two frozen policies, and five continuation verifiers show the benefit depends on dispersion, the gap between the best and mean candidate return: branching does nothing in low-dispersion settings but raises return by 12.8% in the main Craftax-Medium setting. Most of that gain comes from selecting among different stochastic realizations of the same action instead of choosing better actions, and verifier quality is not the main constraint. Gating branching on predicted dispersion keeps 69% of the improvement with 48% as many branching calls on held-out seeds.

arXiv:2611.13003 PDF

FaST: Factored Synthesis of Transitions for Offline Reinforcement Learning

Daniel Campa, Spandan Das, Srinjoy Das Diffusion-based data generators for offline reinforcement learning model each transition (state, action, next state, reward, terminal) as a single block, so no one factor can be redrawn on its own. FaST (Factored Synthesis of Transitions) splits the transition by the chain rule into state, action, dynamics, reward, and terminal stages, each fit with an estimator suited to it. On twelve D4RL locomotion and maze benchmarks, TD3+BC and IQL agents trained on its data match the SynthER baseline while generating transitions over 90x faster. Drawing ten candidate actions per generated state and keeping the highest-advantage one beats training on real data in 11 of 12 medium and medium-replay settings, by 8.9 D4RL points on average, though the approach degrades on expert mixtures because of the advantage estimate.

arXiv:2611.13013 PDF

GTAGym: Real AAA Game Tasks Environment for Evaluation and Reinforcement Learning

Jiajun Chen, Shicheng Fang, Jiahao Lu, Yuxin Wang, Xipeng Qiu cross-listed Checking whether an agent completed a task in a AAA game usually needs human annotation, which blocks both scalable evaluation and reinforcement learning (RL). GTAGym is an environment for GTA V that exposes screenshots and keyboard-mouse actions through a web server with automatic engine-state outcome checks; its benchmark GTAVBench has 103 online tasks across nine families drawn from main-story missions plus 47 image-based shooting tasks. Among six frontier models, the best configuration reaches 70.9% overall accuracy on the GTAVBench-AGI variant, where instructions appear only in the first screenshot, while aerial navigation, evasion, and interception remain difficult. A task generator supplies automatic rewards for RL, and LoRA RL of Qwen3.8-27B on generated on-foot navigation tasks transfers to other categories and raises native GTA V task success from 15% to 23%.

arXiv:2611.13028 PDF

Building World Models from Intrinsic Motivation

Reece Keller, Declan Kutscher, Xaq Pitkow, Aran Nayebi How different forms of intrinsic motivation interact with world-model representation learning to support downstream control has not been compared systematically. The authors place 24 intrinsic objectives spanning curiosity, entropy, and competency in one model-based reinforcement learning framework and evaluate the resulting world models on zero-shot, few-shot, and multi-task control in locomotion, manipulation, and open-ended game environments under three representation-learning strategies: observation reconstruction, latent prediction, and redundancy reduction. Curiosity-driven exploration combined with observation reconstruction gives the strongest aggregate results, though the best individual pairing varies by experiment. On compositional vision-based tasks, motivation that combines episodic and lifelong novelty achieves the best zero-shot transfer and few-shot adaptation.

arXiv:2611.13113 PDF

Making Strides in Symbolic Imitation Learning: Offline Demonstrations, Online Tuning

Bryce MacInnis, Malcolm I. Heywood Deep reinforcement learning dominates continuous control but yields opaque policies, while interpretable alternatives have lagged in performance. Symbolic Imitation Learning distills an oracle policy from an offline dataset into a decision list whose leaves are linear-model motion primitives, then tunes the constants and primitives online with Augmented Random Search (ARS). Across 20 benchmark reinforcement learning tasks the resulting policies approximate the oracle and exceed it on 11 of the 20 tasks, while staying compact enough for domain experts to read.

arXiv:2611.13182 PDF

Audited Gradients, Anchored Policies: Model-Based Reinforcement Learning for Efficient Flow Control

Ruige Kong, Stefania Fresca Designing feedback controllers for fluid flows is costly because each candidate needs a high-fidelity solver, and differentiable surrogates can track the flow accurately while getting the derivative of the objective with respect to the action wrong. The method audits the surrogate's action gradient against finite-difference estimates from the solver on held-out states, fits a low-dimensional feedback law to whichever gradient source passes, and uses that law to initialize and anchor a neural policy via a quadratic penalty. Across four test cases the controllers match or beat a model-free baseline using 3 to 156 times fewer high-fidelity solver interactions. One surrogate predicted the objective at correlation 0.9996 but its action gradient at only 0.459, showing that prediction accuracy does not identify a usable surrogate, and audit passes were found to hold only locally.

arXiv:2611.13203 PDF

Wrong the Same Way: Cross-Prompt Sharing Both Suppresses and Amplifies Reward Hacking in RLVR

Michael Li, Rohan Pandey, Ray Chen Two competing accounts of reward hacking in reinforcement learning with verifiable rewards (RLVR) — that it is inherited from the base model, or that RL amplifies it — are studied in an environment small enough to compute the full response distribution exactly. At the KL-regularized optimum the hack rate is inherited unchanged, but with shared parameters a wrong answer accepted on few prompts is pushed below the base rate while one accepted on most prompts is amplified, a pattern that holds under Adam, SGD, and sampled GRPO. The practical implication is that verifier fixes should target mistakes the model repeats across prompts: two patches removing equally many verifier errors can differ by up to 38 points of true accuracy.

arXiv:2611.13218 PDF

Off-policy Learning with Excursion Policies

Jiamin He, Mark Rowland, Zhaohan Daniel Guo, Hado van Hasselt, Diana Borsa Off-policy reinforcement learning evaluates a target policy from data generated by a different behavior policy, typically trading bias against variance or stability by linearly mixing the two. Excursion policies offer a structured alternative in which the agent follows the target policy for a multi-step excursion before reverting to the behavior policy. The authors derive dynamic programming results and provably sound off-policy temporal-difference algorithms for them, and find on continuous control benchmarks that excursion policies give an efficiency-stability trade-off comparable to mixture policies, while behaving differently in concept.

arXiv:2611.13275 PDF

Train High, Ship Low: Efficient LoRA Synchronization for Distributed On-Policy Post-Training

Shenghui Li, Li Ju, Thiemo Voigt Sending LoRA adapters to rollout workers can bottleneck distributed on-policy post-training when bandwidth is limited. HiLo instead transmits compressed gradients and has each worker replay the optimizer locally from replicated optimizer state, so learner and workers stay at the same policy version under reproducible execution. Its companion compressor Spar combines spectral truncation, adaptive rank allocation, and error feedback to decouple training rank from communication cost. Across on-policy distillation and reward-guided post-training, learning quality stays close to full synchronization while transmitting as little as 1.6% of the adapter bytes, with end-to-end updates up to 2.2x faster under bandwidth constraints.

arXiv:2611.13306 PDF

SLIC: Reinforcement Fine-Tuning Small LMs for Multi-Turn Analog Circuit Optimization

Yan Xu, Zhengqi Gao, Ching-Yun Ko, Zhang-Wei Hong, Wenjie Lu, Tao Yu et al. Analog integrated circuit sizing is a multi-turn, simulator-in-the-loop task, and proprietary process design kits (PDKs) make hosted LLMs hard to use for it. SLIC applies reinforcement fine-tuning to a 4B-scale, locally deployable language model using structured circuit representations, discretized relative-update actions, gated multiobjective rewards and Reward-Guided Rollout Steering (RGRS), and comes with an open benchmark of 41 amplifier topologies. Trained only on open 65 nm predictive technology models, it reaches 87.1% in-distribution Pass@1 within eight turns, 48.5 percentage points above the 763B DeepSeek-V4.1-Flash Think. It reaches 72.6% on held-out topologies and transfers without further tuning to five unseen technology nodes and a commercial TSMC 65 nm PDK.

arXiv:2611.13452 PDF

Are Dense Rewards Worth Their Tokens in GRPO?

Vikash Singh, Weicong Chen, Debargha Ganguly, Yanyan Zhang, Mohsen Hariri, Sreehari Sankar et al. Dense process rewards have been proposed to improve on group-relative policy optimization (GRPO), which gives every token of a solution the same credit, but comparisons per training update ignore their cost: in this setting the Monte-Carlo continuations alone consume 91% of generated tokens. A matched-compute protocol gives every intervention the same budget in generated tokens and training GPU-hours, tested in preregistered studies across 1,668 runs on six models. On Qwen2.5-1.5B the Monte-Carlo reward leads GRPO per update but ends 14.3 percentage points lower at equal tokens and 7.8 points lower at equal training time, and none of four dense-reward variants beats GRPO. Training on prompts whose predicted success is nearest 1/2 does improve accuracy at matched compute, by 1.7 points on Qwen3-0.6B and 0.8 on Qwen2.5-1.5B, although GRPO with twice the initial learning rate comes within half a point of those gains.

arXiv:2611.13560 PDF

Broad-Persistent Advising in Interactive Reinforcement Learning via Latent State Encoding

Zhuoran Liang, Francisco Cruz Deep Interactive Reinforcement Learning (Deep IntRL) speeds up training with external advice, but typically uses each piece of advice for one step and discards it; Broad-Persistent Advising (BPA) stores advice and reuses it in similar states found by K-means clustering. Encoder BPA replaces the clustering with a trained encoder that maps state sequences to latent representations, retrieving advice by latent comparison with top-k aggregation and a consistency condition for stability. Experiments show it improves early-training learning efficiency and advice utilisation over other Deep IntRL approaches, especially in complex and partially observable environments.

arXiv:2611.13712 PDF

TRUER: Test-Time Reward Correction Using an Error-Aware Rectifier for Reinforcement Learning

Mingxiang Shen, Jinhao Li, Jiaming Xu, Yibo Han, Guohao Dai, Yijia Zhang et al. Test-time reinforcement learning (TTRL) adapts language models using rewards derived from majority voting over their own answers, which can reinforce wrong answers; an audit shows that false positives, incorrect responses receiving positive reward, dominate the reward errors. TRUER uses a small labeled source set to train a reward rectifier that separates correctness from consensus, with emphasis on incorrect majorities and correct minorities. The rectifier is then frozen and used to correct rewards during test-time adaptation, and the authors report higher reward accuracy and model performance than naive TTRL in every evaluated configuration.

arXiv:2611.13714 PDF

Integrating Action Prefix Prediction and Alignment Across Modalities for Multimodal Reinforcement Learning

Sijie Xu, Francisco Cruz Representation learning from multimodal observations is hard in reinforcement learning when interaction data is limited, and the existing Multimodal Joint Predictive Representation method relies on recursive rollouts that compound earlier prediction errors. Improved Predictive Representation (IPR) instead predicts each future representation directly from the observed anchor and the corresponding action prefix, and aligns modalities at that anchor with contrastive multimodal learning. Across five DeepMind Control Suite tasks the IPR-H configuration achieves an average return 10.0% higher than the baseline with visual and depth images and 3.4% higher with visual images and position vectors. In Webots both configurations beat the baseline without LiDAR but not when LiDAR is available.

arXiv:2611.13782 PDF

Temporal Behavior Trees for Reinforcement Learning: Specifications as Rewards, Curricula, and Diagnostics

Vladimir Krsmanovic, Bernd Finkbeiner Reinforcement learning from formal specifications usually splits task definition, reward shaping, curriculum design, and policy diagnostics across separate mathematical objects, losing structure such as subtask boundaries and ordering constraints. TBTCL (Temporal Behavior Tree Centric Learning) uses a single Temporal Behavior Tree (TBT) for all four roles, with algorithms that extract a curriculum graph and dense robustness-based rewards directly from the tree's semantics, and trace segmentation that builds competence profiles and shifts reward toward bottlenecks. The authors argue the formalism is inherently hierarchical and more expressive than LTL, and across six specifications in discrete and continuous domains they report that one tree can serve as specification, reward, curriculum, and diagnostic at once.

arXiv:2611.13993 PDF

Margin-Guided Token Credit for Reinforcement Learning with Verifiable Rewards

Chi Liu, Xin Chen, Yuchuan Weng, Srinivasan Manoharan In reinforcement learning with verifiable rewards, GRPO spreads the reward for a correct final answer equally over every token of the solution. MCPO (Margin-Credit Policy Optimization) instead concentrates credit where the model hesitated, measured by the gap between the top two next-token probabilities, which is read from a forward pass the trainer already performs. Training MiniCPM5-2B-Base with LoRA for 200 steps on 2,048 DAPO-Math problems, it beats the GRPO baseline on all five seeds by 6.55 points of macro avg@32 on average across six competition-math benchmarks, and a Qwen3-4B-Base replication gains 4.75 points. The resulting models also write two to three times longer answers, and the experiments do not separate that effect from the credit scheme.

arXiv:2611.14015 PDF

What a CAGE 4 Score Measures: An Audit of a Learned Cyber-Defense Benchmark, Its Reference Defenders, and the Tables Built on Them

James Rice, Ji Hwan Park, Andres Kwasinski cross-listed An audit of CAGE Challenge 4, the benchmark used to compare learned multi-agent cyber defenders, reproduces its two reference defenders, KEEP and H-MARL Expert, in a pinned container. The simulator silently discards actions sent to an agent that is still busy while still charging for a discarded Restore, and the documented observation does not expose the busy flag that both defenders read in their own wrappers; switching that check off at evaluation costs KEEP 99.58 points per episode and H-MARL Expert 145.98. Evaluation seeds the environment but not the policy, so 88.9% of per-episode outcome variance comes from the policy's own sampling draw, and none of five published comparison tables states enough to make its numbers comparable. The authors release their instrument, patches, traces, checkpoints, and a five-item disclosure checklist.

arXiv:2611.14068 PDF

Opportunity-Cost Option-Critic: Adaptive Switching Costs for Robust Hierarchical Reinforcement Learning

Khaled Ammoura, Aya Kayal, Naseem Daher Deliberation costs in hierarchical reinforcement learning (HRL) keep options from collapsing into primitive actions, but fixed or duration-based penalties are only indirect proxies for an option's usefulness. The Opportunity-Cost Option-Critic (OC2) framework instead sets a state-dependent cost equal to the positive value gap between the active option and its best alternative, so the agent persists when its current option is clearly better and switches quickly otherwise. Because learning such costs directly collapses to a zero-cost solution, an auxiliary objective aligns the learned cost with the value-gap target. In the Pinball goal-shift task, OC2 reaches a 52% first-ten-episode success rate versus 18% for a fixed-cost baseline, with further evaluation on Four Rooms under stationary and shifted conditions.

arXiv:2611.14099 PDF

An Augmented Lagrangian Approach to Lifted Bellman Linear Programming

Jongchan Park, Hyukjun Yang, Narim Jeong, Donghwan Lee Lifted Bellman Linear Programming (LBLP) expresses the in-sample optimal values of offline reinforcement learning as a linear program over the dataset, avoiding bootstrapped targets, but its neural solver ALBUM uses hinge penalties with coefficients fixed in advance. The authors solve the same program with the method of multipliers, so the coefficients become multipliers learned by dual ascent with a single penalty parameter, and show that in the convex case the fixed points are the program's KKT pairs. The resulting critic works with Gaussian actors, flow actors, and best-of-N sampling. Using a single Q and V and no target network, AL-LBLP with best-of-N reaches 60.1% average success on the 50 state-based OGBench tasks, including 81.5% on humanoidmaze-giant, and does not collapse up to batch size 8192.

arXiv:2611.14183 PDF

Recurrent Modular Structure in Spiking Neural Network Reinforcement Learning

Vikram Ramavarapu, Himanshu Udupi, Ashish Gautam Spiking neural networks (SNNs) for reinforcement learning are usually feed-forward, unlike biological circuits, which are modular and recurrent. Community Recurrent SNNs (CoRe SNNs) partition a recurrent network into communities with fixed roles (a controller, one community per action, and latent state communities), so that the chosen action is the community that spiking activity reaches, with all learning local and three-factor. Three ways of obtaining modularity are compared: a new learning rule called Map-STDP, a stochastic block model initialization, and fully sealed communities. On CartPole, Map-STDP improves learning and matches full modularity on two of three neuron models, and replacing the feed-forward actor of a TD-STDP actor-critic with recurrent action communities improves results on CartPole and LunarLander and matches them on Acrobot.

arXiv:2611.14262 PDF

A Theory of Policy Optimization in Post-Training: Coverage Collapse and Reference Resets

Audrey Huang, Nan Jiang, Sham M. Kakade, Sadhika Malladi, Akshay Krishnamurthy, Dylan J. Foster Theoretical foundations of policy optimization for language model post-training are revisited to find algorithms whose sample and gradient costs are governed by how readily the base model already generates optimal responses, a property the authors call coverage-preserving. They show that REINFORCE, the basis for PPO and GRPO, can drive the probability of optimal responses to near zero even as reward improves, a coverage collapse that can require exponentially many steps to escape and that persists with exact gradients and KL regularization to the base model. Repeatedly resetting the KL reference to the most recent policy mitigates this, formalized as Approximate Policy Mirror Descent (APMD), a meta-algorithm that approximately solves a sequence of KL-regularized subproblems. For softmax linear policies over pretrained features, APMD yields the first global convergence guarantees that stylized PPO and GRPO variants with reference resets, and squared-loss algorithms implementing them implicitly, are coverage-preserving.

arXiv:2611.14467 PDF

AutoScaffold: AI-Guided Scaffolding for Reinforcement Learning of Language Models

Zijian Zhang, Rizhen Hu, Jiangweizhi Peng, Shiyang Li, Sujay Bhatt, Sihan Zeng et al. Reinforcement learning with verifiable rewards (RLVR) collapses each training attempt into a single outcome number, discarding most of the information in the student's experience. AutoScaffold has a frozen teacher model periodically inspect that experience, diagnose bottlenecks, and write training-time prompt scaffolds, both strategies shared across a task class and hints for individual hard tasks, which are adapted as the student improves and removed at evaluation. Applied unchanged to ALFWorld, search-based question answering, mathematical reasoning, and GPU kernel generation, students evaluated without scaffolds beat the corresponding RL baselines in all four domains, including gains of 7.2 points in ALFWorld success and 4.0 points in average exact match on question answering. On math, the student surpasses the reported accuracy of QuestA with roughly one-third as many training steps.

arXiv:2611.14515 PDF

3 more specialized papers

Safety & Alignment 24

Train on prefill, monitor at decode: Using sparse autoencoders for topic monitoring

Rishikesh Mallagundla, Pranav Patel Activation-based monitors that run during generation have mostly been limited to detecting harmful content with two or three classes. TopicScope trains causal classifiers on sparse autoencoder (SAE) features to assign one of seven topics to every token, so the label can change mid-response, using synthetic multi-topic conversations and three detector architectures. On Gemma-3 1B and 4B, monitors trained on prefill hidden states generalised to decoding and lost at most 5 macro-F1 points under inserted asides, removed topics, or reordered sections. Ablations found that classifiers matching this performance without the SAE were several times larger.

arXiv:2611.12935 PDF

AI Governance Should Prioritize Control and Knowledge Boundaries Over Limiting Intelligence

Hendrik Luuk A position paper proposes the competence circuit, which decomposes an AI agent's real-world impact into intelligence-mediated discovery, procedural know-how, and direct control, and uses it to formalize the conditions required for an intelligence-driven fast takeoff. Treating intelligence as hypothesis-selection efficiency, stochastic simulations on NK fitness landscapes and an analytic bound indicate that intelligence-driven speedup is capped by the ratio of competing intelligence levels regardless of task difficulty, while absolute timescales remain limited by the time needed for capability-critical tests. Simulations also suggest mandatory authorization delays disproportionately slow highly intelligent agents. The authors argue that restricting control and enforcing knowledge boundaries are stronger safety levers than limiting intelligence, and note that purely digital domains weaken these bottlenecks.

arXiv:2611.12941 PDF

Monitoring One Channel Is Not Communication Control: Causal Evidence for Adaptive Reallocation of Joint Message--Route Codes

Liu Hung Ming (PARRAWA AI, cyril.liu@gmail.com, equal contribution), Tian Xia (PAII, TianXia0209@gmail.com, equal contribution) The study tests whether monitoring a visible channel actually controls communication between learned agents, using a game where a sender and receiver learn a discrete protocol and a routing policy without communication supervision. In two independent 20-seed batches, partial monitoring raised the share of traffic on the unmonitored shadow route by .701 and .575 over no monitoring, while task accuracy stayed near .91 and the messages' contribution was preserved. Interventions show the route itself carries meaning: route alone retains .388 accuracy against .25 chance. An explicit cost on sending silenced 18 of 20 policies but failed the strict complete-suppression test, and the authors limit the claim to adaptive reallocation of a joint message-route code, not emergent encryption, open-world route discovery, or a universal defense.

arXiv:2611.12959 PDF

TrustMod-SM: A Multi-Axis Benchmark for Evaluating Trustworthiness of LLMs in Social Media Content Moderation

Aizan Zafar, Sreekantam Sai Venkata, Gathik Jindal, Ritu Raj Sharma, Zishan Ahmad, Tulika Saha et al. Large language models used as social media content moderators are usually judged on overall accuracy, leaving other aspects of trustworthiness unmeasured. TrustMod-SM is a benchmark of roughly 29,000 instances from eight datasets that scores moderators on five dimensions: trustfulness, fairness, safety, robustness, and context integrity. Testing 16 open-weight models from 0.5B to 284B parameters plus one closed frontier model shows that higher detection accuracy often coincides with greater demographic disparity, and that models comply with concealment or exaggeration prompts, are confidently wrong, rarely escalate ambiguous cases, and confuse hate speech with counter-speech or reclaimed language, especially below 8B parameters. Scaling does not uniformly fix these issues, and no single model excels across all five dimensions.

arXiv:2611.12970 PDF

AI Alignment Is Neither Solvable nor Sufficient

Gerrit Gro{\ss}mann A position piece arguing that the goal of aligning AI systems with human values is misguided as a route to safe AI. The authors make three claims: alignment is not well defined because it presumes a stable, identifiable set of human values; where a concrete notion can be specified, it is not solvable; and where it appears solvable, solving it would not necessarily prevent harm. Their central reason is that alignment is framed as a property of individual systems, whereas emergent risks from interactions and feedback loops in hybrid human-machine societies are likely to become a major source of harm. They conclude that AI safety is an ongoing challenge and not something settled by verifying a single model property.

arXiv:2611.13167 PDF

Enforcing Axioms for AI Alignment under Loss-Based Rules

Alexandros Hollender, Sonja Kraiczy cross-listed Reward models for alignment methods such as reinforcement learning from human feedback (RLHF) are trained by minimizing a loss on pairwise comparisons, and earlier work showed that an optimal linear reward model can violate Pareto optimality (PO), rewarding a response that is worse on every principle. Treating each principle as a voter represented by a linear direction in a fixed embedding space, the authors ask whether richer reward models or different comparison data can restore the property. They show that Pareto optimality can still fail for polynomials of any fixed degree and for one-hidden-layer ReLU networks of any fixed width of at least two, under suitable finite parameter bounds. It is recovered for linear rewards in an idealized setting with uniform comparison directions and equal comparison lengths, which the authors take as motivation for dataset design.

arXiv:2611.13369 PDF

When Reasoning looks like Hidden Understanding: How Chain-of-Thought Fosters Anthropomorphism, Perceptions of Epistemic Superiority, and Cognitive Deference

Oliver Lack, Lana Tikhomirov, Humphrey Thompson, Carolyn Semmler, Melissa McCradden A position piece argues that users mistake the fluent form of chain-of-thought (CoT) output for evidence of understanding and for a transparent explanation of how a model reached its answer. The authors contend that this fosters anthropomorphism and a perception of epistemic authority, which together drive cognitive deference and lead users to generalize from narrow task performance to broad competence, with medicine used as the worked example. They conclude that explainability operationalised through CoT undercuts the ethical principles it is meant to serve, and recommend discouraging user-facing reasoning while granting that CoT may still help accuracy and moderation.

arXiv:2611.13456 PDF

Massive Activations Confound Sparse Autoencoders

Nathan Delisle, Chenhao Tan Sparse autoencoders (SAEs) are usually judged by variance explained and ablation damage, and both measures are distorted when massive activations push up to 99.9% of activation variance onto the first principal component (PC1). Because PC1 has no unique sparse decomposition, each SAE run rebuilds it from different features, so an average of 47 of the 50 most important features have no match in a separate run, and a dictionary can report 99.96% variance explained while leaving 33% of the non-PC1 variance unexplained. The proposed fix trains the SAE on activations with PC1 removed and adds PC1 back to the reconstruction. In affected models this raises top-50 feature sharing from 4-39.5% to 73-99.5%, with up to 44% less model degradation at the same TopK sparsity.

arXiv:2611.13548 PDF

TERM: Target-Free Jailbreak Defense via Entropy-Guided Residual Modulation

Yunkai Dang, Wanyu Zhang, Meiyi Zhu, Qi Fan, Tianyu Ding, Wenbin Li cross-listed Representation-based jailbreak defenses for large language models (LLMs) typically depend on predefined steering directions or learned safety regions, and can over-refuse benign requests. Target-free Entropy-guided Residual Modulation (TERM) is an inference-time defense that calibrates attention entropy against each prompt's own prefill baseline, applies scale normalization and temporal accumulation to the deviations, and uses the resulting signal to apply bounded scaling to the model's own residual updates. Compared with seven defenses on seven jailbreak attacks and two over-refusal benchmarks across Llama, Qwen, and Gemma models, including Qwen3 from 0.6B to 32B, it improves jailbreak robustness while keeping over-refusal low with modest inference overhead.

arXiv:2611.13577 PDF

Mechanistic Interpretability of Deep Neural Networks: Methods, Frameworks, and Perspectives

Lorenzo Cirillo, Irene Amerini, Andrea Cavallaro Internal activations of deep neural networks are hard to interpret, in part because individual neurons frequently represent several underlying concepts at once. This overview article surveys advances in mechanistic interpretability, a field that aims to identify the components whose activations are causally responsible for specific model behaviors, covering methods, frameworks, and perspectives across image, audio, and language domains.

arXiv:2611.13868 PDF

Transparency Disclosures for Chain-of-Thought Monitorability

Jacob Davies, Theo Bearman cross-listed Chain-of-thought (CoT) monitoring is used to oversee advanced AI systems, and the authors argue that frontier developers should publicly report specific indicators so that threats to it can be tracked. They propose five indicators, one for each link in a five-step chain that CoT monitoring depends on: third-party CoT access and base architecture specifications, CoT naturalness and faithfulness, and monitor detection and uplift. An audit of current disclosures from Anthropic, OpenAI, and Google DeepMind finds that no developer covers the full set, but every indicator has been reported by at least one of them. Because disclosures rest largely on developers' own evaluations, the authors also recommend repeating the audit with deeper access through embedded evaluators or an Independent Verification Organization.

arXiv:2611.13873 PDF

Are Causal Neuron Circuits Unique and Transferable?

\"Omer Tekin Yavuz (\.Ileti\c{s}im Yaz{\i}l{\i}m) Interpretability demonstrations often show that intervening on a group of neurons changes a model's answer, without establishing how much of the group is needed, whether other groups would work, or whether the effect generalizes. This study exhaustively measures all 4,096 subsets of 12 selected MLP neurons in small Qwen3, Llama-3.2-1B, and Gemma-3-1B models on a capital-city task and on two-hop relation chains. Patching 12 neurons with first-hop activations shifts the answer to the bridge entity in roughly 8 or 9 of every 10 or 11 chains while random neurons never do, yet the full group is rarely necessary: a few neurons often suffice, and disjoint subsets can each flip the answer on their own. Group effects transfer to rephrasings of the same chain but are much weaker on other chains, and on the capital task the ablated neurons turn out to have nearly constant activations with no evidence of question-specific content.

arXiv:2611.13885 PDF

Specify Privacy Yourself: Assessing Inference-Time Personalized Privacy Preservation Ability of Large Vision-Language Models

Xingqi Wang, Xiaoyuan Yi, Xing Xie, Jia Jia cross-listed Large Vision-Language Models (LVLMs) are reasonably well aligned to protect universally sensitive data such as credit card details, but privacy boundaries vary by person and context. SPY-Bench tests whether models can follow privacy limits a user specifies in language at inference time, using 32,700 samples of image-question pairs with personalized privacy instructions across 67 categories and 24 real-world scenarios. Among 21 models, even the top-performing o4-mini reaches only about 24% compliance accuracy, and contextual privacy understanding is limited. An initial alignment method, a variant of Noise Contrastive Alignment, reaches 96.88% accuracy while keeping reasonable general performance.

arXiv:2611.13997 PDF

Right for the Wrong Reasons? Evaluating Plausibility and Behavioral Change in Explanation-Guided Learning

David Dembinsky, Adriano Lucieri, Andreas Dengel Explanation-Guided Learning (EGL) trains models to agree with human-provided rationales, and this study asks whether better agreement actually means the model changed how it decides. Using a constructed dataset with controlled synthetic shortcuts of localized, background-based, and global kinds, it compares EGL objectives on both behavior and explanations. Rationale-based data augmentation gives the clearest robustness gains, up to +4.8 percentage points in average accuracy, while attribution-constraining objectives raise explanation plausibility by up to +14 points but improve accuracy by at most +1.6 points. Neither improved robustness nor improved plausibility reliably implies higher perturbation-based fidelity.

arXiv:2611.14134 PDF

Representational Atoms in Neural Networks

Chenhui Hu, Pengfei Cao, Kang Liu, Jun Zhao To pin down how neural networks organize learned knowledge and how that relates to human concepts, the authors formally define fundamental representational units (FRUs) and propose two criteria for evaluating them, fidelity and stability. They prove that threshold-activated sparse autoencoders (TSAEs) can identify FRUs with high fidelity and stability under appropriate conditions, and back this with large-scale experiments. Further analysis finds that FRUs with higher stability are more monosemantic and can be composed across layers to recover semantic organization matching human conceptual structure.

arXiv:2611.14168 PDF

Causal Concept Explanations for Deep Neural Models

Josh Rountree, Pulkit Verma, Oswin So, Chuchu Fan, Julie A. Shah Concept-based explanations of neural models are typically correlational, so a cited concept may be neither necessary nor sufficient for the output it supposedly explains. Causal Concept Wrapper Network (CCW-Net) uses interventional mediation analysis, borrowed from economics, medicine and epidemiology, as a training objective that shapes concept representations to be necessary and sufficient for the outputs they are credited with and independent of other concepts' contributions. The same analysis serves as an evaluation tool that quantifies how much of any model's output is causally attributable to each concept explanation. The method is evaluated on aircraft control, driving and image classification; the abstract reports no quantitative results.

arXiv:2611.14407 PDF

Mechanistic Interpretability Needs a Logic of Inference

Zhimin Hu, Chen Jiang, Hanbo Xie, N. Apurva Ratan Murty, Sashank Varma Mechanistic interpretability (MI) can now identify neurons, features, circuits, population structures, and dynamical patterns tied to model behavior, which shifts the question from what can be observed inside a model to how those observations support mechanistic explanations. Drawing on neuroscience's parallel struggle, the authors identify recurring inferential bottlenecks in inferring mechanisms from complex neural systems. They propose five principles organized into a recursive logic for mechanistic inference, meant both to turn local interpretability findings into cumulative knowledge about mechanisms inside models and to make those mechanisms precise enough to generate testable hypotheses about domains beyond the models. They add that evidence from outside MI is ultimately required to establish what model mechanisms reveal about the world.

arXiv:2611.14483 PDF

7 more specialized papers

Reasoning 18

The Shape of a Program: Path Signatures for Trace-to-Program Induction

Mohamed Ghanem, Bernd Finkbeiner Synthesizing a program from its execution traces is treated here through rough path theory: traces are viewed as high-dimensional paths and encoded with the path signature transform, a tool from stochastic analysis whose expectation provably characterizes the distribution of traces a program induces. On the Karel, List, and RASP program-induction domains, the log-signature trace encoder outperforms LSTM, Transformer, and Fourier baselines in almost every setting, with the largest margins when training data is scarce. The pipeline also tends to output valid programs shorter than the ground truth, which the authors turn into a bootstrapping procedure for exposing redundancy in the supervision without an external oracle. The encoder's latent space is also more interpretable, with each state's nearest neighbors lying on its own execution trace.

arXiv:2611.13005 PDF

Trust, but verify, audit and attribute: a human-in-the-loop framework for checking AI-produced research results

Gy\"orgy Eigner, Patrik P. S\"uli, Roland Holl\'os AI systems now generate research results in bulk, and a passing machine check certifies only the artefact that was checked, which for proofs translated into Lean may state something different from the author's text. The proposed human-in-the-loop framework has validated, agent-usable tools with different grounds of trust check each result, a fidelity audit ask whether each check addressed the claim as stated, a human decide at fixed points, and a dialog report attribute work to human and AI. Tool layers are implemented for mathematics (Wolfram Language, MATLAB, Lean) and statistics (Python and R). The mathematical one is demonstrated on a two-claim example, in which an ambiguous claim was split into two readings, one refuted and one proved.

arXiv:2611.13042 PDF

Catch It Where It Happens: Real Time Error Correction During LLM Reasoning

Rotem Elimelech, Michael Shalyt, Ido Kaminer A single wrong step in a large language model's (LLM's) reasoning tends to propagate, and existing safeguards either check only the final answer or call a second verifier model that costs GPU time. The authors instead verify reasoning steps deterministically with code as they are written, flagging a failed check and letting the model continue from the same spot, and demonstrate this on symbolic mathematics using a computer algebra system that checks expressions in milliseconds. On the ASyMoB benchmark with three open-weights models, accuracy improves by 1.4 to 5.6 percentage points, a 1.8 to 14.9% net reduction in incorrect answers, for only 6 to 7% more tokens. Because the verifier needs no model call, it can be attached to any model.

arXiv:2611.13107 PDF

Agreeing on the Answer but Not the Reason: Measurement Limits of LLM Reasoning Failure Taxonomies

Madhu Sudhan Rao Beesetty Papers on large language model (LLM) reasoning often report what fraction of errors fall into named failure categories and compare those fractions across models. The authors test how stable such numbers are by defining one category, Compositional Collapse (a model notices a required sub-part of a problem, then drops or misapplies it), and measuring it across twelve models and eleven languages on a multilingual math benchmark. Half of all failures carry evidence for more than one category, so the category's share ranges from 26.4% to 76.4% depending on the order in which the classifier tests them, and native speakers agree with the automated inspector at kappa = 0.932 on answer correctness but only kappa = 0.093 on the category, a gap largely attributable to the inspector's prompt. They recommend four checks before comparing failure rates: report label co-occurrence, permute the tie-break order, validate on a shared independent sample, and audit systematic disagreements.

arXiv:2611.13140 PDF

What drives the emergence of symbolic mechanisms in large language models?

Melody Zixuan Li, Taylor Webb Prior work found emergent symbolic mechanisms supporting abstract reasoning in large language models, but not what causes them to form. Training networks from scratch on abstract sequence tasks while varying architecture and data distribution, and probing them with representational, attentional, and causal mediation analyses, the authors find that greater data diversity yields stronger symbolic signatures. The same signatures appear in modified transformers and even multilayer perceptrons, leading to the conclusion that data diversity, not architecture, is the primary driver of emergent symbolic computation; an entity-attribute binding task shows consistent patterns.

arXiv:2611.13270 PDF

F3-AGI v4.0: Bioenergetically-Gated Active Inference, Metaplastic Preconditioning, and Residual Program Synthesis for Sample-Efficient Abstract Reasoning

Henrique Oliveira Few-shot abstract reasoning on ARC-AGI is under-constrained by its 2 to 5 demonstration pairs, so program search tends to find spurious programs that fit the examples but fail on test inputs. F3-AGI v4.0 is a CPU-only neuro-symbolic and active inference architecture combining a gated-plasticity optimizer (F3-Metabolic-Optim), complexity-guided program synthesis with residual patching, and exploration driven by Expected Free Energy. On the 400-task ARC-AGI-1 public evaluation set, run offline on CPU with no LLM calls, it solves 57/400 tasks (14.25% pass@2) at a mean 1.62 s per task, of which 38/350 (10.86%) are on tasks the authors did not inspect. It also reports 91.67% final accuracy on a synthetic 5-task continual learning stream and completion of 15/15 levels in a local ARC-AGI-3-inspired interactive suite.

arXiv:2611.13325 PDF

Free Lunch Uncertainty: Sequence-Level R\'enyi Entropy from Particle Decoding for Adaptive LLM Reasoning

Erfan Baghaei Potraghloo, Milad Heidari Sorkhehdizaj, Shahab Rahimirad, Aliasghar Pourghani, Seyedarmin Azizi, Souvik Kundu et al. Estimating uncertainty over a whole LLM completion usually means choosing between cheap token-local signals and costly self-consistency sampling. Sequential Monte Carlo (SMC) decoding for power sampling already estimates the normalizing constant of the power distribution, which yields the sequence-level Rényi entropy (SRE) of the completion distribution at no extra cost, and a chain rule extends this to stepwise entropies of any order. The order-1/2 variant is the best zero-cost selective-prediction signal in 10 of 13 settings and beats log-probability on MATH500 by 0.022 to 0.044 AUROC across four models. An 8-particle probe identifies problems that benefit from more compute, and a two-pass policy matches 64-particle accuracy on MATH500 with 34.2% fewer particles.

arXiv:2611.13414 PDF

Evaluating Compositional Generalization in Transformers: The Role of Composition Equivalence and Module Coverage

Purva Pruthi, Andrew Yuan, Alexander D'Amour, David Jensen Results on whether transformers generalize to unseen combinations of known modules (such as sort and filter) have been inconsistent, largely because train-test splits were chosen ad hoc. A systematic evaluation isolates two factors: composition equivalence, where different module sequences reduce to the same end-to-end function, and module coverage, where modules appear in different contexts at train and test time. Models trained only on final outputs owe most of their apparent success to exploiting composition equivalences, and their performance drops significantly once those are removed, while accuracy degrades monotonically with coverage shift for both direct and step-by-step models. A case study with Llama3-8B on the code-output prediction task of LiveCodeBench shows the same equivalence effect.

arXiv:2611.13498 PDF

How Post-Training Influences Verbalized Algorithmic Computation in Language Models

Niall Dalton The study asks how successive post-training stages change a language model's use of internal information and of its written chain of thought (CoT), using the SFT, DPO, and RLVR checkpoints of Olmo-3-7B-Instruct on multi-digit addition, where carry values serve as known intermediate state. Probes recover carry information from activations about equally well in the SFT and RLVR checkpoints, but manipulating that information does not reliably control outputs. In contrast, CoTs written by the RLVR checkpoint yield higher accuracy across receiving checkpoints than those from SFT or DPO. DPO shows similar sensitivity to carry-related statements without the matching accuracy gain, so greater sensitivity to reasoning content does not necessarily track better performance.

arXiv:2611.13552 PDF

Curriculum proof repair with learned counterexamples

Yuechun Sun, Jun Yang, Kexin Pei cross-listed Verifier-in-the-loop proof repair usually gets only pass/fail feedback, and generating counterexamples that explain a failure normally requires costly iterative refinement at inference time. TopoVerus trains a language model to emit a structured source-level counterexample, a repair rationale and a repaired Verus proof in a single trajectory, optimized with reinforcement learning from verifier feedback, a proof-code safety gate, and a curriculum whose reward follows the topological order of invariant dependencies. On seven Verus benchmarks with 471 problems it reaches 61.6%/72.2% weighted-average Safe-Pass@1/Safe-Pass@3, beating Claude Sonnet 4.5 by 5.7/6.3 points. Its learned counterexamples also improve existing proof-repair approaches' success rate by 64.3% compared with chain-of-thought diagnostics.

arXiv:2611.13627 PDF

A computational framework for understanding embodied plans

Aalap D. Shah, Basil Saeed, Joshua B. Tenenbaum, Tobias Gerstenberg, Hyowon Gweon, Ilker Yildirim Inferring people's goals and plans from their bodily actions is formalized with GenEA, an agent representation built on a multilevel state space that couples the coarse biomechanics of an articulated body with its interactions with nearby objects, extending task-and-motion planning from robotics with simulation-based intuitive physics. Embedding this state space in a generative model yields a joint probability distribution over goals and embodied plans. Across experiments in object manipulation and navigation with multiple behavioral measures, the authors report that GenEA reproduces human judgments about embodied plans and outperforms heuristics, standard task-and-motion planning and a state-of-the-art vision-language model.

arXiv:2611.13669 PDF

Ceci n'est pas une pipe -- Toward a Symbolic Solver for the Abstraction and Reasoning Corpus Using Evolutionary Computation

Luca Favole, Giorgio Mongardi, Giovanni Squillero, Alberto Tonda Evolutionary algorithms (EAs) have been largely absent from the main contests on the Abstraction and Reasoning Corpus (ARC), a benchmark of generalizing from a few examples, partly because program search depends heavily on choosing the right grid representation. The authors describe the first fully symbolic, evolutionary program synthesis approach submitted to the ARC-AGI-2 main contest. It evolves the grid parser and output projector together with the transformation program so that a task-specific language emerges during search, grounds the parser's symbols in core knowledge theory from cognitive psychology, and uses an object-aware fitness function that grades partially correct outputs beyond pixel-level differences.

arXiv:2611.13831 PDF

Where to Approximate in Neurosymbolic Inference?

Samy Badreddine, Emile van Krieken, Luciano Serafini, Antonio Vergari Probabilistic neurosymbolic methods combine neural predictors with symbolic constraints through weighted model counting (WMC), which is #P-hard to compute exactly and so is usually approximated. The authors unify existing approaches under a three-step bottom-up approximate compilation pipeline and ask where the approximation budget is best spent, instantiating two steps with tensor train decompositions that exactly multiply approximated factors and then recompress the products using SVD- or interpolation-based schemes. Across six tasks in Sudoku, arithmetic reasoning, and scene understanding, the new pipelines give WMC estimates orders of magnitude more accurate than prior methods, which translates into stronger downstream learning on tasks where symbolic factors interact tightly.

arXiv:2611.13862 PDF

Evaluating LLM-Generated World Models for Neuro-Symbolic Reasoning

Souhaila Djaffal, Amna Jamal, Gagana Madhugiri Devaraja, Anidipta Pal, Jobst Heitzig Formal world models support planning and verification, and this study tests whether five LLMs can generate them as partially observable stochastic games from natural-language situations, and whether mechanical validity implies faithfulness. A benchmark of 44 scenarios across game, situational and cultural settings is scored on mechanical validity, structural agreement with references and semantic criteria, over 410 first-pass generations. In the reference-backed game track, 134 of 135 generated models pass a strict mechanical audit, yet mean structural F1 is only 0.366 and just 28 exactly match the authored reference. In the cultural track, naming the country changes the generated structure for most LLMs while automated semantic scores barely move, so executability is necessary but not sufficient.

arXiv:2611.14419 PDF

Re-Search: Dissecting LLM Ideation Through a Two-Player Questioning Game

Muhan Gao, Hangxiao Zhu, Xinqi Su, Yu Zhang, Daniel Khashabi, Kuan-Hao Huang Evaluations of large language model (LLM) research ideation usually hide the iterative search that produces an idea. Re-Search adapts Twenty Questions to ideation: given only the research problem of one of 200 ICML 2026 papers, a model must rediscover the paper's core idea by asking yes-or-no questions answered by an oracle, which separates ideation from execution. Across six recent LLMs, nearly all progress happens within the first 30 of 100 questions and most models then plateau, failing either by over-exploiting one direction or over-exploring too many. Controlled interventions on candidate questions and the hypothesis space show that the bottleneck is generation rather than judgment: models recognize a good idea or question when offered one but often cannot produce it themselves, suggesting that autoresearch systems gain more from externally sourced ideas and questions than from better single-model generation.

arXiv:2611.14487 PDF

3 more specialized papers

Robotics 15

Embodied Science Agents: Making Embodied Experimental Design Learnable

Yihang Che, Zhongwei Yu, Yang Li, Xingyu Lu, Hongxia Yu, Jun Wang Self-driving laboratories (SDLs) increasingly use foundation models as disembodied planners that emit symbolic workflows for separate robots, which the authors argue blocks learning of physical experimental skill. The Embodied Science Agent (ESA) combines a Digital Agent using memory-based reasoning and Bayesian optimisation, an Embodied Agent for generalisable instrument control, and a unified interface coordinating the two. In autonomous wastewater treatment experiments spanning over 200 million possible conditions, 15 continuous rounds raised final solution transmittance from 82% to 99.7%, a 98.5% reduction in residual dye concentration, and the learned visuomotor policies generalised across different instrument placements.

arXiv:2611.12958 PDF

PhysMoE: Fast Spatial Grounding with Single-Token Coordinate Representation in Modality-Decoupled VLMs

Ruimeng Liu, Denan Liang, Ruofei Bai, Fuming Zhang, Kuan Xu, Shenghai Yuan et al. cross-listed Embodied vision-language models used for robotic planning predict coordinates either as discrete tokens, which breaks spatial continuity and is slow to decode, or by point regression, which fails on multi-contact affordances and irregular shapes. PhysMoE shares lower layers for multimodal context and routes upper layers to separate spatial and linguistic experts, encodes each coordinate as a single token by reusing the visual encoder's positional embeddings, and supervises pointing with a differentiable Signed Distance Field (SDF) loss instead of a mask decoder. Compared with its base model it achieves a 15.7% average gain and an 8.31x throughput speedup across 9 coordinate-based grounding benchmarks, while preserving performance on 11 text-based embodied benchmarks.

arXiv:2611.13000 PDF

APS-VLA: Action-Guided and Future-Aware Token Pruning for Efficient Vision-Language-Action Models

Xiaoyang Xing, Chun-Mei Feng, Anthony Cohn, Xiaoliang Gong, Asoke Nandi Vision-language-action (VLA) robot policies are slowed by dense visual tokens, and existing pruning methods rank tokens using static or immediate heuristics. APS-VLA keeps the pretrained backbone frozen and trains a lightweight estimator that scores tokens from the visual observation, language instruction, and proprioception, supervised by matching both full-input policy predictions and demonstrated actions, with an auxiliary objective that predicts future visual features and is dropped at deployment. Tests with OpenVLA-OFT on LIBERO and with pi0.5 and XR-1 on RoboCasa365 show faster inference with modest changes in task success. On five real-world tasks across two dual-arm platforms, mean inference latency falls from 137.25 ms to 93.25 ms, a 1.47x speedup with 61% average success versus 60% unpruned.

arXiv:2611.13030 PDF

How Much of Your VLA Is kNN

Denis Grachev, Damien LaRocque, Omar Moured cross-listed Vision-language-action (VLA) models now exceed 95% success on the LIBERO simulation benchmark, and the authors ask how much of that requires generalizing beyond the training demonstrations. Retrieving nearby demonstrations and replaying their action chunks reaches 91.7% success on LIBERO with no policy training, and shows the smallest drop of all evaluated policies under LIBERO-Plus perturbations. Across nine public VLAs, fine-tuning mostly turns the vision-language backbone into a better retrieval key while the predicted actions stay close to the demonstrations. They also introduce Local Action Consistency (LAC), an offline metric of whether observations that are close in representation space lead to similar future actions, which tracks retrieval success and can guide training of a retrieval key.

arXiv:2611.13160 PDF

Cognitive Multi-Agent Search and Rescue: Language-Grounded Hierarchical Reinforcement Learning for Collaborative Robotics

Sepideh Ravanbakhsh, Hamidreza Moradi Language-guided reinforcement learning has mostly targeted single-agent instruction following, leaving team coordination in partially observable search and rescue (SAR) settings unaddressed. The proposed LG-HRL (Language-Grounded Hierarchical Reinforcement Learning) framework uses a large language model to turn human instructions into structured semantic cues that drive high-level strategy selection and role assignment across aerial and ground robots, which coordinate through a shared cognitive workspace. In a simulated disaster environment the authors report 2.3 times faster learning and 20-30% higher victim rescue rates, along with emergent role specialization and immediate adaptation to verbal updates. Ablations indicate that both the shared workspace and the language grounding are needed for these gains.

arXiv:2611.13161 PDF

Curie's Jarvis: Autonomous Robotic Chemistry Experiments from Printed Protocols

Rumi Loghmani, Bozhang Song, Victor Alvarado, Aliyah Redding, Amir Barati Farimani cross-listed Lab automation usually relies on scripted procedures and containers at fixed positions or with fiducial markers, unlike a human who works from a printed sheet and reads labels. Curie's Jarvis lets a robot arm run chemistry experiments on a bench set up for people: it reads the steps from a photo of the printed protocol, finds unmarked containers, distinguishes identical ones by reading their labels from four camera views, re-locates its target before each step, and adapts motion speed to what it is handling. On a Franka Panda arm it completed 5 of 5 runs of one children's chemistry kit protocol and 4 of 5 of another without human intervention, and identified the container correctly in all 160 label-reading trials.

arXiv:2611.13251 PDF

Discrete Model-Based Diffusion for Combinatorial Optimization in Multi-Robot Task Allocation

Haejoon Lee, Thomas Marshall Vielmetti, Xinyi Wang, Dimitra Panagou cross-listed Task allocation in multi-robot systems leads to NP-hard problems over discrete variables that must be solved within seconds, and existing diffusion-based solvers either work only in continuous spaces or need offline training. Discrete MBD extends Model-Based Diffusion (MBD), a training-free form of sampling-based optimization (SBO), to discrete variables by approximating the categorical reverse transition distributions with Monte Carlo sampling. The authors prove a finite-step lower bound on the probability of finding a global minimizer that converges to one as denoising steps increase. On generic integer programs and heterogeneous multi-robot coalition formation and scheduling, it beats sampling-based baselines in nearly all settings and is competitive with mixed-integer programming and reinforcement learning methods at a fraction of their computational cost.

arXiv:2611.13382 PDF

Ego Guidance: Enhancing Robot Policy Learning with Action-Free Human Egocentric Videos

Lin Long, Tianren Ma, Zhong Chu, Di Wu, Maosong Cao, Qixiang Ye et al. cross-listed Human egocentric videos without action labels are studied as a training signal for robot policies. Building on a pretrained vision-language model, Shared Prediction on Action and Representation (SPAR) is an architecture in which predicting future visual states provides a learning signal that is independent of and complementary to action prediction. The authors find that representations learned from egocentric video improve performance on out-of-distribution robot tasks, with further gains as video data scales, and they also analyze the effects of objective choices and data curation.

arXiv:2611.13685 PDF

One Policy for 100 Household Tasks: Our Solution for the 2026 BEHAVIOR Challenge

Ilia Larchenko, Gleb Zarin, Atindra Nair, Keerti Manney, Vitaly Kleban, Mirko Kiefer et al. cross-listed A single vision-language-action (VLA) checkpoint built on pi0.5 covers all 100 long-horizon household tasks of the 2026 BEHAVIOR Challenge in the OmniGibson simulator, finishing 9th on the final leaderboard. Most of the work went into training longer with less gradient noise: a migration from 50 to 100 tasks with warm-started task and stage embeddings, variance-reduced multi-sample flow matching, and a 200k-step run over the 211M demonstration frames, which lifts the Q score on the 50 tasks shared with 2025 from 0.260 to 0.371. Fourteen tasks are instead routed to a GPT-6 Astra coding agent that drives the robot through observation and motion tools, where it reaches a mean Q of 0.639 against 0.108 for the learned policy. The public submission, which included only part of the agent results, scored 0.446 Q with 27.3% full-task success.

arXiv:2611.14432 PDF

6 more specialized papers

Unclassified 1

Visual Similarity in Adult Word Extension and Artificial Word Learning

Patricio M. Vera No summary available — see the abstract on arXiv.

arXiv:2611.14116 PDF