AI Paper Digest

Thursday, November 5, 2026

531 papers cs.AI · cs.LG · cs.CL ← 2026-11-042026-11-06 →

Aug Sep Oct Nov

Highlights

The role of injected noise in diffusion model generalization

Highlight Vincent Emmerling, Felix G\"unther, David Kappel

Diffusion models trained on disjoint datasets produce nearly identical images from the same noise input, a phenomenon known as strong generalization. This workshop study separates the starting noise from the noise injected at every sampling step, and finds that the injected noise sequence, not the starting noise, mostly decides which sample comes out.

  • Using the pretrained UNet models from Kadkhodaie et al. (7.6M parameters, 10^5 downsampled CelebA images) with a fixed T = 250 steps and 100 trials, the authors explicitly control the per-step noise z_t in an Euler–Maruyama-style sampler and measure sample agreement with cosine similarity.
  • When two runs share the injected noise sequence, their samples reach 0.91 average cosine similarity versus 0.02 when they don't, and similarity stays at 0.93 / 0.89 whether the starting noise is the same or different, so the starting noise barely matters.
  • With shared injected noise, intermediate states converge to a cosine similarity of ≈1.0 within about 20 steps (~10% of sampling); with different noise, the denoised predictions still agree for roughly the first 50 steps, when the model produces coarse "mean face" features.
  • Replacing one of five equal segments of the noise sequence shows the samples are most sensitive in the middle "speciation" regime (regime II, from Biroli et al.'s three-regime account) and significantly more robust in the first and last 20% (paired t-tests, α = 0.05).
  • The core effect holds in an ablation with DDPM sampling on Smithsonian Butterflies, using two models trained on only 500 images each, but the three regimes are much less distinct there; the whole study uses small models and a single similarity metric.

arXiv:2611.03764 PDF

Aligning Training and Rollout Policies for End-to-End FP4 RL on Mixture-of-Experts LLMs

Highlight Large Language Models Hei Yi Mak, Hoang Le, Mehran Taghian Jazi, Yunke Peng, Huan Lin, Zhengzhi Tang et al. The authors study end-to-end FP4 reinforcement learning post-training for Mixture-of-Experts (MoE) language models and find that recovery recipes that work for dense models give no gain over naive FP4 here. The cause is the router's discrete expert choices: when quantization corrections are applied only on the rollout side, the probabilities and expert selections during training and inference drift apart sharply. Their recipe, MoEResQ, applies the corrections symmetrically on both sides, restores training-inference alignment and substantially closes the accuracy gap to full precision.

Low-bit (FP4) quantization makes RL post-training cheaper. However, the residual-correction method that recovers accuracy for dense LLMs (RolloutResQ) fails on Mixture-of-Experts models. The router's discrete top-k expert choice turns small numerical differences between training and rollout into different expert selections. The proposed MoEResQ fixes this by applying the same residual correction on both the training side and the rollout side, so the two policies stay aligned.

  • MoEResQ splits each FP4 linear layer's computation in two: one product with the quantized activation and a second with its quantized leftover error, using the same correction in rollout and in the training forward pass. It also applies the correction to the output gradients and inputs in the backward pass, while keeping the routers (under 0.5% of parameters) in BF16.
  • On Qwen3-30B-A3B trained with GRPO on DAPO-Math-17K in HiF4, asymmetric RolloutResQ scores below the naive FP4 baseline on three of four benchmarks; for example, AMC23 drops to 62.34 against 65.23 for plain FP4 and 85.10 for BF16.
  • Alignment metrics explain the failure: correcting only one side leaves the correlation between training and inference probabilities near FP4's 0.94. Correcting both sides restores it to 0.994 (BF16 is 0.995) and brings middle-layer exact top-k expert matches back from about 0.62 to about 0.87.
  • Quantizing both attention and expert layers, MoEResQ narrows the gap to BF16 from 19.87 to 0.65 points on AMC23 and from 24.27 to 3.75 points on AIME24. Quantizing only the expert layers (up to 94.66% of parameters) brings every benchmark within 1.5 points of BF16.
  • All experiments use simulated quantization on one model and on math benchmarks only, with no measured speed or memory gains, so the efficiency benefit is assumed rather than shown.

arXiv:2611.04376 PDF

LoopVQ: Sub 1-bit KV Cache Quantization for Looped Transformers

Highlight HF pick · 2▲Large Language Models Minsoo Cheong, Sungjoo Yoo Looped transformers reuse the same layers across multiple loop steps to match much larger models with fewer parameters, but each loop step stores its own keys and values, so the key-value (KV) cache grows linearly with loop count and becomes the memory bottleneck. LoopVQ exploits the observation that this cache is highly redundant along the loop axis while loop steps differ in quantization sensitivity, jointly vector-quantizing channels from several loop steps and fine-tuning the codebooks to match full-precision outputs. At 0.5 bits per element it beats KV-cache vector quantization methods built for standard LLMs on general, long chain-of-thought, and long-context retrieval tasks with Ouro and Huginn, and composed with TaSQ it retains much of BF16 accuracy at 0.25 bits. Implemented in SGLang, it nearly halves decode latency at a 32K-token context and raises peak serving throughput by up to 4.3x and 6.1x on the 1.4B and 2.6B Ouro models.

Looped transformers reuse their weights across loop steps, but every loop step still writes its own keys and values, so the KV cache grows with loop count and becomes the main memory cost. LoopVQ exploits the strong correlation of the cache across loop steps by vector-quantizing channels from several loop steps jointly under one shared index, which pushes KV storage to 0.5–0.25 bits per element.

  • In Ouro-1.4B, a key channel correlates with the same channel at the adjacent loop step at about |r| = 0.85, but with the other channels in its sub-vector at only about 0.16, so each 8-bit code spans G/4 channels across 4 consecutive loop steps instead of G channels of one step, which roughly halves k-means reconstruction error at no extra bit cost.
  • Because quantizing the first loop step degrades the output about 5× more than quantizing the third (measured as KL divergence from the full-precision output), the Fisher-weighted k-means codebooks are then fine-tuned end-to-end to minimize that KL divergence through the loop, which roughly halves the KL again in about 2–6 minutes on one GPU.
  • On Ouro-1.4B at 0.5 bit, the average over GSM8K, MATH500, HumanEval and MBPP rises from 32.3 (CQ) and 43.6 (TaSQ) to 68.3 for LoopVQ and 73.6 for TaSQ+LoopVQ, against 75.7 for BF16, with similar gains on Ouro-2.6B and Huginn-3.5B.
  • On long chain-of-thought reasoning with Ouro-2.6B-Thinking at 0.5 bit, TaSQ+LoopVQ scores 63.8 vs. 64.9 for BF16, while the baselines collapse to 16–28 and often run out the 32K-token budget; LoopVQ also leads on RULER needle-in-a-haystack retrieval up to 64K context.
  • An SGLang kernel roughly halves decode latency at 32K context and raises peak throughput 4.3× and 6.1× on the 1.4B and 2.6B Ouro models, but prefill is slightly slower, serving speed is compared only against BF16, and there is no accuracy comparison with the scalar looped-model KV quantizers FlashLoop and ResidualQuant.

arXiv:2611.05304 PDF

Loop Before You Leap: What Recurrent Depth Can and Cannot Buy in Data-Limited Pretraining

Highlight Large Language Models Daichi Hotta, Yohei Kobashi, Takeshi Kojima, Yusuke Iwasawa, Yutaka Matsuo Multimodal models have often scored below text-only baselines in the BabyLM challenge, and the authors test whether looped (recurrent-depth) Transformers, which reuse layers to add computation without adding parameters, can close that gap under a 100M-word budget. In a pre-registered text-only comparison at matched inference compute, a 12-layer stack looped four times beats a standard 48-layer Transformer by 5.4 BLiMP points. Adding images through a single frozen DINOv2 prefix vector lowers BLiMP by about 3.8 points in both architectures, and recurrence does not clearly help. Control runs suggest the drop comes mainly from changes in the language data and how it is presented, rather than from the visual input itself.

Multimodal models in the BabyLM Challenge (pretraining on a 100M-word budget) often score below text-only baselines on grammar benchmarks. The authors test whether looping (reusing a Transformer's layers for extra computation without adding parameters) closes that gap. They also run controls showing that much of the gap comes from changes to the language data, not from the images.

  • A 12-layer LLaMA-style stack (97M parameters) is applied four times with the text embeddings re-added after each pass, and is compared against standard 12-, 24- and 48-layer Transformers; in the multimodal runs, each image enters as a single frozen DINOv2 feature vector placed before the caption, and caption words count toward the 100M-word budget.
  • In a pre-registered comparison at matched inference compute, the looped model beats the 48-layer standard model on BLiMP by 5.4 points in text-only training (76.16 vs 70.80) and 3.7 points with multimodal training, but its gain over the parameter-matched 12-layer model is only about 1 point (p≈0.12).
  • Deeper standard models reach lower training loss but score progressively worse on BLiMP (75.21, then 73.84, then 70.80 as depth goes from 12 to 24 to 48 layers), which suggests that sharing weights across loops limits overfitting at this data scale.
  • Multimodal training costs about 3.8 BLiMP points in both architectures, so looping gives no measurable mitigation (the gap between the two architectures' drops is −0.01 points, 95% CI [−1.44, 1.41]); a looped no-image control trained on the multimodal language data scores 5.71 points below the text-only model, and adding the image vector recovers only 1.85 points (95% CI [−1.67, 5.37]).
  • Limitations: the models do use the images (shuffling image features raises caption loss by about 1.5 nats per token), but runs use only three seeds, each image is reduced to one global vector, the effect of looping is not separated from the effect of re-adding text embeddings, and the batch-matched no-image control was run only for the looped model.

arXiv:2611.06206 PDF

Mass Producing Training Data with Surprising Side Effects

Highlight Safety & Alignment Yaowen Ye, Jacob Steinhardt, Sarah Schwettmann, Cassidy Laidlaw Training data can change language model behavior far outside its domain (insecure code can cause broad misalignment, for example), but only a few dozen hand-found examples of this exist. The authors automatically search for datasets with surprising side effects (DSSEs) that elicit a target behavior, read like ordinary post-training data, and would surprise human auditors, scoring each objective with LLM judges and optimizing via evolution, reinforcement learning, and subliminal learning. The pipeline yields 6,575 DSSEs across 8 behaviors, many of which generalize across model families and sizes. The automated metrics track ML researchers' judgments, and frontier models predict the side effects better than humans do.

Fine-tuning data can change a language model's behavior far outside its training domain: insecure code can induce broad misalignment, and outdated bird names can induce a 19th-century persona. Only a few dozen such cases are known, each built by hand or found by accident, so the authors search automatically for datasets with surprising side effects (DSSEs) to build a large collection for studying generalization.

  • A DSSE is scored on three objectives, each judged by an LLM: it must elicit a target behavior, read like ordinary post-training data, and look surprising to a human auditor reading it.
  • The pipeline searches for high-scoring datasets three ways, using evolutionary search, reinforcement learning, and subliminal learning, then re-scores and filters candidates with held-out metrics to avoid overfitting to the judges.
  • The result is 6,575 DSSEs spanning 8 target behaviors, and many of their side effects carry over across model families and sizes.
  • The automated metrics track the judgments of ML researchers, and frontier models outperform humans at predicting side effects from the data alone, which suggests side effects could be flagged before deployment.
  • All three objectives are scored by LLM judges, so "surprising" and "ordinary-looking" are proxies for human judgment, and the collection covers only 8 behaviors.

arXiv:2611.06324 PDF

What Does Multi-Turn Reinforcement Learning Really Improve in LLM Agents?

Highlight HF pick · 15▲Agents Kejian Zhu, Zhuoran Jin, Shangqing Tu, Tianyi Men, Yupu Hao, Jiachun Li et al. Multi-turn agentic reinforcement learning (RL) looks strong on long-horizon tasks, but the authors ask where those gains actually come from and how much they depend on interaction limits. Using a Pass@(K,T) metric that varies both the number of samples K and the number of turns T, they find that trajectory-level RL gains depend on generous turn budgets and shrink sharply when budgets tighten. Agentic RL does not uniformly improve single-step decisions; its benefit lies mainly in recovering from errors and staying on track over long horizons. They briefly examine process supervision signals and propose BudgetMix RL, which varies interaction budgets during training and performs robustly across execution settings.

Multi-turn agentic RL lifts LLM agents' success rates on long-horizon tasks, but those gains depend on having enough interaction turns. The authors evaluate along two axes, sampling breadth K and turn budget T, using Pass@(K,T). They find RL mainly improves error recovery and long-horizon execution rather than individual decisions, and they propose BudgetMix RL to make the gains hold up under tight budgets.

  • Across DeepCoder-MT, BFCL, InterCode-SQL and FrozenLake, GRPO-trained Qwen3 agents beat their base models' Pass@K boundary as K grows, but the advantage shrinks and can even reverse when the turn budget T is small, even on tasks where plenty of headroom remains.
  • When base and RL policies are compared from the same states, RL picks the better next action on only 46.8% of DeepCoder-MT states and 42.8% of FrozenLake states, yet after an injected error it wins 65.7–67.3% of the time and finishes more often from those states (51.3% vs. 38.6% on FrozenLake, 38.5% vs. 22.9% on DeepCoder-MT).
  • A likelihood analysis on Terminal-Bench suggests RL learns to chain plausible steps into complete trajectories rather than produce new kinds of actions: RL actions stay locally plausible under the base policy, but successful RL trajectories reach a cumulative log-likelihood ratio of ~228 vs. 26 for successful base-model trajectories. During training, trajectory reward rises +0.331 while per-step reward falls 0.097.
  • BudgetMix RL leaves the reward unchanged and instead samples each task's turn budget uniformly from 10%–100% of the maximum, stating the budget in the prompt. Across GRPO, GiGPO and PPO it gains an average of +7.5, +5.3 and +4.0 points over full-budget RL at 25%, 50% and 75% budgets, and roughly matches it at the full budget.
  • Accurate process supervision (oracle Q-values, GiGPO) also reduces budget dependence on FrozenLake, but it brings no improvement on BFCL, where no exact step-level oracle exists. The authors therefore present their mitigation discussion as preliminary, with experiments limited to 4B and 8B Qwen3 models.

arXiv:2611.06483 PDF

EmoReward: Self-Evolving Reward Calibration for Emotional Video Generation RL

Highlight HF pick · 11▲Vision Mingzhe Zheng, Yue Wu, Jing Tan, Chang Zou, Dengyang Jiang, Zeyue Tian et al. Reinforcement learning with Group Relative Policy Optimization (GRPO) works well when correctness can be checked, but emotional expression in generated video has no single correct target, and fixed rewards invite reward hacking. EmoReward plans grounding cues from the emotional prompt, scores videos with several signal primitives, and fuses them through instance-level gates. A RewardMemory component then estimates each primitive's reliability from the finished training run and fixes bounded weight multipliers for the next run, without updating the reward models themselves. Experiments on HunyuanVideo 1.5 show gains across multiple RL algorithms, with transfer results on SD 3.5 and FLUX.

Fixed reward fusions handle emotional video generation poorly: facial, motion and color cues are each reliable only in some scenes, and a fixed mix makes reward hacking easier. EmoReward calibrates which reward providers to trust at two levels: inside each video through rule-based gates, and across training runs through a frozen reliability table learned from the previous run's reward traces. No reward model is retrained.

  • A planner turns each prompt into an EmoScript (valence and arousal targets, facial action-unit intensities, body cues, an expression window), and seven frozen providers score the video, including a VLM judge, facial action units, CLIP-AUTT, R1-Omni and motion naturalness. Reflective Reward Gating then masks or down-weights providers using phase, face-confidence, conflict and subtlety gates, and renormalizes the weights over the providers that remain active.
  • RewardMemory correlates each provider's scores with the reference outcome inside each emotional regime of a completed run, then turns these correlations into bounded multipliers (temperature 0.5, log-cap 1) that stay fixed for the next run; memory can boost a provider but cannot revive evidence the gates have marked invalid.
  • On HunyuanVideo 1.5, VideoAlign Overall improves over every optimizer baseline: +0.32 for DanceGRPO (gating only), +0.23 for FlowGRPO, and +0.56 for SAGE-GRPO (−0.83 to −0.26). PickScore and CLIPScore changes are small and mixed.
  • In a seven-policy controlled comparison under SAGE-GRPO, the full system has the best reliability selectivity (0.428) and composite scores (0.922 core, 0.884 emo). In an 18-person study (540 judgments per cell), it raises Good ratings for emotion fidelity to 60.0% versus 53.0% for SAGE-GRPO and cuts Bad ratings to 15.4% versus 31.5%.
  • Caveats: VideoAlign serves as both the training reward and the memory's outcome signal, so the gains partly measure agreement with the objective rather than independent emotional fidelity; only one reference-to-calibrated run transition is tested, there is no matched global-memory control, and transfer to SD 3.5 and FLUX gives mixed results that depend on the objective.

arXiv:2611.06518 PDF

Interactive World Simulation with Cosmos for Physical AI

Highlight Robotics NVIDIA, :, Anis Ahmad, Jon Allen, Malik Aqeel Anwar, Jesse Archer et al. NVIDIA adapts its Cosmos world foundation models into a shared framework for real-time interactive world simulation, where visual observations stream back in response to actions and scene controls. The framework can either predict the visual consequences of actions directly or render observations from physical state kept by an external simulator. Demonstrations cover camera navigation, robot manipulation, controllable video transfer, and multi-view autonomous vehicle (AV) simulation, and code and model checkpoints are publicly released as a base for building virtual environments to train and evaluate embodied agents.

Physical AI agents need virtual environments they can act in and get visual feedback from in real time. The framework adapts Cosmos world foundation models into streaming, interactive world simulators that produce new visual observations in response to actions and scene controls.

  • The framework supports two modes: it either predicts the visual consequences of actions directly, or renders observations from physical state held by an external simulator, so the generative model can serve as a visual layer on top of a conventional physics engine.
  • It is demonstrated on camera navigation and robot manipulation, controllable video transfer, and multi-view autonomous vehicle (AV) simulation, all built on one shared foundation.
  • Users and embodied agents can both interact with the environment as it evolves, which positions the framework as infrastructure for training and evaluating Physical AI policies rather than only for offline video generation.
  • Code and model checkpoints are publicly released through NVIDIA's Cosmos GitHub repository and a Hugging Face collection.
  • The abstract gives no quantitative results such as latency, frame rate, visual fidelity, or physical-consistency metrics, so how well the system performs in real time and how accurately it simulates physics cannot be judged from this material.

arXiv:2611.06627 PDF

Turns Matter: Advancing Streaming Video RL via Turn-Level Credit

Highlight HF pick · 5▲Reinforcement Learning Shuai Dong, Kanghui Tian, Linghan Chen, Wenqi Shao, Hongrui Wu, Haowen Hou et al. In streaming video understanding, models repeatedly decide whether to respond or keep watching, and existing Group Relative Policy Optimization (GRPO) methods give every turn the same trajectory-level credit, which can reward redundant replies and penalize useful follow-ups. TurnWise assigns credit per turn by scoring each decision's remaining return through silence-completed prefixes, and averages token losses within each turn before averaging across turns so short and long decisions count equally. Without a learned critic or extra rollouts, it outperforms prior RL approaches for streaming video understanding.

Streaming video assistants trained with GRPO give every respond-or-wait decision in a trajectory the same advantage, so a redundant reply can inherit credit from an earlier success, and a useful follow-up can be penalized for earlier failures. TurnWise gives each decision its own credit based on the reward still to be earned from that turn onward, and weights every turn equally in the loss.

  • How credit is assigned: each trajectory prefix is re-scored with all later turns set to silence (NO REPLY), and subtracting that score from the full reward gives each turn's remaining return. The turn's advantage is that remaining return compared against the other rollouts in its group, computed from cached judge scores with no learned critic, extra rollouts or extra judge calls.
  • How the loss is weighted: token losses are averaged within each turn and then across turns, so a long response and a short wait count equally regardless of output length.
  • Main results: TurnWise gets the best aggregate scores on ProactiveVQA, OVO-Timing and StreamingBench with both backbones. On Qwen2.5-VL-3B it reaches 43.56 PAUC, 21.59 Timing F1 and 49.43 accuracy, ahead of the best RL baseline by +3.96, +3.87 and +0.69 points. On Qwen3-VL-4B the margins are +6.26, +1.92 and +2.38.
  • Why both parts are needed: combining the two cuts the repetition rate by 61.5% compared with credit assignment alone (15.53 to 5.98) and adds 2.96 Timing F1. In training logs, turn-level credit flips 59.83% of repetitive replies that had a positive trajectory-level advantage to a negative one.
  • Limitations: the experiments are small (3B–4B models, 3,983 RL instances, 4 rollouts per group) and use a single reward from MMDuet2. The StreamingBench gain on the 3B model is under one point, the α ablation is non-monotonic, and the method relies on scoring trajectories with silent suffixes, which needs an explicit wait action.

arXiv:2611.06728 PDF

FrontierPhysics: Evaluating Agents for End-to-End Physics Research

Highlight HF pick · 22▲Agents Bingran You, Wenjun Ke, Ria Rosenauer, Bo Liu, Wenbo Chen, Min Chen et al. Existing physics benchmarks test fixed questions, isolated sub-tasks, or reproductions of known results, and rarely judge the quality of the research itself. FrontierPhysics has 56 end-to-end tasks built by 42 physicists from their own research, spanning nine theoretical, computational, and experimental domains. In each task, agents must review the literature, plan and carry out the research, and write a manuscript, which is graded by deterministic verifiers and expert-written rubrics. Across 22 models in 26 configurations, no configuration passes more than 46 of its 168 rollouts. Manual review traces the failures mainly to insufficiently rigorous reasoning and weak expert intuition, rather than missing physics knowledge or execution errors.

Physics benchmarks so far mostly test fixed questions, isolated subtasks, or reproduction of a paper the agent is given, and rarely judge research quality alongside correctness. FrontierPhysics instead gives agents open-ended research projects, built by 42 physicists from their own work, and grades both whether the numerical results are correct and whether the paper the agent writes would be publishable.

  • The benchmark has 56 tasks, accepted from 182 submissions, across nine domains spanning theoretical, computational, and experimental physics; each starts from a PI-style brief and asks the agent to review the literature, plan and run the research, and deliver both numerical outputs and a paper.pdf.
  • Grading combines deterministic verifiers on the numerical outputs with expert-written rubrics scored by a reviewer agent (GPT-5.6 Sol in Codex CLI); a rollout passes only if every verifier test and every "blocker" rubric item passes, and non-blocker items then set a 0–100 quality score.
  • Across 22 models in 26 configurations, each attempting every task three times, no configuration passes more than 46 of 168 rollouts; the best is Claude Opus 5.5 in Claude Code at 27.4% Pass@1, and pass rate generally rises with cost, though not across the Claude models.
  • Experts manually reviewed all 168 GPT-6 Astra rollouts and found that agents rarely fail on physics knowledge; the main failures are insufficiently rigorous reasoning (48%, for example self-checks that rely on the agent's own wrong assumption), missing expert intuition (34%), weak literature synthesis (27%), and instruction following (26%).
  • The grading itself makes mistakes: nine GPT-6 Astra rollouts were wrongly failed, and correcting them raises its pass rate from 14.9% to 20.2%; this correction and the failure taxonomy cover only that one model, an Opus 5.5 reviewer grades its own model family more leniently, and with few tasks per domain, per-domain comparisons are uncertain.

arXiv:2611.06910 PDF

Applications 116

BEND 2.0: Benchmarking the Capabilities Needed to Design DNA

Andreas Bjerregaard, Rachael DeVries, Dionysia Danai Brilli, Panagiotis Antoniadis, Yan Li, Anders Krogh et al. cross-listed Genomic foundation models are increasingly used to propose DNA sequences and edits, but existing benchmarks do not test the capabilities that design actually needs. BEND 2.0 (BENchmark for DNA language models) covers eight tasks built on five experimental datasets and tests both how well models predict the effects of DNA changes and whether those predictions can guide sequence design. Linear readouts fitted to frozen representations can match or beat models' native outputs and transfer well across assays and genes, and functional post-training substantially improves results. However, strong prediction of average effects can hide weak prediction of context-specific effects and mutation interactions, and design tasks still leave a lot of room for improvement.

arXiv:2611.03806 PDF

Thinking Fast, Trading Late: System One Decision Models vs the Market

Amir Dehkordi (Independent Researcher), Alireza Kargarzadeh (TailState Intelligence Ltd), Nariman Khaledian (Independent Researcher), Navid Parvini (Zanista AI Ltd), Sid Ghatak (Increase Alpha, LLC) et al. cross-listed Backtests suggest that large language model (LLM) readings of company news predict stock returns, and the authors test how much of that survives once look-ahead bias, memorization and selection effects are ruled out. Their protocol, OLIVIA, uses point-in-time inputs with company names masked, selection on validation data only, an analysis plan frozen by cryptographic hash, and a test period unblinded once. A fine-tuned open-weights model's signal held up on validation but fell to an insignificant level on the blinded 2026 test. Commercial models' readings predict only the overnight move to the open, so the news is already priced in by the time a trader can act, and no pre-registered strategy survives trading costs. Commercial models also identify the masked companies in 62 to 74% of cases.

arXiv:2611.03985 PDF

What Does 4-Bit Microscaling Cost a Wearable Activity Recognizer?

Megan Healy, Hossein Valavi The 4-bit block-scaled formats MXFP4 and NVFP4 have been evaluated on general benchmarks and large models, but not on tiny on-device sensor models. The authors train a 159K-parameter 1D CNN to recognize basketball actions from wrist accelerometer data in Hang-Time HAR, evaluated on unseen players with player-grouped cross-validation. With both weights and activations at 4 bits, MXFP4 loses 5.14 macro-F1 points while NVFP4 loses only 0.75 and wins in all five folds. Ablations trace most of MXFP4's loss to its round-down scale rule, which clips block maxima; rounding the scale up recovers over 3 points.

arXiv:2611.04185 PDF

Anomaly Detection on Resource-Constrained Devices with Reservoir Computing

Marco Angioli, Mathias Persson, Hiroto Tamura, Hadi Banaee, Amy Loutfi, Gouhei Tanaka et al. The authors adapt MDRS (Mahalanobis Distance of Reservoir States), an unsupervised time-series anomaly detector based on reservoir computing, for microcontroller-class devices. They replace random reservoir connectivity with a circular shift (MDRS-Circ), show that the anomaly score for a whole series can be computed from the final reservoir state alone, and use the reservoir's temporal memory to make input preprocessing cheaper. On the MLPerf Tiny anomaly detection benchmark, only MDRS and the reference autoencoder meet the performance target, and MDRS needs fewer operations and a smaller model. Deployed on a NUCLEO-F401RE board, it reaches an ROC AUC of 0.869 with a 24 KB model, about 24 µJ per inference and 0.32 ms latency, a 75.9x speedup over the autoencoder with standard preprocessing.

arXiv:2611.04835 PDF

Can saliency maps flag chest radiograph classifier errors at a new hospital? An external validation study

Aashish Chapagai, Angad Singh Gurtatta, Erik Cambria cross-listed Saliency maps are often presented as a safeguard for medical image classifiers, so the authors test whether they flag individual errors when a model is used at a new hospital. They trained a convolutional network and a vision transformer on NIH ChestX-ray14 and evaluated 303,768 decisions per model at the training hospital and two external sites in the USA and Vietnam, scoring explanations by how concentrated their attribution was in segmented thoracic regions. Grad-CAM localized findings less often than radiologists (hit rate at most 0.575 versus 0.888), and adding saliency to a confidence-based error detector never improved its AUC by more than 0.011, with the relationship reversing across hospitals for the convolutional network. The authors conclude that model confidence, not saliency, should anchor deployment monitoring.

arXiv:2611.05011 PDF

Multi-Objective Generative Optimization via Expected Hypervolume Fine-Tuning

Juan Guevara, Riccardo De Santi, Andreas Krause, Raul Astudillo Scientific design problems need sets of candidates that cover trade-offs among competing properties. Steering pretrained diffusion or flow generators toward such sets is hard: inference-time search is costly, and standard fine-tuning uses fixed per-candidate rewards that ignore how designs complement one another. Hypervolume Fine-Tuning (HVFT) fine-tunes a generator to maximize the expected hypervolume of its samples while limiting drift from the pretrained model. The authors show this objective cannot be written as a fixed per-candidate reward. Its first variation, however, equals a candidate's expected marginal hypervolume contribution, which yields an adaptive reward with an unbiased sample-based estimator and a sequence of KL-regularized scalar-reward subproblems with convergence guarantees. Across therapeutic-peptide and molecular design tasks using both discrete diffusion and flow-matching models, HVFT achieves state-of-the-art expected hypervolume and trade-off coverage.

arXiv:2611.05207 PDF

LLM-based NPCs in Video Games: A Systematic Mapping Study

Dilan Mian, Jeremy Bradbury, Yann-Ga\"el Gu\'eh\'eneuc, Fabio Petrillo, Cristiano Politowski cross-listed Large language models (LLMs) promise non-player characters (NPCs) that converse openly and adapt beyond scripted behaviour, but no one has mapped what has actually been built and how well it works. This systematic mapping study screened 377 papers and analysed 42 against ten sub-questions covering architecture, performance, safety, and reproducibility. Short-term memory (60%) and world-state grounding (69%) are common, while long-term memory (31%), reflection (38%), and retrieval-augmented generation (14%) remain rare. Reported response times range from 2.6 to 15 seconds. The authors identify six structural barriers to commercial adoption and give evidence-based recommendations for developers and researchers.

arXiv:2611.05427 PDF

Hybrid LLM-Deterministic Pipelines for Enterprise-Scale Code Migration at Bloomberg

Mario Corchero, David Williams, Laszlo Kiss Kollar, Haowen Liu, Shivam Bairoliya, Gala Malbasic et al. cross-listed Bloomberg's Automated Uplift initiative applied LLM-assisted migration pipelines to thousands of internal repositories in two campaigns: a varied Python packaging migration (about 1,900 candidate repositories) and a uniform C++ build-system migration (about 6,700). The architecture does as much as possible deterministically, limits the LLM to bounded file transformations within a pre-computed context, and feeds in runtime production telemetry such as execution logs and usage traces. The campaigns produced 2,333 pull requests and 729 successful migrations. Generating the changes turned out to be the easy part. The bottleneck was the cost to owning teams of reviewing, testing, and deploying them, and that cost fell when changes could be verified inside the pipeline (for example, by compiling and testing C++) and when repositories were similar enough for fixes to carry over from one to the next.

arXiv:2611.05457 PDF

Can Large Language Models Review the Scientific Literature? A Test Against a 102-Article Benchmark

Abdullah Al Kalbani, David M. Schultz, Alexandra K. Anderson-Frey, Zhonghua Zheng, Catherine McGuire Six AI tools, including Codex and Claude Code, were tested on a full literature-review workflow, using a published review of the Spanish plume weather pattern and its 102-article corpus as ground truth. The best tool found 42 articles on its own, and institutional library access raised discovery to 82 for Codex and 79 for Claude Code. Given all 102 PDFs, three tools correctly identified all 18 articles that discuss the topic substantially, but they misclassified 40% to 79% of passing mentions as substantial. Two domain experts, grading blind, found the tools' syntheses largely accurate but limited in insight, with some unsupported claims, quotation errors, and overstated consensus. No single tool was best across all stages, and the authors give eight recommendations for researchers.

arXiv:2611.05459 PDF

Decision Models as Scalable Economic Surrogates that Capture Human Variation

Maxime Saxena, Fabio Caccioli, R. Maria del Rio-Chanona cross-listed Decision Models (DMs) are a newer class of AI model that respond quickly and return a probability distribution over a fixed set of options, which makes them candidate stand-ins for human participants in economic simulations and surveys. Ten DMs and eight LLMs were compared on a household spending survey, several classic economic games, and a labour-market field experiment. Frontier DMs reproduced the variation in human responses more closely than LLMs in all three settings. They matched LLMs at predicting the most common human answer on the survey but did worse on the games. They were also roughly ten times faster than the fastest LLM and up to over a hundred times cheaper.

arXiv:2611.05553 PDF

TACT: A Pragmatic Approach to Qualitative Analysis with an AI Council

Rudrajit Choudhuri, Christian Bird, Carmen Badea, Anita Sarma cross-listed Software teams often leave thousands of qualitative stakeholder responses unanalyzed or hand them wholesale to a large language model. TACT (Traceable AI Council for Thematic analysis) runs thematic analysis at scale through a council of models while leaving interpretation and judgment to the team. TACT-Lite is the core council procedure, and TACT-Full adds models from other families that critique each output, plus a decision model that flags uncertain code assignments. On 256 Stack Overflow threads coded by other researchers, TACT's final coding reached an F1 of 0.71–0.72, and its agreement with the authors' own manual coding was Cohen's kappa of 0.79–0.85. Model calls cost $13.77–$33.59 per 100 threads, and the authors list pitfalls and safeguards for AI-assisted qualitative analysis.

arXiv:2611.05702 PDF

A Hybrid Approach for Automated Security Pattern Recovery

Anh Minh Nguyen, Tuan Dat Tran, Alzubair Hassan, Liliana Pasquale cross-listed Finding where security design patterns are implemented in a codebase is hard because the patterns are abstract, span many components, and can be built in many valid ways. SecPatRec treats each pattern as a reference architecture: a ReAct-style LLM agent with semantic and lexical code search finds which roles are present and maps them to code entities, and static code property graph analysis recovers how those roles interact. On a human-validated dataset of 100 pattern instances across 93 Java and Python repositories, averaged over four LLMs, it reaches 0.84 role-detection accuracy, 0.84 role-grounding similarity, and a 0.66 interaction-recovery score. It also beats a one-shot end-to-end LLM baseline at lower estimated API cost, and handles repositories too large for that baseline's input limit.

arXiv:2611.05823 PDF

Does Your Trial Simulator Know What the Drug Does? An Intervention-Blind Test of Interventional Validity on 7,207 Registered Trials

Alicem Koyun cross-listed Patient world models are pitched as clinical trial simulators, but whether they actually understand what an intervention does has not been measured. The authors build a benchmark of 9,996 within-trial contrasts from 7,207 registered trials and define interventional skill as the error a model removes compared with an intervention-blind copy of itself trained without any intervention-identifying fields. A registry-trained gradient-boosting simulator beats a trivial baseline but has negative interventional skill (-0.12): it does better when the intervention is hidden. A frontier language model reading only the trial protocol reaches 0.20 and reverses its prediction when the arms are swapped in 86% of contrasts, while smaller and open-weight models sit at the limit of detection.

arXiv:2611.05907 PDF

Beyond Risk Cues: A Mechanism-Level Benchmark and Verifiable Debate for Scam Video Detection on the Web

Yifan Wang, Junyu Lu, Shun Zhang, Chaozhuo Li, Fanliang Bu, Rong Yin et al. Scam-video detection research mostly treats the problem as binary classification, which says little about how scams work and is weak at separating scams from look-alike content such as advertisements. MechScam is a benchmark of 6,614 videos covering scam mechanisms and non-scam categories, annotated with evidence-backed explanations and misleading arguments built from genuine evidence. The authors also propose ScamCourt, in which trained prosecution and defense debaters argue opposing cases and a trained judge decides from a transcript with verified citations. State-of-the-art multimodal large language models (MLLMs) struggle to tell scams from similar legitimate videos, while ScamCourt outperforms larger MLLMs and fine-tuned baselines, with fewer false positives and greater resistance to misleading arguments.

arXiv:2611.06110 PDF

The Split Protocol Decides the Result: Re-Evaluating Encrypted Traffic Classification

Mayan Sharma, Kriti Saini, Devansh Behl cross-listed Encrypted traffic classifiers routinely report 95-99% accuracy, and this study shows that much of that comes from the evaluation split. On one public corpus, the same model and code score macro-F1 0.992 under the standard protocol but 0.567 when flows from the same packet capture stay on one side of the split. On a month of backbone QUIC traffic, server-disjoint splitting was the only one of five protocols that mattered, and the authors propose single-identifier probes (for example, a decision tree trained on the capture ID alone) to detect such leakage before any model is trained. Under corrected evaluation, a metric-learned flow embedding beats four deep baselines but not gradient-boosted trees on twenty packet sizes, and the paper openly reports that five of its own pre-registered architectural claims failed.

arXiv:2611.06290 PDF

When Does Risk-Controlled Throughput Forecasting Actually Control Risk? A Reproduction Study on LEO Satellite Link

Mayan Sharma, Kriti Saini, Devansh Behl, Suganya R, S. Vijayaragavan cross-listed This reproduction study re-examines BG-CFQS, a state-of-the-art method that turns forecasts of Starlink LEO satellite throughput into conservative bounds whose overestimation rate is held at a user-set risk budget. Its risk guarantee reproduces under random calibration splits but holds on only 1 of 3 datasets under a contiguous temporal split, which is what a deployed terminal actually faces, and it cannot serve budgets below 0.15. The authors also confirm that static conformal calibration only shifts an operating point and cannot improve ranking. Online recalibration tracks the budget within 1%, while conditioning on covariates such as satellite geometry made results worse on three of four datasets.

arXiv:2611.06307 PDF

EnzymeRSI: A dual-loop recursive self-improvement system for enzyme design

Genoria AI Team, Yuejie Hou, Lixiang Yang, Yisi Liu, Weiting Tang, Yueyuxiao Yang et al. Enzyme engineering usually optimizes which candidates to test inside a fixed model and workflow, which offers no escape when the method itself is the bottleneck. EnzymeRSI is a dual-loop recursive self-improvement system: an outer loop evolves harness assets without changing model weights, an inner loop develops prediction and selection algorithms, and a cloud lab is specified to close the physical loop. The outer loop raised mean ProteinGym development-set Spearman correlation by 0.044 across five assays, and the inner loop reached a PCC of 0.849 when self-improving ThermoMPNN. In prospective wet-lab tests, 10 of 32 not-pre-specified RLuc luciferase candidates showed at least twice wild-type activity, and path selection raised the catalytic efficiency of polymerase KOD-E 31.1-fold.

arXiv:2611.06321 PDF

Does Frontier AI Strengthen Cyber Defense? Evidence from Project Glasswing

Thomas Mosk cross-listed Using Project Glasswing, which gave a small group of large US tech firms early access to the vulnerability-finding model Claude Mythos Preview, as a natural experiment, this economic study measures how frontier AI affects defenders, external researchers, and attackers across all publicly disclosed vulnerabilities. Firms with access more than doubled their vulnerability disclosures relative to Chinese and Russian firms without access, disclosed about 150% more high-impact flaws, and patched as quickly as before. Bug-bounty researchers shifted roughly half their reports to other firms' code, and the share of flaws exploited within a month of disclosure fell by about two thirds. The author concludes that restricted release creates a security asymmetry between firms in different countries.

arXiv:2611.06629 PDF

Replicability, Reproducibility, and Robustness in Political and Social Science Research with Large Language Models

Joan C. Timoneda, Sebasti\'an Vallejo Vera LLM outputs vary across runs, models, and versions, which undermines reproducibility when they are used as measurement tools in political and social science. The chapter separates replicability, reproducibility, and robustness, proposes judging results by their distribution rather than identical outputs, and introduces a standardized prompt script plus an error-structure analysis of where models fail and whether their errors align. Across three datasets, the choice of measurement model can flip the sign of downstream regression coefficients; design-based corrections with gold-labeled subsamples restore validity at the cost of precision, which a stacked ensemble of open-weight models partly recovers.

arXiv:2611.06639 PDF

Template Gap Reconstruction for Shell Output Simulation with Small Language Models

Jarrod Ragsdale, Rajendra Boppana cross-listed Small language models can simulate shell output without executing commands, but decoding the full output forces them to reproduce recurring utility formatting along with the parts that actually vary. Template gap reconstruction (TGR) selects a reusable output template containing fixed text and placeholders, and the model generates only the gaps, using a template bank built from 18.7 million outputs across varied environments. TGR matches or beats full-output decoding on output edit similarity, reaching 0.964 versus 0.812 on shorter gap sequences, and cuts generated tokens by 26.3-38.3%.

arXiv:2611.06731 PDF

96 more specialized papers

Large Language Models 73

Contextual Linear Activation Steering of Language Models

Brandon Hsu, Daniel Beaglehole, Adityanarayanan Radhakrishnan, Mikhail Belkin cross-listed Linear activation steering controls a language model by adding a concept vector to its activations at inference time, but existing methods apply the same strength to every input. Contextual Linear Activation Steering (CLAS) splits steering into a fixed, interpretable direction and a steering coefficient that depends on the context. This also allows selective steering of different behaviors in different parts of a single generation. Across many steering benchmarks and models, CLAS consistently beats standard linear steering and matches or exceeds ReFT and LoRA when labeled data is limited, and its directions also work as reliable concept monitors, which fine-tuned directions do not.

arXiv:2604.24693 PDF

Hallucination Is Linearly Decodable from Mid-Layer Hidden States in Quantized LLMs

Aizierjiang Aiersilan cross-listed The authors test whether quantized open-source LLMs carry a linearly readable truthfulness signal in their hidden states. They use Llama-3.1-8B, Mistral-7B, and Qwen2.5-7B loaded in 4-bit NF4, evaluated on TruthfulQA, HaluEval-QA, FEVER, and a synthetic set. A linear probe on a single mid-network layer reaches 0.904–1.000 AUROC, while sampling-based detectors such as self-consistency and INSIDE EigenScore stay at or below 0.541 under the same protocol. MLP probes rarely beat linear ones, the best layers fall in a consistent mid-depth band across model families, and first-block attention entropy adds a free complementary signal. Everything reproduces on a single 8 GB GPU.

arXiv:2606.02628 PDF

A Year in LLM Serving: Workload Evolution, Caching and Load-Balancing

William Nixon, Jon Durbin, Florian Standhartinger, Haryadi S. Gunawi, Juncheng Yang cross-listed Existing studies of LLM serving workloads cover short periods and offer little visibility into how users interact with models in production. The authors characterize a full one-year production trace from a single serving provider spanning many models and users, including long-tail models, and analyze it at the aggregate, temporal, model, and user levels. The analysis surfaces how the workload changes over time and how users group around models, structure that aggregate views usually hide, with implications for caching and load balancing. The complete one-year trace is released publicly so serving systems can be benchmarked on real rather than sampled or synthetic traffic.

arXiv:2608.13573 PDF

Mixture of Wide Experts: Sparse Routing for Interpretable Native Coordinates

Tavish Mankash, Vardhaman Kalloli, Keshava Prasad, Deepan Muthirayan The goal is a language model whose internal coordinates are interpretable from the start, instead of recovering features afterwards with sparse autoencoders (SAEs). Mixture of Wide Experts (MoWE) is a sparse feed-forward layer whose selected experts read the full representation but write to fixed, disjoint output coordinates. In 10-billion-token training runs at roughly matched compute, MoWE reaches lower loss than a dense model (though higher than a standard mixture-of-experts model) and earns a native-coordinate interpretability score of 0.523, against 0.375 for dense and 0.268 for standard MoE, comparable to a TopK SAE trained on the dense model. Case studies show that specific coordinates can be copied in to change a prediction, or used to detect hazardous requests and steer the model toward refusal.

arXiv:2611.03655 PDF

EpiLoRA: Fusing Many LoRA Adapters into Low-Precision GEMMs for Serving and RL

Hieu Tran, Hong Yu cross-listed As base models move to 8-bit (FP8) and 4-bit formats, their matrix multiplications get cheaper, but Low-Rank Adaptation (LoRA) adapters still run as separate small GPU kernels and take up a growing share of each decoding step. EpiLoRA, a plugin for vLLM, runs each adapter's down-projection first and then adds every tenant's up-projection in the final output stage (the epilogue) of the low-precision base matrix multiplication. On NVIDIA Hopper GPUs it decodes with up to 24% lower time per token than vLLM's fastest configuration and beats SGLang and TensorRT-LLM. The same FP8 base can also be used for LoRA reinforcement learning without a 16-bit copy of the weights, so a 27B model trains on a single 48 GB GPU.

arXiv:2611.03899 PDF

FlagRot: Selective Rotation for Low-Bit KV Caches

Kunming Shao, Xijie Huang, Jierun Chen, Jiaqi Shao, Yanli Wang, Jiangnan Yu et al. cross-listed 2-bit KV cache quantization could shrink long-context memory about sixfold, and the usual fix for its error is to rotate keys before quantizing. The authors observe that rotation flattens spiky parts of a key but can create new peaks in flat parts, so FlagRot decides for each 32-channel piece whether to store it rotated or unrotated based on calibrated attention error, and rotates the query instead at read time. Per-channel caches such as KIVI need only one decision per head, while per-token caches store the choice in existing sign bits. At 2.5 bits per value, FlagRot cuts KIVI's twelve-model KL divergence from 18.7 to 12.4 ×10⁻³, and as a vLLM plugin it improves RULER at 128k tokens by 2.9 points.

arXiv:2611.03992 PDF

Testing Processor Contracts in Speculative Decoding Engines

Tianyu Ding, Yaocheng Tong, Muzhi Yang, Lei Yang, Fei Jiang cross-listed Speculative sampling should leave the output distribution unchanged, but only if logits processors, which apply constraints or adjust token scores, receive the correct token prefix and state at each verification step. The authors formalize this processor-prefix contract and introduce BADC (branch-aware differential conformance testing), which compares against enumerated references on tiny random models, checks what processors receive, and puts confidence bounds on output probabilities. On selected Transformers and llama-cpp-python releases, they confirm two failure causes: processors that read unverified draft tokens and processors that advance state without rolling it back. On one prompt, ordinary and prompt-lookup decoding differ by at least 60 percentage points in a first-token probability, although plain permutation testing detected each bug family at least as often as BADC.

arXiv:2611.04042 PDF

HybridSpec: Throughput-Aware Speculative Decoding for Hybrid GDN-MoE Models

Haiyu Wang, Sai Qian Zhang, Zechun Liu, Steven Li cross-listed With cheap parallel drafters such as DFlash, verification by the target model dominates the cost of speculative decoding, and for hybrid models that combine Gated DeltaNet (GDN) and mixture-of-experts (MoE) layers, verification time grows steeply with the number of draft tokens checked. HybridSpec picks the verification length on the GPU from acceptance estimates and measured latency, decides whether to keep one extra token after seeing which experts the MoE router assigns it, and uses block-parallel GDN kernels that rebuild the accepted state in one launch. Those kernels alone are 1.54x to 4.04x faster than the defaults. Across three GDN-MoE models, HybridSpec is 1.75x to 1.88x faster than autoregressive decoding and 1.30x to 1.33x faster than DFlash.

arXiv:2611.04055 PDF

When Strong Reasoners Join the Pool: Small Union Premiums and Costly Runaway Generations in LLM Routing

Teng-Ruei Chen Routing benchmarks often compare routers against an oracle that counts a query as solved if any model in the pool gets it right, which inflates the ceiling by rewarding multiple stochastic samples. The study measures that premium carefully on MATH-500 with 30 seeded samples per problem from Llama-3.3-70B-Instruct, Nemotron Nano 9B v2, and Nemotron Super 49B v1.5, for 45,000 scored responses on a single B300 GPU. Both reasoning models reach 96.1% accuracy, and the multi-model premium is only 0.215-0.588 percentage points, far below an earlier 3.5-point estimate for small models. Resampling Nano three times nearly matches calling all three models once. Only 31 responses hit the 65,536-token cap, yet they account for up to 60.2% of a model's output-length-squared cost proxy, a measure of compute.

arXiv:2611.04098 PDF

Separating Common and Unique Directions for Model Merging

Shiting Wang, Yingjie Zou, Ke Li Arithmetic model-merging methods combine fine-tuned models without training data, but they apply one element-wise rule to all parameter directions, even though some directions merge better than others. In a shared directional basis, the authors find two regimes in task vectors: a low-dispersion common bulk shared across tasks and a high-dispersion, task-specific tail that carries disproportionate energy. Common-Unique Subspace Decomposition (CUSD) is a data-free pre-merging step that merges the common components with an existing arithmetic operator and keeps the unique residuals separate. Across 11 base models, four model families, and four merging operators, CUSD improves 42 of 44 model-operator combinations, by up to 22.5%.

arXiv:2611.04108 PDF

LoopKV: Loop-Aware KV Cache Compression for Serving Looped Transformers

Tianle Zhou, Yang Zhang, Tianyu Wu, Zhuohan Wang, William Nixon, Juncheng Yang Looped transformers such as Ouro reuse one block several times per token and keep a separate key-value (KV) cache for each loop, so cache memory limits the batch size during serving. Measurements show that sensitivity to quantization and to cache sharing differs across (loop, layer) cells. LoopKV uses a one-time offline profile to assign each cell 4-bit storage, 2-bit storage, or reuse of the final loop's cache, and at decode time it fetches the low bits only where attention weights call for full precision, all inside a single fused vLLM kernel. Serving Ouro-1.4B on one RTX PRO 6000, it stores 35% less KV than uniform 4-bit KIVI-4 and sustains 1.7-2.2× its goodput under a latency target, with no benchmark average significantly below BF16.

arXiv:2611.04137 PDF

Tokenizer Myths and Where To Find Them

Rylan Schaeffer, Sanmi Koyejo cross-listed The authors challenge three common beliefs in tokenization research. Using untrained models, they show that Bits-Per-Byte (BPB) is not tokenizer-agnostic: it prefers some tokenizers and vocabulary sizes even when the model has learned nothing, and on trained models it correlates only weakly with downstream accuracy. The reported 0.989 correlation between unigram-normalized loss and BPB comes from a confounder and falls to 0.057 once fixed. They also prove a trilemma: no metric can rank only the true distribution best, give identical predictions the same score, and tie all untrained models, so choosing a metric amounts to choosing a research question.

arXiv:2611.04215 PDF

GC-MoE: Overlapping Expert-Parallel Token Dispatch with FP8/FP4 Grouped GEMM via CUDA Green Contexts

Yong Li, Kan Liu, Tao Lan cross-listed In mixture-of-experts (MoE) decoding with expert parallelism, token dispatch latency and weight traffic dominate. Fused operators such as Mega MoE force dispatch and compute to share every streaming multiprocessor (SM), so they compete for the same resources. GC-MoE is an open-source FP8/FP4 MoE inference operator that uses CUDA Green Contexts to split the GPU into a small communication partition and a large compute partition, each with its own kernel and synchronization. A ring handshake starts computation as soon as tokens arrive, and double-buffered counters remove the barrier between graph replays. It speeds up DeepSeek-V4.1-Flash decode by 1.188x on 8 H20 GPUs and beats Mega MoE at all 54 measured points on B300 while reserving only 20 of 148 SMs for dispatch.

arXiv:2611.04222 PDF

Coarse Semantic Blocks for Efficient Training-Free Sparse Attention

Sosuke Kobayashi, Shoichiro Yamaguchi, Seiya Tokui The authors present a training-free sparse attention method that speeds up both prefill and decoding by clustering keys into large semantic blocks that can span noncontiguous positions. Each query scores block centroids and attends to the original keys and values in the blocks it selects, and queries that select the same block are processed together, so large blocks keep selection cheap and allow a larger token budget. On one NVIDIA H200, Qwen3.5-4B and Qwen3.5-35B-A3B reach whole-model speedups over dense attention of 2.3-2.4x for prefill and 1.7-2.0x for decoding at 512K tokens, and 2.9-3.0x for prefill at 1M tokens. The 35B model stays close to dense accuracy on RULER, LongBench-v2 and MRCR, while the 4B model loses accuracy mainly on multi-key retrieval but still beats a dense Qwen3.5-2B of similar latency.

arXiv:2611.04313 PDF

Anchor: Augmented Cross-Layer Recurrent State for Long-Context Retrieval in Test-Time Training

Vimal William, Yuanzhe Li, Danella Zhao, Huanrui Yang Linearization converts a pretrained Transformer into a cheaper hybrid of sliding-window attention (SWA) and a constant-size recurrent memory. Studying an SWA plus Test-Time Training (TTT) hybrid, the authors find that long-context retrieval is concentrated in a few anchor layers and that the other layers' recurrent states hold similar information under layer-specific transformations. Anchor places TTT memory only at those anchor layers and shares one cross-layer recurrent state that each layer reads through its own learned map. On Llama-3.1-8B and Mistral-7B it achieves 2.3-3.1x higher recall and 6.7-8.9 points higher MMLU accuracy than LoLCATs with a 10x smaller recurrent state, and it cuts end-to-end latency by up to 4.3x compared with softmax attention.

arXiv:2611.04327 PDF

Aligning Training and Rollout Policies for End-to-End FP4 RL on Mixture-of-Experts LLMs

Hei Yi Mak, Hoang Le, Mehran Taghian Jazi, Yunke Peng, Huan Lin, Zhengzhi Tang et al. The authors study end-to-end FP4 reinforcement learning post-training for Mixture-of-Experts (MoE) language models and find that recovery recipes that work for dense models give no gain over naive FP4 here. The cause is the router's discrete expert choices: when quantization corrections are applied only on the rollout side, the probabilities and expert selections during training and inference drift apart sharply. Their recipe, MoEResQ, applies the corrections symmetrically on both sides, restores training-inference alignment and substantially closes the accuracy gap to full precision.

arXiv:2611.04376 PDF

Near-Zero Token Embeddings Silently Stall Training in Gated Sequence Models

Xin Liu, Boyuan Sun, Yuxin Wang, Haoran Song, Mingjian Zhao, Mengyu Liu In a survey of open-weight checkpoints, 52.6% of untied gated and hybrid models contain reachable input embeddings with norms below 1e-6, a side effect of weight decay shrinking embeddings for tokens that never receive gradients. In gated sequence models such as Mamba, Gated DeltaNet and xLSTM, such a near-zero token can amplify gradients by up to 35 orders of magnitude through vanishing gates and Pre-RMSNorm, until FP32 global-norm clipping in PyTorch, Hugging Face Trainer and FSDP2 overflows to infinity and silently zeroes the batch's gradients. In continual pretraining of Nemotron-3-Nano-4B on Telugu, this discarded updates on roughly 20% of steps. The authors propose GroupClip, which clips input embeddings separately with a 64-bit or rescaled norm, for released checkpoints, and GateFloor, which adds a small offset to the SiLU gate, for pretraining.

arXiv:2611.04449 PDF

SaMA: Morpho Adaptation via Asymmetric Expansion of Kronecker Product

An Nguyen, Anh Tong Low-rank adaptation (LoRA) has flat directions in its loss landscape that slow adaptive optimizers, and its effective rank grows only linearly with the parameter budget. SaMA (Scalable Morpho Adaptation) is a parameter-efficient fine-tuning method that generalizes the Kronecker product by relaxing the shared-block constraint in its perfect-shuffle factorization. Its practical rank grows quadratically with the parameter budget, up to full rank, and its invariance set reduces to a diagonal group, which makes it easier to optimize. It outperforms strong parameter-efficient baselines on commonsense and arithmetic reasoning benchmarks while using fewer parameters.

arXiv:2611.04462 PDF

Accelerating Offloading-Based MoE Inference by Interleaving Concurrent Requests

Hongseung Yu, In Gim, Kyunghan Lee, Lin Zhong cross-listed Expert offloading lets large mixture-of-experts (MoE) language models run on local hardware, but the GPU stalls while missing experts are fetched. The common fix, predicting and prefetching experts, turns out to be counterproductive when an expert cache sits in GPU memory, because prediction gets less precise as the cache fills with the experts that are actually needed. Bevy instead runs concurrent requests ahead: a request whose experts are already loaded advances to later layers and reveals which experts it needs next, so fetching overlaps with computation without any prediction. A self-balancing scheduler batches ready requests at each layer and prioritizes lagging ones. On two local platforms, Bevy reaches up to 3.46x the decode throughput of llama.cpp and up to 1.55x that of the fastest prefetching baseline.

arXiv:2611.04794 PDF

Training Specialist LLMs on Your Phone: Forward-Only Optimization via Block-Spectral Coordinates

Liyan Tan, Yequan Zhao, Ziyue Liu, Yupeng Su, Ruijie Zhang, Zheng Zhang Adapting an on-device language model into small task specialists is hard because backpropagation needs too much memory and users rarely supply labels. MobileSpec trains with forward passes only, using random-perturbation (evolution-strategies-style) updates, and learns from automatically checked task outcomes, as in reinforcement learning with verifiable rewards (RLVR). Its core algorithm, FuRA-S, updates only the singular values of column blocks of the pretrained weights, which mimics how RLVR rotates singular vectors while largely preserving singular values, and it works on INT4-quantized models without modifying their weights. With 1–5 MB specialists that update 0.015–0.05% of parameters, it raises a phone-use agent's success rate from 6–23% to 89–100% and trains a personalized 8B model on an iPhone 16 Pro overnight.

arXiv:2611.04797 PDF

Feeling of Knowing in Large Language Models

Zichuan Fu, Xian Wu, Jingtong Gao, Wenlin Zhang, Jiaxuan Li, Binhao Wang et al. cross-listed Post-training updates change which questions a large language model (LLM) can answer, but its own judgments about what it can answer may not update with them. The authors frame this as a mismatch between knowledge and a functional feeling of knowing (FoK), which supports selective generation, deciding when to retrieve, and model routing, without implying subjective experience. They train two channels in three stages: on-policy supervised fine-tuning (OP-SFT) updates knowledge, knowledge-delta FoK training (KΔ-FT) adapts answerability judgments across intermediate knowledge states, and FoK-guided policy optimization (FGPO) trains the answer or explanation that follows. On factual question answering, across datasets and over three sequential updates, explicitly aligning FoK improves judgment quality on updated knowledge, and MMLU and ARC-Easy are used to check that general capability is preserved.

arXiv:2611.04834 PDF

FourSight: Stable FP4 RL Post-Training via Sequence Filtering and Sparse Channel Compensation

Yining Wei, Tianyuan Wu, Gu Gong, Yuechen Tao, Yinghao Yu, Xiaowen Chu et al. cross-listed Using the NVFP4 4-bit weight-and-activation format during reinforcement-learning (RL) post-training degrades rollout quality and widens the training-inference mismatch (TIM) between engines, which destabilizes training. The authors find that these failures are concentrated: badly mismatched tokens cluster in a few outlier sequences, and quantization error clusters in a few high-variance activation channels. FourSight filters out drifted trajectories from policy updates, keeps only the top-variance channels in BF16, and adds a fused SpliceGEMM kernel with separate prefill and decode paths. It sustains near-BF16 accuracy while improving step-to-accuracy by 1.76-2.60x, and its step time stays within 2.3% of pure FP4 on models up to 120B parameters.

arXiv:2611.04887 PDF

Unifying and Beyond the Theory of Pervasive Capability Degradation in LLM Post-training SFT from Feature Interference

Hang Chen, Jiaying Zhu, Wenya Wang Supervised fine-tuning (SFT) of a large language model on a narrow task often erodes its general capabilities, a phenomenon the authors call pervasive capability degradation. They explain it through feature interference under superposition: narrow SFT data disrupt the nearly orthogonal features learned in pre-training and push activations across decision boundaries, flipping predictions on general tasks. The theory groups existing mitigations into six paradigms and shows that one-hot cross-entropy alone cannot fix the problem. Their proposed AEGIS loss directly limits drift in activation margins along competing token directions, and across 3 model families and 14 SFT tasks it beats representative methods on both target-task and general capability at about 1.3x the cost of standard SFT.

arXiv:2611.05003 PDF

Going Deeper With Memories

Afshin Khadangi Byte-level language models avoid tokenizers but are expensive because they process every byte at full resolution. DeepMem is a causal multiscale architecture that uses a learned invertible subband transform to compress sequences, while a lossless carrier and detail sidebands preserve exact byte identity. Fast-weight memory blocks and iterative computation run at the compressed bottleneck, and an attention-based decoder restores byte resolution. It beats a matched byte-level Transformer on FineWeb-Edu and PG19, has the best macro average across FLORES languages, and improves PIQA. Its bottleneck refinement uses about 7,600 times fewer FLOPs than equivalent full-resolution reasoning.

arXiv:2611.05255 PDF

LoopVQ: Sub 1-bit KV Cache Quantization for Looped Transformers

Minsoo Cheong, Sungjoo Yoo Looped transformers reuse the same layers across multiple loop steps to match much larger models with fewer parameters, but each loop step stores its own keys and values, so the key-value (KV) cache grows linearly with loop count and becomes the memory bottleneck. LoopVQ exploits the observation that this cache is highly redundant along the loop axis while loop steps differ in quantization sensitivity, jointly vector-quantizing channels from several loop steps and fine-tuning the codebooks to match full-precision outputs. At 0.5 bits per element it beats KV-cache vector quantization methods built for standard LLMs on general, long chain-of-thought, and long-context retrieval tasks with Ouro and Huginn, and composed with TaSQ it retains much of BF16 accuracy at 0.25 bits. Implemented in SGLang, it nearly halves decode latency at a 32K-token context and raises peak serving throughput by up to 4.3x and 6.1x on the 1.4B and 2.6B Ouro models.

arXiv:2611.05304 PDF

A Circuit for In-Context Learning

Andrew Gao, William Gao cross-listed In-context learning with arbitrary labels forces a model to acquire a mapping it cannot have memorized, so whatever learning happens must happen inside one forward pass; prior work identified candidate components but not how they compose. Studying few-shot sentence classification in three transformers of 3-4B parameters, the authors describe the full mechanism as a key-value memory over label-token positions: middle-layer attention heads write the class of the preceding demonstration into each label token as a key, upper-layer heads match the query's class against the stored keys and copy out the matching label, an early-layer signal encoding whether any label has repeated gates that readout, and a separate head group suppresses the most recently seen label. Every step is established causally by activation patching, head ablations against random-head baselines, and steering with the model's own signals, and the named heads alone reproduce the behaviour. The same circuit appears from 0.6B to 8B within one family, survives instruction tuning, transfers to semantic concepts and meaningful labels, and regrows in a transformer trained from scratch.

arXiv:2611.05344 PDF

Aligning Language Models with General Preferences: An Empirical Study of Direct Nash Optimization

Austin Tsai, Saurav Nanda, Ganapathy Parthasarathy, Bhuvnesh Kumar, Parivesh Choudhary Direct Nash Optimization (DNO) aligns a language model to a general preference function by approximating the Nash equilibrium of a two-player preference game, and its practical variant iterates on-policy sampling, LLM-based rating of samples against a teacher's gold response, large-margin preference-pair selection, and a DPO-style contrastive update. An independent re-implementation on Llama 3.1 8B Instruct and Mistral 7B Instruct v0.1 lifts the length-controlled win rate against GPT-4o on AlpacaEval 2.0 from 31.7% to 45.3% and from 6.8% to 18.4% over three iterations, with the tuned 8B model scoring above Llama 3.1 70B Instruct on every reported win rate. Replacing the proprietary annotator with a same-family open-weight model retains 62% and 82% of the gain, so the rating step does not require a closed model; ablations show removing gold responses, the data look-back, or non-teacher-preferred pairs costs 6 to 10 points, while sample count and margin settings move results by at most about 2 points. The authors report pipeline cost, where the practical algorithm departs from the theory, and caution that all runs are single and GPT-4o serves as both teacher and judge.

arXiv:2611.05373 PDF

Same Origin, Different Verdicts: Text Production Conditions Confound Machine-Generated Text Detection in Russian

Stanislav Kirichenko Detectors of machine-generated text treat origin as something visible in the text. This study tests how much the conditions under which a text was written move detector scores when origin is held fixed. On a Russian corpus of 1,882 documents, the authors compared pairs written by the same model for the same task that differ only in how the task was worded. They scored machine-likeness with four procedures: an interpretable index, a trained classifier, a zero-shot metric, and a judge model. Changing the task mode shifted all four scores, but three procedures shifted negatively while the judge model shifted positively, and no confidence interval covered zero. A baseline using only the source identifier reached a median AUROC of 1.0, and removing emphasis markers alone flipped a confident machine verdict to human. The authors conclude that detector scores reflect corpus and preprocessing choices as much as authorship.

arXiv:2611.05385 PDF

SRRM: A Retrieval-Reconstruction Benchmark for Evaluating Long-Context Memory in Small Language Models

Christopher Alberto Hamonangan Small language models (SLMs) run on consumer hardware are mostly tested on retrieval, which may hide how well they keep the full content of long contexts. The authors introduce SRRM (Simple Retrieval-Reconstruction Memory), a lightweight benchmark that scores both targeted retrieval and reconstruction of all the given information. They also propose the Memory Preservation Ratio (MPR) to relate the two scores. Testing Qwen2.5 at 0.5B, 1.5B, and 3B parameters on 10 to 100 factual statements, they found that reconstruction degraded much faster than retrieval as context grew. Notably, Qwen2.5-1.5B outperformed Qwen2.5-3B on every measure (MPR 0.53 vs. 0.26), suggesting that more parameters do not guarantee better long-context memory.

arXiv:2611.05386 PDF

AGORA: Reliability-Aware Aggregation of LLM Judges via Item-Response Theory

Junhyung Ahn Combining several LLM judges is the usual way to offset any single judge's bias, but simple averaging weights weak judges as heavily as strong ones, and confusion-matrix models such as Dawid-Skene need more labeled items than most validation sets contain. AGORA (Aggregation Of Reliability-weighted Assessments) uses item response theory: it places every item on one shared latent quality scale and gives each judge a sensitivity parameter and ordered rating thresholds. This lets it learn judge reliability from all items jointly, without human labels. Across three benchmarks it matches or beats Dawid-Skene, and its advantage grows as the evaluation set shrinks to a few hundred items. A label-free statistic also predicts when reliability weighting will help.

arXiv:2611.05556 PDF

Verified but Vacuous: Three Failure Surfaces of LLM-Driven Verified Code Generation

Aditya Vempaty, Himaneesh Sompalle, Manan Tayal, Mowafak Allaham, Prasenjit Dey, Ashish Jagmohan et al. LLM-generated verified code is usually scored with a single verification number. The authors argue that this number hides three separate properties: whether the verifier accepts the code, whether the proof obligation actually binds to the specification, and whether the specification captures the informal requirement. Reproducing a Dafny-as-intermediate-representation pipeline with Claude Opus 4.7 on HumanEval, they find 99.1% verifier acceptance but only 1.9% of certificates binding to a spec-defined predicate or function. With their own VCG pipeline on Rust, where spec binding holds by construction, 11.48% (MBPP-Rust) and 14.82% (VeriContest LC slice) of structurally verified samples still fail hidden tests.

arXiv:2611.05590 PDF

RAGCompass: Generalizable Retriever Routing via Reranking

Valentin Six, Mufei Li, Dongqi Fu, Yinglong Xia, Hong Li, Pan Li No single retriever performs best for every query in retrieval-augmented generation, and existing routers predict from query features alone, so they cannot handle a retriever or corpus they were not trained on. RAGCompass routes by reading the passage sets each retriever returns. It uses a reranker contrastively fine-tuned on the answers each evidence set produces, plus a small comparison head, and neither part has retriever-specific parameters. It improves answer F1 by 28% relative to the best fixed retriever and beats fusion baselines, three query-only routers, and a prompted 8B model. Most of the gain holds with an unseen corpus, retriever, or reader, while query-only routers drop to the best fixed retriever or below.

arXiv:2611.05596 PDF

Hybrid Spark: Adaptive Speculative Decoding That Lets the Content Choose the Drafter

Ekagra Ranjan, Bharat Venkitesh Speculative decoding speeds up large language model (LLM) serving. Model-based drafters work on any content, while model-free drafters such as n-gram lookup excel at repetitive outputs like code editing, but adaptive verification has so far supported only one model-based drafter. The proposed framework runs both drafter types on every step and gives each request the better chain, entirely on the GPU. It does this by pricing each verified draft token at the throughput that slot would earn elsewhere in the batch, using acceptance estimates for the n-gram drafter that are calibrated once on general text. Implemented in vLLM with DSpark and an n-gram drafter, it reaches 1.10–1.62x the output throughput of DSpark on code editing at batch size 4 while staying at 0.97–1.01x on chat, so one deployment can serve both workloads.

arXiv:2611.05611 PDF

Pairing Expert Critique with LLM-Driven Refinement: An Iterative Approach for Curating Quality AI Evaluation Datasets

Honson Ling, Kidus Admassu, Stephanie Rosenthal cross-listed Having experts write domain-specific evaluation data from scratch is slow and expensive. SME-C (Subject Matter Expert Critique) instead collects lightweight expert critiques of candidate responses and uses an LLM to refine both the responses and the evaluation rubrics over several rounds. In financial business briefs, the authors gathered 3,853 critiques from 411 finance professionals. They compared refinement guided by written critiques with refinement inferred only from preference votes. In a blind survey of 100 more professionals, both strategies significantly improved on the experts' top-ranked baseline responses, with critique-based refinement adding domain grounding and vote-based refinement better at removing irrelevant content.

arXiv:2611.05749 PDF

Scheduling Data to Compute: Hierarchical Data Pipelining for LLM Training on AMD GPUs

Gengyu Rao, Ghadeer Ahmed H Alabandi, Su Ann Chong, Thomas Huber, Kailash Gogineni, Subhajit Dutta Chowdhury et al. cross-listed In large-scale LLM training on AMD GPUs, compute units often sit idle waiting for data, and this work treats that waiting time as the main bottleneck. The authors build kernels and framework support around AMD's hardware for explicit data movement: the Tensor Data Mover (TDM) for loading data from high-bandwidth memory onto the chip, transpose-load instructions that rearrange data into the needed layout as it is loaded, and System DMA and Remote DMA for moving data between GPUs. TDM gives an average 1.88x kernel-level speedup, transpose-load gives 1.19x, and SDMA and RDMA give 1.16x and 1.33x speedups for computation overlapped with communication. The work is integrated into the open-source Primus and Primus-Turbo libraries.

arXiv:2611.05751 PDF

Matching the Training-Time Weight Rule in Inference-Time Expert Deactivation

Hao Cheng (Zhejiang Gongshang University), Zhiyu Yang (Chang'an University), Bo Ma (Zhejiang Gongshang University), Hanzhang Jia (University of Sussex) Turning off experts in mixture-of-experts models at inference time changes both token routing and expert weights, and results depend on how this is implemented. The authors compare removal, freezing, renormalizing the remaining experts, and unnormalized re-routing on five instruction-tuned checkpoints plus a base OLMoE, and build a weight-sum model that ties the outcome to the weighting rule used in training. With half the experts disabled, perplexity rises 1.4 to 73 times over the intact model when the training rule is kept, but 14 to 28,740 times when the rule is mismatched. On OLMoE, GSM8K accuracy falls to zero under all four implementations, and the penalty from extra renormalization grows as fewer experts are active.

arXiv:2611.05804 PDF

EchoKV: Self-Corrective KV Cache Eviction via Iterative Query Feedback

JianKun Zou, Tongtong Yuan Query-driven KV cache eviction methods such as LAQ build a temporary workspace to generate lookahead queries, so an early error in that workspace can distort later queries and reinforce bad eviction decisions. EchoKV starts from a bootstrap workspace built from proposals at future decode positions, then uses early decode-generated queries to repair the workspace before the remaining lookahead steps. It uses the same number of workspace forward passes as LAQ, plus one tensor re-scoring operation. On LongBench with Mistral-7B-Instruct-v0.2 and Qwen2.5-7B-Instruct, it improves notably on tasks that need dispersed evidence, such as multi-hop cross-document QA, while staying comparable on the other subsets.

arXiv:2611.05903 PDF

On the Design of Hyper-Connections: Simplification, Stability, and Scalability

Xiangdong Zhang, Xiaohan Qin, Sunan Zou, Tuo Dai, Bizhao Shi, Zhuo Xia et al. cross-listed Hyper-Connections, an alternative to standard residual connections used in some frontier language models, add training overhead, and it has been unclear which of their components matter. The authors find that learned read/write gating without mixing, or unconstrained mixing with fixed gates, each performs about as well as mHC, and that repeated access to the expanded residual state is the main source of overhead. The resulting SimpleHC removes residual mixing and shares memory access across layers. It adds only 2.4% training time over a plain residual baseline, matches mHC on MoE models from 15B to 30B, and shows smoother gradients and stable training under stress.

arXiv:2611.05984 PDF

Adapting Qwen3 to Tajik from 0.6B to 8B: Lower Text Loss Did Not Mean Better Understanding

Tohir Saidzoda, Madina Abdusalomzoda, Nasimjon Rajabov, Akhrorkhon Ikromov, Komron Giyosiev, Abubakr Abdullozoda et al. The authors adapt Qwen3 models from 0.6B to 8B parameters to Tajik and find that small LoRA adapters lower Tajik text loss at every size without improving reading comprehension on Belebele. At 8B, continued pretraining on about 375 million tokens of raw Tajik text, followed by supervised fine-tuning, yields a model 11.4 points above Qwen3-8B on Belebele with English MMLU down only 1.2 points. The same heavy recipe applied to the 0.6B base model brought its text loss close to the 8B model's while its multiple-choice answers stayed near chance. The report covers 15 trained models and documents what went wrong, including lost training records and a 4-bit release that contains no transformer layers.

arXiv:2611.06062 PDF

ProcessPrefBench: Evaluating Cross-Task Transfer of Individual Process Preferences from Naturalistic Human-AI Interaction

Yiwen Yang, Guanming Liu Users reveal individual process preferences, such as standards for evidence, depth, and delivery, through corrections and requests in conversation. ProcessPrefBench tests whether a preference revealed in one dialogue carries over to a later task that does not restate it. The benchmark has 66 source-target cases drawn from three months of real dialogues by two knowledge workers. Compared with target-only prompting, full-history and distilled-preference prompts raise preference realization by about 0.2 points on a 0-2 scale, with no reliable benefit from distillation. They also nearly double fabricated-evidence failures, from 5.8% to over 10%, which exposes a tension between personalization and factual reliability.

arXiv:2611.06065 PDF

Accelerating Long-Context Decoding in Large-Scale Hybrid LLMs Without Retraining

Jo\~ao Eduardo Batista, Emil Vatai, Mohamed Wahib Hybrid language models replace most full-attention layers with linear attention, but the few full-attention layers that remain still dominate decoding cost at long context lengths. Without any retraining or weight changes, the authors replace those layers with combined sparse attention: each layer attends to a sliding window of recent tokens, a pinned prefix, and a fixed budget of the most relevant tokens, with window size and budget chosen per layer by a multi-objective search. Across six hybrid models from 4B to 397B parameters, the chosen configurations keep baseline accuracy on 21 long-context datasets. Decoding at 256k tokens becomes up to 2.22x faster on Qwen3.5-4B, 2.23x faster on Qwen3.5-122B-A10B, and 2.06x faster on Ornith-1.0-397B.

arXiv:2611.06133 PDF

AssameseLM-E2B: An Open-Weight Assamese Language Model for Efficient Consumer-Grade Deployment

Pragyan Phukan Large language models degrade on low-resource languages such as Assamese because of inefficient tokenization, unstable generation, and interference from related languages. AssameseLM-E2B is a 2-billion-parameter model built on Gemma-4 with an expanded byte-level BPE vocabulary, adapted through parameter-efficient continual pre-training with hybrid LoRA and then supervised fine-tuning for instruction following and retrieval-augmented generation (RAG). It reaches 0.7488 bits per byte on FLORES-200, an overall RAG score of 96.7%, and 100% refusal accuracy on adversarial unanswerable prompts. General-purpose baselines, by contrast, degrade on Romanized Assamese and drift into Bengali script, and the model is sized for deployment on consumer hardware.

arXiv:2611.06170 PDF

A.S.E 2.0: Understanding How Repository Context Impacts the Security of LLM-Generated Code

Jiongchi Yu, Keke Lian, Junqing Wang, Zhikang Fan, Zhaoxuan Li, Yongheng Huang et al. cross-listed LLM coding assistants and agents gather repository context before writing code, but that context is built for functional correctness, and its effect on security has not been measured. The authors build CASE-Bench, 145 Dockerized repository-level tasks drawn from real CVEs across 51 projects, each with functional tests, a validated proof-of-concept exploit, and annotations of the code that enforces or undermines security. Across 13 context-construction methods and six LLMs, functionality-oriented context raises functional pass rates by up to 8.2 percentage points but secure pass rates by at most 3.7. Withholding a safeguard or inserting a single insecure example measurably hurts security, and coding agents look at the relevant safeguard in only 27-36% of their trajectories.

arXiv:2611.06294 PDF

Presentation Sensitivity in Typed Decision Models for Subjective Text Evaluation

Jatin Agrawal, Radhika Mamidi Typed decision models, which return structured choices, scores, and probabilities, are used as cheap automated evaluators, but it is unclear how much their verdicts depend on how the inputs are presented. This controlled audit collects 702,000 judgments from Jev, Clef Flash, and Perplexity Decider on 9,000 human-annotated humor, helpfulness, and summarization items. The interventions change positions, identifiers, authorship claims, endorsements, framing, and rubrics while leaving the evaluated texts unchanged. The effects are large: reversing reader endorsements changes Perplexity's chosen humor text on 40.40% of pairs, rejection framing cuts Clef's agreement with human summary preferences by 23.36 points, and Jev shows a second-position advantage of up to 6.11 points. The authors conclude that repeatability, robustness to presentation, and alignment with humans are separate properties that each need validation.

arXiv:2611.06384 PDF

Where Should a KV-Cache Repair Budget Go? Distractor Structure and Contextual Compatibility in Non-Prefix Reuse

Yuxing Peng, Zhiqing Tang, Weijia Jia Non-prefix KV-cache reuse encodes retrieved chunks independently and concatenates their caches, which saves computation but loses cross-chunk context. Across three 7B-8B models and four datasets, reuse loses 19-100% of the context-dependent questions that full prefill answers, and the losses are worse when the surrounding chunks resemble the answer-bearing one. Because the damage is spread across those distractor chunks, a small repair budget is better spent on the target chunk: recomputing it against the reused cache beats inserting its full-prefill KV in all seven settings tested. Attention-based selection rarely finds the target, but retrieving the top chunk, moving it next to the question, and recomputing it beats InfoFlow-style reordering at a 5% budget in six of nine settings. On dialogue memory, chunk recomputation never significantly beats token-level selectors.

arXiv:2611.06397 PDF

Mind Control of LLMs via Closed-Loop Stimulation: Out-of-Distribution Robustness

Kasra Fallah, Rudramani Singha, Hannah Ghanei, Paul Sajda, James Anderson, Erfan Zabeh cross-listed Borrowing from closed-loop neurostimulation, the authors treat LLM steering as a feedback-control problem in which the model's internal state determines how it is perturbed, rather than applying a fixed intervention. To account for the gap between tractable linear dynamical models and the actual nonlinear transformer, they design an H-infinity robust controller that limits sensitivity to disturbances and linearization errors. Across steering benchmarks the controller beats existing baselines, with especially large gains under out-of-distribution conditions and varying sequence lengths. They also introduce robust steerability, a pre-deployment measure of how much dynamics mismatch steering can tolerate, which shows task-specific structure and varies systematically with model scale.

arXiv:2611.06418 PDF

D3D: Decoupled Diffusion Drafting for Asynchronous Speculative Decoding

Baoyou Chen, Ruiqiao Mei, Guancong Lin, Tao Song, Weihao Yuan, Qun Liu et al. Diffusion drafters for speculative decoding propose a whole block of tokens in one pass, but each new block waits for the target model's updated hidden states, so drafting and verification cannot overlap. D3D is a standalone diffusion drafter that combines blockwise denoising with auxiliary autoregressive supervision, achieving long accepted prefixes without needing target hidden states. This lets it draft the next blocks during verification in one packed pass, enabling asynchronous speculative decoding, and layer pruning cuts drafting cost while keeping over 99% of acceptance length. Across seven benchmarks it averages 5.00x and 5.71x decoding speedups with Qwen3-8B and Qwen3-32B targets, against 4.03x and 3.67x for DFlash, and reaches 7.48x on GSM8K.

arXiv:2611.06425 PDF

ArcRAG: Agentic Compositional Retrieval over Knowledge Graphs

Jyotin Goel, Abhinav Arun, Tejas Prakash Agarwal, Bhaskarjit Sarmah, Stefano Pasquali Graph-based retrieval-augmented generation (GraphRAG) systems usually design retrieval around the knowledge graph their own pipeline builds, and rarely test whether retrieval still works when the graph is built differently. ArcRAG infers a query-local graph interface at inference time (the node labels, relation descriptors, and property keys reachable from the entities a question names), then plans multi-hop traversals and executes and repairs them as Cypher queries. With its configuration fixed, it is evaluated on MuSiQue, HotpotQA, and TriviaQA, each built both as a schema-guided graph and as an open information extraction graph. The results show that retrieval need not be co-designed with graph construction, although the construction still affects absolute accuracy.

arXiv:2611.06481 PDF

Model Merging Does Not Require Uniform LoRA Ranks

Zirong Xie, Zainab Afolabi, Nikita Kozodoi, Jack Butler cross-listed When merging separately fine-tuned experts, the usual practice is to train every LoRA adapter at the same rank, even though each domain has its own best rank. Across 36 experts in math, instruction following, and multilingual tasks at three model scales, merged with six methods, giving each expert its own rank costs averaging and utility-based selection nothing and gains up to 1.8 percentage points, while conflict-resolution methods such as TIES are inconsistent and can lose up to 4 points. The cause is that magnitude-based component selection discards the instruction adapter, which has the smallest singular values, so IFEval drops by 16 to 25 points. Selecting components by their effect on loss keeps all experts and leaves accuracy nearly flat across an 8x range of ranks.

arXiv:2611.06486 PDF

NaijaFacts: A Framework for Disentangling Language from Knowledge in Facts about Nigeria

Zainab Afolabi, Nikita Kozodoi, Jack Butler, Babs Khalidson Multilingual benchmarks for African languages usually translate generic questions, which tests language ability rather than whether models know facts about Africa. NaijaFacts poses factual and statistical questions about Nigeria, grounded in official statistics, in English, Nigerian Pidgin, Yoruba, Igbo, and Hausa, and pairs them with similar UK and US questions to separate a language effect from a knowledge effect. Across eight models, the knowledge gap dominates: English-query accuracy is about 50% for US/UK questions but 20% for Nigeria, and no model exceeds 31% on Nigerian statistics. Extending the evaluation to all 54 African countries shows the same pattern, with high accuracy on culture and politics but only 26% on statistics.

arXiv:2611.06504 PDF

Reasoning over Non-Parametric Memory in Multilingual LLMs: A Cognitive-Based Analysis of Relational Knowledge

Zohreh Abbasi Liasi, Batoul Hammoud, Anas Madkoor, Ehsaneddin Asgari Retrieval-augmented generation (RAG) assumes models reason over relations given in context rather than relying on memorized associations, and Rel-Bloom tests this directly with 450 multiple-choice questions built from 30 Wikidata relations, posed at five Bloom-taxonomy levels and under factual, counterfactual, and anonymized conditions. Across 43 models, six languages, and 226,800 predictions, open-weight models dropped sharply from Apply to Analyze questions, and even Gemini 3.1 Pro, at 99.5% on other levels, scored only 60.7% on Evaluate items. Difficulty rankings held across languages, and counterfactual contexts were generally harder than factual ones.

arXiv:2611.06715 PDF

LLMs Don't Know What They Know: Benchmarking Temporal Self-Knowledge for Financial Backtesting

Lingfei Qian, Yan Wang, Xueqing Peng, Yupeng Cao, Yuechen Jiang, Zhuohan Xie et al. LLM backtests in finance are only valid if the model did not already know the outcomes, and practitioners rely on stated knowledge cutoffs, the model's own self-reports, or instructions to answer as of a past date. FinDendro tests these safeguards with 2,226 dated U.S. facts from 1999 to 2026 (macroeconomic series, month-end stock prices, and Federal Reserve decisions), each asked in several forms, with controls that keep chance accuracy near 0.27. For most frontier models, self-reports understate what they know: GPT-5 answers 76% of controlled facts correctly but claims to know only 38%, and answers 61% of the facts it says it does not know. Claims of knowledge can extend up to seven quarters past the measured knowledge boundary, and declaring a past date removes most claims of prior knowledge without lowering accuracy, so none of these safeguards can certify a backtest as leak-free.

arXiv:2611.06808 PDF

LooM: Loop-Invariant Memory for Causal Encoder-Decoder Language Models

Lele Yang, Muxi Diao, Zhonghao Yan, Jiankang Chen, Bin Wen, Tingting Gao et al. Looped Transformers gain effective depth by re-running shared layers, but their memory grows with the loop count, compute can only change in whole-block steps, and the loop count is hard to change after training. LooM (Loop-invariant Memory) loops only the decoder layers after the first in a causal encoder-decoder, and these layers read a global memory written once per token and publish a token's key-value (KV) entry only on their first visit, so the cache does not grow with the loop count. The layers loop independently, giving any budget from 3 to 21 layer executions per token, and a single model trained on randomly sampled budgets covers all of them without a router. With enough loops, LooM beats a single-pass baseline trained on the same data on language modeling, zero-shot accuracy, and long-context perplexity at the same cache size.

arXiv:2611.06835 PDF

Context--Answer Metamodels: Forecasting LLM Activations Before the First Answer Token

Tommy He, Thomas Jiralerspong cross-listed Under stochastic decoding, one prompt can lead a language model to many different answers, and characterizing that spread normally requires many expensive rollouts. The authors train context-answer metamodels that map the activations produced while reading the context to point or distribution forecasts of the mean-pooled activations of the eventual answers. The metamodels forecast answer activations with high fidelity and capture rollout variability; they improve as training grows to 500,000 prompts and generalize zero-shot to out-of-domain datasets. Their forecasts also retain signal that predicts behavior, and context activations from Qwen2.5 can predict answer activations from Qwen3.5 and Gemma 4.

arXiv:2611.06837 PDF

Balancing hierarchical representation and quantization: A performance trade-off in text embeddings

Kai Ponel, Marco Polignano, Erasmo Purificato, Aylin Julia Temizkan, Marco Levantesi, Giovanni Semeraro et al. Text embeddings can be made cheaper in two separate ways: hierarchical models that support shorter, variable-length vectors, and quantization, which stores each number at lower precision. This benchmark study measures how the two interact. It tests four transformer embedding models on 38 tasks from the Massive Text Embedding Benchmark (MTEB), with 168 configurations per task, and checks significance with a Friedman test and Nemenyi post-hoc analysis. int8 quantization is statistically indistinguishable from float32 on pair classification and re-ranking and loses at most 1.5 points overall, while binary quantization is significantly worse in every task category. Combining an 8x cut in dimensions with int8 gives 32x smaller per-vector storage at 94.4% of baseline effectiveness, though retrieval degrades much faster than the other task categories.

arXiv:2611.06859 PDF

18 more specialized papers

Other 61

One Mistake, Three Layers: Persistence Rules Under Depth Correlation

Maryam Rahimimovassagh, Michael Brown, Sanjeda Sara Jennifer, Niloofar Yousefi Layer-by-layer probing of frozen representations commonly reports the first layer where a signal becomes decodable, often by requiring a score to stay above a threshold for three consecutive layers. On synthetic curves, correlation between neighboring layers makes this rule report the onset nineteen stages too early, so the authors propose an estimator that calibrates its error rate using label permutations held fixed across depth, per-layer thresholds and a calibrated persistence requirement. Applied to pocket information in a frozen protein co-folding model across 1,820 targets, the three-stage rule fires on 12 to 17% of null trajectories against a nominal 5%, which reproduces the failure on real data. Separately, a widely used public co-folding checkpoint silently loads with 128 trunk parameters missing, so representations extracted from it are partly random noise.

arXiv:2611.03968 PDF

Deep Learning Compilers: Architecture, Optimization, and Empirical Characterization

Jinghao Wen, Ruixuan Wang, Xun Jiao The authors survey deep learning compilers, organizing them into frontend, middle-end, and backend stages and reviewing representative systems and optimization techniques. They also measure TorchInductor in PyTorch on an RTX A6000. Speedups depend heavily on the workload, reaching 8.21x for fusion-friendly graphs and staying small for others. Compilation overhead can take thousands of runs to pay back, and dynamic-shape compilation cuts shape-transition costs by about 70% at a modest cost in steady-state speed.

arXiv:2611.04273 PDF

What Is the Influence of Task Order in Continual Learning? An Empirical Study and Order-Aware Evaluation for Motion Forecasting

Zirui Li, Yunlong Lin, Haohan Yang, Xiaocong Zhao, Chen Lv Continual learning benchmarks report aggregate scores that implicitly average over a task-order distribution that is rarely stated. The authors frame algorithm comparison as an expectation over a declared order distribution and test it on motion forecasting with four road-layout tasks from the INTERACTION dataset, covering all 24 orders, four algorithms and three replicates (288 runs). Task order explains 65% to 91% of the variation in current-task error and backward transfer, almost all of it from which task comes first, because the smallest task cannot be learned from a cold start. After the first stage, the order effect is about as large as ordinary training variability, and the leading algorithms usually differ by less than one replicate standard deviation. The authors recommend that benchmarks declare their order distribution and report order-resolved and starting-task-resolved errors.

arXiv:2611.04309 PDF

Koopman Generative Operators for Efficient Probabilistic Time-Series Forecasting

Raz Marshanski, Liran Nochumsohn, Mayank Jauhari, Boris N. Oreshkin, Omri Azencot Probabilistic time-series forecasters usually trade off between structured latent dynamics with limited generative flexibility and expressive diffusion or flow-matching models that need slow iterative sampling. The Koopman Generative Operator (KGO) combines three parts: a Koopman patch embedding for consistent latent trajectory extrapolation, Koopman flow matching that generates samples in a single step through a closed-form matrix exponential, and an adaptive gate that sets uncertainty per variable and per forecast horizon. It achieves state-of-the-art accuracy on 13 of 17 ProbTS benchmarks by continuous ranked probability score (CRPS) and 16 of 17 by normalized mean absolute error (NMAE), while cutting inference time at least 25-fold compared with iterative generative models.

arXiv:2611.04760 PDF

ZEMA: confidence-calibrated signed descent with only one state

Ali Dadsetan, Domenic Rosati, Frank Rudzicz Building on the view of Adam as sign descent scaled by gradient uncertainty, ZEMA (Z-score Evidence Momentum Adaptation) computes a per-coordinate z-score from per-example gradients in each batch, converts it into the probability that the gradient's sign is correct, and smooths the result with a single exponential moving average. Because the probability is computed before smoothing, ZEMA needs only one state per parameter instead of Adam's two, at a 4.6x increase in per-step time, and the per-example variance can be estimated from a small fraction of the batch. In fine-tuning and pretraining experiments it is competitive with Adam, reduces the variability of validation loss, and in some settings tracks the sign of the full-dataset gradient more consistently.

arXiv:2611.04987 PDF

Beyond Foundation Models: PFN Priors Improve Tabular Hyperparameter Optimization

Artem Chistyakov, Renat Sergazinov, Artem Babenko Synthetic priors from Prior-data Fitted Networks (PFNs) have so far been used only to pretrain tabular foundation models. PriorTune uses them instead to meta-tune hyperparameters for conventional models that are trained from scratch on each dataset. Offline, it fits a pool of configurations on datasets sampled from a published PFN prior and greedily selects the subset with the highest winrate. Online, a new dataset is fitted only with the selected configurations, at no extra cost. On the full BeyondArena benchmark, LightGBM with TabPrep features and PriorTune configurations ranks above TabPFN-3 and RealMLP while being much cheaper at inference. The authors also observe, and explain theoretically, that scaling the number of configurations and the number of synthetic datasets improves results by equal amounts.

arXiv:2611.05141 PDF

Unlocking Turing Tensor Cores: A Measured Throughput Account of From-Scratch Pre-training on a Single 22G RTX 2080 Ti

Dezhou Shen cross-listed The RTX 2080 Ti has full Tensor Cores, but the driver throttles FP16 matrix multiplication. A kernel module called turing-unlock removes this limit, and published measurements report cuBLAS FP16 throughput rising from 55.7 to 91.4 TFLOPS. The authors measure what this means for real from-scratch pre-training on a single 22 GB card, across five sizes of Chinese encoder-decoder models. The gain grows with matrix size: +11.8% throughput (9% less wall time) for the 75.7M-parameter tier, +4.7% for the next size down, and at most +0.3% for the three smallest. Peak memory was bit-identical across all tiers, and final quality metrics were bit-identical or matched within 10^-3, so the unlock speeds up training without changing its dynamics.

arXiv:2611.05419 PDF

Just the Facts: How Dialogues with AI Reduce Conspiracy Beliefs

Thomas H. Costello, Gordon Pennycook, David G. Rand cross-listed Two experiments with 2,193 U.S. adults removed possible causes one at a time to find out why personalized dialogues with AI reduce conspiracy beliefs. The effect held when the AI disclosed its persuasive goal, when the exchange was framed as a debate, when responses were halved in length, and when the dialogue was replaced by a single tailored message. Only instructing the AI to persuade without factual evidence and arguments substantially weakened the effect. Content analyses linked larger belief reductions to factual counterevidence rather than rapport, storytelling, or empathy, which suggests that humans could be similarly effective if they could supply comparable evidence.

arXiv:2611.05572 PDF

VIPBench: A Human-Aligned Benchmark for Voice Identity Perception in the Age of Voice Cloning

Xudong Tang, Matthew Groh cross-listed Speaker-recognition benchmarks test whether a model recovers who produced a recording, not whether humans hear two voices as the same person. The Voice Identity Perception benchmark (VIPBench) collects 124,876 same-or-different judgments from 1,290 listeners on 9,800 recording pairs from 100 public figures, covering real recordings, text-to-speech voice clones, and morphs between two speakers. It defines four tasks that measure how well speaker models align with these judgments. In 41% of responses, listeners judged a clone and a real recording of the same speaker to be different people, and no tested model reaches the noise ceiling; the best of eleven explains 61% of the explainable variance. Models trained with speaker labels align better than self-supervised ones, but one model with a 12% speaker-label error rate aligns as well as models with error rates below 2%.

arXiv:2611.05638 PDF

Familiar letters, not related languages, let unseen languages into a shared space of meaning

Yuliang Liu, Xinyu Wang, Xiang Bai Hundreds of languages have written text but almost no translations, and the paper asks whether a model learning such a language from untranslated text benefits from a related language or from a familiar script. GlyphMind is a tokenizer-free vision transformer that reads text as pixels; it was pretrained to align translations across 27 writing systems. Adapted on untranslated Lao, it retrieved English translations seven times as often as the strongest simple baseline. Controlled script swaps show that familiar letters, not linguistic relatedness, determine whether an unseen language enters the shared meaning space, which suggests transliteration into familiar scripts as a route for low-resource languages.

arXiv:2611.06006 PDF

Today's Work, Tomorrow's Workforce: Incentivizing Learning in AI-Assisted Work

Moritz Diener, Daniel Hendriks, Philipp Spitzer, Gerhard Satzger, Joshua Holstein, Wei Sun cross-listed AI assistance lets workers accept AI answers instead of solving tasks themselves, which saves effort but can displace the practice needed to build and keep skills, including the skills needed to oversee the AI. Extending an accept-or-solve model to sequences of tasks where skill evolves, the authors show that myopic novices can get stuck at low skill and experts can lose skill even while AI-assisted accuracy stays high, whereas strategic workers sometimes solve tasks themselves despite a more accurate AI. They characterize the payments for correct independent work needed to sustain skills, show that independent work becomes self-sustaining once skill is high enough, and find that a higher skill target can sometimes lower long-run incentive costs.

arXiv:2611.06346 PDF

50 more specialized papers

Agents 58

Do agents need to sleep?

Ido Haber, Giulio Tononi cross-listed LLM-based agents that persist across sessions and keep accumulating memory show symptoms of unchecked plasticity: semantic drift, worse retrieval precision, propagation of stored errors, and memory saturation. Drawing on the selectionist "ringing, resonance and renormalization" (RRR) theory of sleep, the position paper maps each of its three phases onto an operation that digital memory architectures can perform. The authors note that engineers building stateful agents are independently rediscovering the problems sleep evolved to solve, namely what to forget and when. They make the falsifiable prediction that selectionist memory consolidation will outperform instructionist alternatives as memory stores grow large.

arXiv:2611.02216 PDF

WASP-Mem: Let Agent Organize Procedural Memory. Human Spend Effort on Diagnosis

Peizhi Niu, Ge Shi, Vaibhav Gorde, Sarthak Ghosh, Christos Faloutsos, Min Cheng et al. cross-listed LLM agents struggle with enterprise procedures because fixed, pre-built tooling cannot capture proprietary policies or recovery paths that only appear during execution. WASP-Mem (Wiki-based Agent-Synthesized Procedural Memory) distills execution traces and feedback into persistent, interlinked wiki pages that the agent maintains, which lets it keep learning without training. On TheAgentCompany, it raises the checkpoint pass rate by 4.6 to 6.1% over the strongest baseline while using 57 to 61% less storage. A controlled study finds that human-imposed schemas and deduplication rules hurt performance, while a single round of human diagnosis of failures beats four rounds of self-judgment by 7.0 to 10.9%.

arXiv:2611.04003 PDF

Computer-Use Agents Gain Little Benefit from a World Model When Conditioned on Tutorials

Anh Nguyen, Tianyi Xu, Phat Tran, Harry Yu, Huazheng Wang, Stefan Lee Recent computer-use agents (CUAs) combine retrieved tutorials with look-ahead planning, in which a world model predicts future states and a multimodal LLM judge picks actions, but each component's contribution had not been measured separately. The authors build two tutorial datasets, OSWorld-Tutorial and ScienceBoard-Tutorial, and a policy-judge framework, RAG-PolicyJudge (RAG-PJ), then compare it with and without world-model rollouts, including an oracle that returns the true future screenshots. When tutorials are present, the agent gains little from rollouts, even perfect visual ones, which suggests earlier gains came mainly from understanding the tutorials and from the judge's action selection. Error analysis places the bottleneck in the judge's ability to reason over future states and recognize when a task is complete, not in how accurate the rollouts are.

arXiv:2611.04073 PDF

ModelBlaster: Closing the Performance Gap through Learning-Augmented Agents Across Models, Kernels and Hardware Backends

Shengjun Kris Dong, Dima Nikiforov, Sahil Modi, Sana Damani, Edward C Lin, Siva Kumar Sastry Hari et al. ModelBlaster is an LLM-agent framework that optimizes both individual kernels and whole models across different accelerators. It uses Memory-Augmented In-Context Reinforcement Learning (MAIC-RL): a persistent knowledge base, updated during inference by Actor-Critic Textual Gradient Descent (AC-TGD), records which actions worked in which performance states and lets that experience carry across kernels, models, and hardware. The same loop targets CUDA on NVIDIA GPUs, OpenCL on Qualcomm GPUs, and RVV intrinsics on RISC-V vector cores. It reaches geometric-mean speedups over PyTorch of 1.85×, 5.58×, and 5.18× on KernelBench Levels 1-3 on a B200 GPU, 5.8× over the vendor SDK on Adreno GPUs, and 2.7× over ExecuTorch with XNNPACK on RISC-V.

arXiv:2611.04104 PDF

SKILL2STEER: One-Pass Hypernetwork for Mapping Natural-Language Skills to Low-Rank Post-Block Activation Shifts

Ran Li, Lei Chen Some methods compile a natural-language agent skill into a small set of parameters so that the skill text no longer has to sit in the prompt. The authors test whether these compiled artifacts actually carry the skill's content and find that they do not. On a retrieval agent, a wrong skill document performs as well as the correct one, and a single averaged bundle matches per-skill routing, so the compiled shift acts as a domain-general behavioural prior rather than skill-specific knowledge. They then introduce SKILL2STEER, a hypernetwork that reads a skill document once and produces low-rank activation shifts added after every decoder block. Each skill takes 2.36M parameters, 9.25 times fewer than a weight-space adapter, and the method reaches 37.9 exact match on SearchQA, 9.8 points above the frozen model.

arXiv:2611.04497 PDF

Who May Overrule the Agent? Evidence-Gated Authority in LLM Review Institutions

Fuad Ali Multi-agent LLM systems often add a reviewer to check an upstream agent, but a reviewer can overwrite correct decisions as well as fix wrong ones. The authors frame this as a question of when a reviewer should have binding authority. They compare unrestricted override with an evidence gate that allows replacement only after the reviewer produces a mechanically valid record of a constraint violation and a compliant repair. Across two locally run open-weight models, 96 spatial-policy tasks, and a 96-task portfolio benchmark, the gate sharply reduced corruption of correct proposals while keeping most legitimate repairs. In a follow-up where the reviewer could retain, replace, or escalate, Qwen stopped corrupting correct proposals entirely, but Mistral still replaced all 48. Interface design therefore changes reviewer behaviour but does not protect against it independently of the model.

arXiv:2611.05390 PDF

Agent Harness Awareness: Turning Verified Telemetry into Operational Control for Tool-Using Agents

Manu Agrawal, Shrey Nagpal Tool-using agents choose actions from tool descriptions, while the execution harness running them often knows from monitors and recent failures which tools are currently unhealthy. The authors call this gap operational blindness. Their design, Agent Harness Awareness, has the host give the model a small, fresh, verified state snapshot before a consequential action, while the host keeps enforcement and final dispatch authority. On the ToolRoute isolation benchmark with GPT-5.6 Sol, Claude Sonnet 5, and Gemini 3.7 Flash, verified tool-health state cut mean policy regret by 96.9% and raised single-dispatch completion from 69.4% to 95.8%. In a live end-to-end replication with OpenTelemetry traces and Toxiproxy-injected faults, it completed all 9 model-and-fault cases, against 5 of 9 for the control.

arXiv:2611.05391 PDF

Uncertainty-Aware Resource Orchestration for Agentic AI Systems

Masoud Shokrnezhad, Tarik Taleb The compute, bandwidth, and latency an LLM agent pipeline will need only become known as it runs, which leaves infrastructure orchestrators choosing between under- and over-provisioning. The authors formalize this as the problem of reserving resources at dispatch time so that they cover actual demand and meet the deadline with a prescribed probability. They propose CRO (Conformal Reservation for Orchestration), an online algorithm that uses conformal prediction over a sliding window of past execution traces to give distribution-free coverage guarantees. In simulations with multi-modal, heavy-tailed workloads, CRO was the only method that met target coverage at every risk level. It also roughly halved resource waste compared with a Chebyshev-bound baseline.

arXiv:2611.05392 PDF

Challenge-Aware Knowledge Recruitment with Meta-Agent Supervision for Medical Multi-Agent Reasoning

Riccardo Riccio, Haitz S\'aez de Oc\'ariz Borde, Pietro Li\`o Multi-agent debate on hard medical questions often fails to beat strong single-model prompting. The authors argue this is because experts and knowledge are chosen from the question alone rather than from gaps that appear during reasoning. They add an external Challenger agent that inspects the evolving reasoning trace and flags specific factual or expertise gaps, which then trigger targeted retrieval or recruitment of a specialist mid-discussion. On the hard subset of MedAgentsBench, retrieval during the discussion beat retrieval up front, and Challenger-triggered recruitment raised GPT-4.1-mini accuracy on MedQA from 46% to 54%. Naively combining retrieval and recruitment hurt performance, which shows the two intervention channels need coordinating.

arXiv:2611.05393 PDF

LIBRA: Programmatic Induction of Executable Tool Libraries

Zekai Li, Zhenke Liu, Yonghao Li, Junhan Liu, Xulong Xiao, Jiapeng Zhang cross-listed Tool retrievers for LLM agents usually rank each tool independently by its similarity to the request. On multi-tool tasks this finds the tool that matches the goal but misses the prerequisite tools the workflow needs. LIBRA turns a flat tool catalog into an executable library before any queries arrive. An LLM proposes a contract for each tool (its invocations, reusable outputs, state changes, and success conditions), and compositions are tested in a sandbox so that only dependencies that hold at runtime are kept. It needs no dependency annotations, trajectories, or task-specific training, and on AppWorld, BFCL, and SciToolEval it improved one-shot Recall by 7.9 points and NDCG by 3.7 points over the same dense retriever, with further gains inside the agentic retriever ToolReAGt.

arXiv:2611.05445 PDF

Towards Robust Long-Lived Agents: Benchmarking Experience Lifecycles in Evolving Environments

Shijue Huang, Hangyu Guo, Hengheng Zhang, Guo An, Zhitao He, Qi Tian et al. Long-lived agents have to learn how unfamiliar environments work, reuse that knowledge across successive tasks, and keep it current as the environment changes. ReScopeBench tests this across 18 interactive environments. It separates structural experience (how an environment is organized and accessed) from rule experience (reusable principles and constraints). It builds task streams around shared experience and inserts silent environmental changes, then tracks both task performance and how each agent's saved experience changes over time. Across 20 models, retained experience improves performance, but agents struggle to keep that experience up to date when the environment changes, and they sometimes fail tasks even when the needed knowledge is correctly recorded.

arXiv:2611.05450 PDF

SlackBench: Benchmarking Agents on Collaborative Projects Grounded in Real Code Repositories

Mingxuan Li, Huanzhi Mao, Ruitao Zheng, Jarod Stanbury James, Yunze Li, Natanya Anderson et al. SlackBench evaluates agents on realistic collaborative research and engineering work, where answering a question means reasoning over workplace chat together with a code repository whose state changes over time. It is built from privacy-preserving reenactments of real academic projects, with multi-party dialogue, decisions that evolve, and matching changes in GitHub repositories and Google Docs. Its 140 queries cover search across mixed sources, abstaining on unanswerable or false-premise questions, goal-directed summarization, and inferring circumstances from long conversations. Across 33 frontier models and agent harnesses, the best system scores around 80% and most score below 66%.

arXiv:2611.05512 PDF

Iterated creation, grading, and revision of data science projects by language models

Rohan Alexander Seven models from five developers each act as a writer and as a grader. The writer builds an end-to-end data science project (download a dataset, clean it, analyze it, and write a short paper), and the grader scores the result over three rounds of revision. Of 147 initial submissions, 107 ran and rendered, and average scores rose from 59% to 76% after feedback. Gains ranged by writer from 3 percentage points for Opus 5.5 to 46 for Haiku 4.5. Models scored their own work about 9 points higher than other graders did, and about half of the round-to-round change fell within the grader's own noise. How much the loop helps depends on which model plays which role, and weaker graders generally did not improve stronger writers.

arXiv:2611.05548 PDF

Anduin: A sybil-resistant agentic marketplace reputation system

Reza Ramji, Preston Vander Vos, Ziming Zeng, Duc V. Le Autonomous AI agents buying services in permissionless onchain markets need a way to pick providers, but naive volume rankings are easy to game with Sybil attacks (one actor running many fake identities). Using 157 days of data covering 66M agentic transactions worth $6.1M across 16 blockchains, the authors model payments as a buyer–provider bipartite graph and rate each provider by graph paths that connect it to a buyer's existing trusted network. They propose a max-flow rating and two relaxed variants that rate providers more accurately. Against simulated adversaries across 813 buyers, manipulating a rating costs 20% of those buyers' total spend, making it about 23x harder to manipulate than collaborative filtering.

arXiv:2611.05586 PDF

ViBench 2.0: A Benchmark for Long-Horizon Software Building

Daniel Furman, Preeya Kirani, Ibrahim Sheikh, Zhen Li, Peter Zhong, Michele Catasta Existing coding benchmarks test single, fully specified tasks and many are nearly saturated, while users of Replit's agent platform now build whole products through a series of feature requests. ViBench 2.0 contains 17 full-stack web products, modeled on widely used software, that grow over 241 feature requests and are judged only when finished: a browser agent runs 544 test plans, and a build passes only if every plan does. Across nine models, the three strongest pass up to 98% of test plans but deliver a fully working product in at most 65% of builds, and no open-weight model delivers one. Across three releases of one model family, pass@1 on the public products rose from 3% to 53% while cost per build fell by more than half. Eight products are released with the harness and grader, and nine are held out.

arXiv:2611.05620 PDF

Route to Remember: Adaptive Memory Selection for LLM Agents

Zeyu Fang, Jun Yin, Yongsheng Mei, Xu Han, Chaitali Joshi, Zhongkai Sun et al. Agent memory is usually text that is retrieved into context, so the cost of every query grows with the accumulated history. The authors ask whether weight memory, which stores knowledge in model parameters, can serve as a long-term store for facts such as user-specific details. They apply on-policy self-distillation (OPSD): a context-free student is trained against the same model conditioned on the user's memory. A verify-then-regenerate router then sends each query to weight memory or text memory, whichever handles it better. On HaluMem, OPSD matches or surpasses full-context prompting and retrieval baselines while reading fewer tokens per query, and the authors study how training recipes affect performance.

arXiv:2611.05636 PDF

Attributing Agent Failures to Their History Needs an Executed Counterfactual: Toward a Reporting Standard

Egor Pakhomov, Erik Nijkamp Explanations of why long-horizon agents fail usually come from a judge labeling completed transcripts. The authors argue that any claim attributing behavior to an agent's history should instead come from an executed intervention: freeze a mid-trajectory state, alter only the history the agent sees, and continue. The restored state should also carry a measured replay certificate. They note that methods such as DoVer, REFLECT, and Causal Agent Replay already execute interventions but report no such certificate. In a cautionary measurement, hash-identical snapshots restored into identically seeded processes agreed with each other but diverged from the live run in 9 and 20 of 24 states, until the position of the random stream was also restored.

arXiv:2611.05650 PDF

Grounded but Not Immune: Confirmation Bias in Agent-as-a-judge

Drake Benjamin Brown, Khushbu Pahwa, Yubin Ge, Salvatore Romeo, Akhil Agnihotri, Yi Zhang et al. An Agent-as-a-Judge evaluator receives tools and the agent's full trajectory so it can check results against the environment, on the assumption that gathering its own evidence grounds its verdict. The authors show that a single false pre-screening note attributed to an authority overturns up to 40% of correct verdicts, across four judges and five datasets, even when the judge collects decisive evidence. They classify the resulting flips into over-trust (verifying correctly, then deferring to the note), hallucinated support, and under-investigation. Over-trust accounts for 67–75% of harmful flips for the two Claude judges, while smaller judges show all three failure modes. Lightweight mitigations reduce the bias but never remove it.

arXiv:2611.05662 PDF

Agents, Think Only When it Helps

Min-Seon Kim, Seungdong Yoa, Sanghyu Yoon, Ye Seul Sim, Suhee Yoon, Dongmin Kim et al. When large language model (LLM) agents are distilled into smaller student models, the usual practice is to train on full teacher trajectories that mix reasoning thoughts with actions. The authors test whether the thought supervision actually matters by perturbing teacher thoughts and teacher actions independently during distillation, across common agent benchmarks and several model scales. Changing or discarding thoughts has little effect on student performance in most cases. Thoughts become important only when the information they carry is not revealed in the corresponding action, which suggests thought supervision can be dropped selectively to make training cheaper.

arXiv:2611.05709 PDF

Cognitive Bias Signatures in Search-Equipped LLM Agents: A Stage-Resolved Analysis

Yanru Xiao, Yingwei Yu Using classic cognitive-bias paradigms, the authors study how starting conditions such as result order, records of prior failed work, and initial hypotheses shape what search-equipped LLM agents retrieve and conclude. A controlled search harness holds each question and document collection fixed, with matched controls that separate order effects from document composition and the effect of failed work from plain context length. The study tracks queries, retrieved documents, opened documents, and final answers across 10 models and 1,000 synthetic search worlds. Final answers are more sensitive to initial evidence order during active search than when all evidence is given up front. Dead-end histories reduce strategy switching beyond what context length explains, and opening counter-evidence more often does not reliably improve accuracy.

arXiv:2611.05720 PDF

Prospective Memory for Proactive LLM Agents

Yuanchen Bei, Zhining Liu, Tianxin Wei, Weizhi Zhang, Zhichen Zeng, Philip S. Yu et al. Most LLM agent memory is retrospective, but proactive agents also need prospective memory: holding future-directed intentions and acting on them at the right moment without being asked. Pro-Memory, inspired by cognitive theories of human prospective memory, keeps maintaining pending intentions separate from deciding when and how to carry them out. It links each intention to its supporting evidence and activation conditions, updates intention states based on actions actually taken, and distills feedback from finished episodes into an experience memory. On two benchmarks covering delayed-intention tasks and open-ended proactive assistance, it consistently improves proactive task performance across multiple LLM backbones while limiting unnecessary interventions.

arXiv:2611.05731 PDF

PersonaCodeBench: Evaluating Coding Agents in Persona-driven Human-Interactive Sessions

Haixin Wang, Jianheng Hou, Hongyu Shen, Wenzhe Fan, Wei-Chieh Huang, Qianyu Zhu et al. PersonaCodeBench tests whether coding agents can meet a specific user's expectations as well as deliver correct code. It turns persona traits, including role, technical ability, interaction preferences, and practical constraints, into verifiable requirements that surface gradually during simulated multi-turn coding sessions. A controller manages the user simulator and when requirements are revealed. Separate verifiers check functional correctness and persona fulfillment, and a joint pass requires both, within the declared budget. The authors evaluate all major frontier agents and publish a leaderboard.

arXiv:2611.05739 PDF

PAYWORTHY: Can LLMs Predict Which Freelance Jobs an AI Agent Will Complete?

Mohammad Reza Rezaei, Dingxian Wang, Brett Levert, Celeste Fejzo, Andrew Rabinovich, Zhao Chen PAYWORTHY tests whether frontier LLMs can predict, before any attempt, which paid jobs an AI agent will complete successfully. It contains 1,494 real freelance job postings; an agent attempts each one and an LLM judge acting as the client accepts or rejects the result. Eighteen models forecast the agent's success. Told what the agent cannot do, every model ranks jobs well (AUC 0.66–0.86), but told only that it is a generic agent with a browser and code execution, they rank close to random (0.45–0.56). The results hold under judges from four other model families and against two human raters.

arXiv:2611.05759 PDF

BobFlow: Towards Human-Agent Collaborative Data Science Using Dataflows

Jiadong Bai, Yicong Huang, Chen Li cross-listed BobFlow uses dataflows rather than Python or SQL scripts as the shared representation between human analysts and LLM agents doing data science. Combined with the ReAct paradigm, agents build and execute dataflows step by step, while analysts can inspect the analysis and give feedback. A user study found the system helps analysts understand and verify agent-generated analyses. Across proprietary and open-weight LLMs, it achieved accuracy gains of up to 18.6 points at about one-third the inference cost of a script-based ReAct agent.

arXiv:2611.05761 PDF

FlyHash for LLM Agent Memory: An Empirical Validation and a New Application

Mohan Bheemadolu, Chrisogonas O. Odhiambo FlyHash is a hashing scheme inspired by the fruit fly's olfactory circuit, using sparse random projection followed by winner-take-all competition. The authors apply it to retrieval in LLM agent memory and test it against SimHash and HNSW. Under a correctly bit-matched protocol across 10 seeds and four datasets, FlyHash's quality advantage holds, and it also holds on genuine agent-memory relevance labels. However, at equal bits its codes are larger and slower to compare than SimHash's, and HNSW beats both hashing methods at ordinary scale, so its practical niche is narrow. The authors document the methodological pitfalls they found along the way and release their code and a demo.

arXiv:2611.05791 PDF

DARE: Difficulty Analysis for Repair Enhancement

Jikwang Lee, Seungwook Jung, Jaechang Nam cross-listed Solve rates for LLM coding agents hide how difficult the solved tasks actually are. DARE (Difficulty Analysis for Repair Enhancement) defines a task's difficulty as its failure rate across multiple agents and predicts it from 151 static features of the benchmark artifacts. It then identifies which of seven feature groups drive each task's difficulty and uses that to choose task-specific support for the agent. On 2,594 tasks across four benchmark partitions, predictions reach a Spearman correlation of 0.476–0.621 on held-out partitions, and matching support to each task's most salient feature group improves repair success.

arXiv:2611.05797 PDF

Learning When Human Decisions Matter: Executor-Relative Consultation Value for LLM Agents

Shuli Jiang, Khushbu Pahwa, Yuting Zhang, Yi Zhang, Wei Xia, Stefano Soatto When an LLM agent can ask a human only a limited number of times, it has to choose which decisions are worth the human's input, and that value depends on which downstream agent carries out the work. Consultation Value Distillation (CVD) estimates consultation values offline by running Monte Carlo rollouts of the executor agent, then trains lightweight predictors so that no rollouts are needed at deployment. Across five batch and three sequential settings with two frontier executors, CVD broadly outperforms direct prompting and Value of Information baselines. It stays robust with much less rollout supervision and allows nearly lossless approximate allocation at low search cost.

arXiv:2611.05814 PDF

TandemBench: Benchmarking Decision Models for Multi-Agent Collaboration

Baohang Li, Xiaocheng Feng, Bing Qin TandemBench tests how well decision models choose among team plans in multi-agent collaboration, using 1,200 cooking-scenario questions whose reference answers come from execution and exact search. Models get full global information and candidate plans, must pick the best one, and in paired variants must judge whether a better arrangement exists beyond the options given. Across 40 native decision models and 10 general-purpose LLMs, the best native decision model reaches only 62.50% accuracy, and native models are near chance at checking optimality. Some models report better alternatives whether or not the optimal plan is already listed, so plan-selection accuracy alone can hide systematic errors.

arXiv:2611.05827 PDF

Autonomous agents for electricity-system planning

Yoon Pyo Lee, Samrendra Roy, Jay Phil Yoo, Kazuma Kobayashi, Souvik Chakraborty, Syed Bahauddin Alam An autonomous LLM-agent framework handles electricity-system planning end to end. Agents propose investment bounds, storage operating conditions, and feeder interventions, read numerical feedback from automatically executed investment-and-dispatch optimization and feeder verification, revise their designs, and decide when to stop. Across forty carbon-constrained trajectories covering ten scenarios in five electricity systems and four countries, using single- and multi-agent setups, all forty kept candidates that passed the physical and carbon-budget checks. In 28 of the 40 cases, the matched reference designs exceeded the new emission ceilings.

arXiv:2611.05911 PDF

Graph-Mediated Large Language Model Systems in the Agentic Era: Reasoning, Knowledge, Retrieval, Execution, and Coordination

Haihui Liu, Yang Chen, Zice Qiu, Yun Zhang, Jialun Zhong, Wenhua Wang et al. This survey covers how graphs are used in LLM systems for structured reasoning, knowledge grounding, retrieval, memory, execution dependencies, and multi-agent coordination. It introduces an operation-level taxonomy that separates what each graph operation does, which component computes it, and how its output is used later. The survey compares approaches on whether gains in local metrics actually carry over to downstream task performance. It also organizes benchmarks, process diagnostics, and system costs into an evaluation framework, and identifies open problems in adaptive stopping, memory maintenance, and dependency-aware recovery.

arXiv:2611.05953 PDF

Reglos: A Self-Evolving Personal Memory Agent with Recursive Self-Improvement

Sen Yang, Jiaduo Lin, Zehong Huang, Wen Liu Personal memory agents increasingly let coding agents rewrite the harness that turns stored memory into model context, but a patch tested only on the failures it targets can overfit. Reglos is a training-free memory agent with an append-only per-user record and a recursive self-improvement loop. In that loop, a coding agent proposes harness patches from failure traces, and a critic from a different model family screens each patch for benchmark gaming. Once the harness is frozen, a fixed evaluator decides which further patches to accept. Using the same backbone for every system, Reglos outperforms Mem0, MemOS, and EverOS, reaching 84.0% on LongMemEval-S with Qwen3.8-27B and 87.2% with GPT-5.6 Luna.

arXiv:2611.06049 PDF

The Harness Is the Model: Agent Scaffolding Changes a Fixed LLM's Compliance with Banking Conduct Rules, Unevenly Across Models

Mehmet Murat Huyut Banks validate language models on their own, but they deploy them inside harnesses: the surrounding code that controls what the model sees, which tools it can call, and who checks its actions. The authors held each of four open-weight models fixed and ran it in five harnesses on 156 pre-registered cases in a simulated retail bank whose rules paraphrase EU conduct regulations, scoring all 3,120 episodes deterministically from the executed tool trace. Pooled across models, the harness barely moved the correct-case rate (87.3% to 92.1%). Within a single model it could matter a great deal: one model's 11.5-point spread between its best and worst harness moved it from first to last place, while another model showed no detectable effect. A same-model compliance reviewer cut violations from 9.6% to 6.4% without raising over-refusals, rules that need date arithmetic failed in every harness, and 26% of violations came with a customer message denying the action the agent had just executed.

arXiv:2611.06116 PDF

ModalSearcher: Weaving Visual Evidence into Autonomous Cross-Modal Deep Search

Zhihan Yin, Ruyi Li, Chengyu Fang, Yifeng Yao, Nonghai Zhang, Pengfei Xiong et al. Multimodal search agents usually fix where visual evidence enters a trajectory, so they cannot decide for themselves when to search with text, with images, or by reverse-image lookup as their information needs change. The authors build ModalSearch-Pipeline, which samples paths from an interleaved multimodal knowledge graph and writes questions backward, hiding entity names and image targets so the agent cannot take shortcuts. They use the resulting data to train ModalSearcher with supervised fine-tuning and reinforcement learning in a live web environment, and they release ModalSearch-Bench. Under the same agentic workflow, ModalSearcher-8B improves average accuracy by 16.2 points over Qwen3-VL-8B across five knowledge-intensive multimodal QA benchmarks, with consistent gains on separately built active visual-search benchmarks.

arXiv:2611.06131 PDF

HarnessFlow: MCTS-Guided Harness Optimization for LLM Agents

Zhi Hong, Qianlong Wang, Hongchen Wei, Aidan Wu, Daohong Ying, Mingze Kong et al. An LLM agent's performance depends on the harness that governs its tool use and its interaction with the environment, yet harnesses are mostly designed by hand and evaluating each candidate end-to-end is expensive. HarnessFlow keeps the model and environment fixed, stores candidate harnesses and their evaluation histories in a refinement tree, and uses Monte Carlo Tree Search (MCTS) to pick which harness a coding agent refines next, drawing on archived code, execution feedback, and a library of reusable harness mechanisms. With 15 refinement attempts per benchmark across six agentic, coding, and web benchmarks, the optimized harnesses average 69.48%, beating the initial harnesses by 10.89 points and Meta-Harness by 3.95 points, at lower per-task cost than hand-engineered baselines.

arXiv:2611.06254 PDF

MA-Mesh: Resilient Multi-Agent Collaboration via Role-based Self-Healing

JunYong Lee, Woo Kyung Kim, Honguk Woo Existing multi-agent systems rarely model resilience explicitly and instead rely on ad-hoc correction of outputs. Multi-Agent Mesh (MA-Mesh) generates resilience-aware system specifications, turns them into executable multi-agent policy code, and adds a role-based self-healing layer on top of a collaboration pattern. That layer has two mechanisms: cross-agent reflection, in which agents evaluate their peers as well as themselves, and role-based failover, in which agents take over each other's roles when sub-tasks fail or an agent underperforms. On software development, tool use, and scientific reasoning tasks, MA-Mesh outperforms existing multi-agent systems in both task performance and robustness under system uncertainty.

arXiv:2611.06389 PDF

Zero-Shot Coordination among LLM Agents

Adrian Hayler, Shashank Reddy Chirra, Andrei Lupu, Johannes Forkel, Bidipta Sarkar, Siheng Feng et al. cross-listed Zero-shot coordination (ZSC) asks whether independently developed agents can coordinate at test time. It is well studied in reinforcement learning but much less so for LLM agents. Using simple, general-purpose scaffolds in minimal environments built to isolate specific coordination challenges, the authors find that frontier LLM agents struggle to coordinate even in these controlled settings. The main causes are a limited understanding of the coordination problem and weak reasoning about the partner's beliefs, and the agents also fail to exploit shared semantic information. The authors propose Coordination-friendly definitions (CFDs) as a principled fix and show they can be discovered automatically.

arXiv:2611.06409 PDF

Twists and Turns: A Graph-Based Benchmark and Framework for Long-Term Memory in Exploration Tasks

Chongyuan Dai, Yaling Shen, Zirui Song, Junfeng Liao, Jinpeng Hu, Xiuying Chen Long-term memory research has focused on recalling and updating facts. Exploration tasks such as research or software development also require remembering how attempts were tested, revised, abandoned, and revived. The authors model these processes as graphs and use an evolutionary mechanism to make them structurally more complex, producing LExMem, a benchmark of 224 dialogues and 2,018 question-answer instances. They also propose LExGraph, a memory framework that incrementally organizes dialogues into linked memory nodes and retrieves evidence through hybrid retrieval plus neighbor expansion. LExMem proves hard for existing retrieval and memory systems, while LExGraph performs best and generalizes to external benchmarks.

arXiv:2611.06414 PDF

EviSeek: An Evidence-Centric Agentic Framework for Embodied Active Information Seeking

Jintao Li, Zhiyong Cao, Baihua Xiao, Yu Bai, B\"orje F. Karlsson In active embodied question answering, an agent gathers evidence by changing viewpoint and physically interacting with its environment, but agents often act without specifying what each action is meant to reveal. EviSeek is a training-free framework that turns the question into explicit evidence demands and uses them to guide action choice and decide which evidence to keep. On 3,436 ESI-Bench questions it improves accuracy over a single-model agent by 7.1 points with Qwen3-VL-32B and 3.7 points with Gemini 3.1 Pro. It also improves visually supported accuracy and reduces average token usage.

arXiv:2611.06434 PDF

EvoGuard: Mitigating Degradation in Online Skill Evolution

Yanli Wang, Zichen Bai, Tianyue Jiang, Kunming Shao, Junhao Shi, Xueqin Liu et al. Language agents can store reusable skills as natural-language documents, but when skills evolve online, adopting every update can degrade them over time, and validating each update on a held-out set is too slow and costly. EvoGuard trains a pairwise selector on skill pools labeled by their utility on a private test set, using rejection sampling to keep only pairs with a clear utility gap. The selector reads the logits of two label tokens to pick the better skill in a single forward pass. Compared with adopting every new skill, it improves success by 11.0% on ALFWorld, 5.9% on ScienceWorld, and 2.7% on LiveCodeBench.

arXiv:2611.06444 PDF

SUMONE: A Simple and Efficient Summarizer for Plug-and-Play Agent Memory

Jizhe Zhang, Ang Li, Chenheng Zhang, Yisen Wang Long-horizon agents accumulate large interaction histories, and existing learned compaction methods tie their training to specific tasks or agents. SumOne trains a standalone summarizer with GRPO, using a task-agnostic reward: a frozen reconstructor must recover randomly masked parts of the original trajectory from the summary, with a penalty for summary length. This needs no reference summaries and no downstream task runs, and at inference only the lightweight summarizer is kept as a plug-in context compressor. It cuts estimated inference cost by about 78% on DeepSearchQA and 75% on SWE-bench Verified compared with keeping the full context, with reported absolute decreases of 0.04, and the same checkpoint transfers across downstream models and agent harnesses without further fine-tuning.

arXiv:2611.06445 PDF

What Does Multi-Turn Reinforcement Learning Really Improve in LLM Agents?

Kejian Zhu, Zhuoran Jin, Shangqing Tu, Tianyi Men, Yupu Hao, Jiachun Li et al. cross-listed Multi-turn agentic reinforcement learning (RL) looks strong on long-horizon tasks, but the authors ask where those gains actually come from and how much they depend on interaction limits. Using a Pass@(K,T) metric that varies both the number of samples K and the number of turns T, they find that trajectory-level RL gains depend on generous turn budgets and shrink sharply when budgets tighten. Agentic RL does not uniformly improve single-step decisions; its benefit lies mainly in recovering from errors and staying on track over long horizons. They briefly examine process supervision signals and propose BudgetMix RL, which varies interaction budgets during training and performs robustly across execution settings.

arXiv:2611.06483 PDF

Tools Need Not Speak in Text: Latent Context Compression for Agents

Yuetong Liu, Zhehao Zhou, Yitong Zhang, Haorui Chen, Jia Li Large language model (LLM) agents usually append raw tool outputs to their context, which inflates token usage and spreads the model's attention across noisy observations. Coca-Cola instead compresses tool observations into continuous latent tokens injected directly into the agent's input embedding space, and is trained with Action-Preserving Training so that the agent's next-action distribution under compressed observations matches its distribution under the full text. On web-search agents across four QA benchmarks and two model families, it reduces token consumption by 70-89% and end-to-end runtime by 24-56% while also improving task performance.

arXiv:2611.06500 PDF

Using Experience Twice: Dynamic Process Rubrics for Privileged Token-Level Credit Assignment

Shumao Wu, Kongcheng Zhang, Jianqiang Huang, Baisheng Lai Agent trajectories mix good decisions with failed attempts, and sparse trajectory-level rewards cannot tell which steps deserve credit. Online Rubric Self-Distillation (ORSD) has the policy generate natural-language process rubrics from differences across a group of trajectories, uses a frozen verifier to choose the rubric bundle that best matches task success, and treats that bundle as privileged information. The change in each token's log-likelihood with and without the rubric becomes a token-level credit signal added to the GRPO advantage. On ALFWorld and WebShop with Qwen3-1.7B and Qwen3-4B, ORSD speeds up policy improvement and raises final performance under the same interaction budget.

arXiv:2611.06543 PDF

When Do Summaries Become Instructions? Unsupported Policy in LLM Agents' Self-Written Handoff Summaries

Wenhui Chu Long-running LLM agents compact their context into handoff summaries, and the authors check whether those summaries invent permissions or restrictions that the record does not support, and whether this changes what the next agent instance does. Using synthetic tool-use tasks with a ground-truth permission ledger and replays that swap suspect sentences for neutral filler, they find unsupported restrictions in 1.2-2.4% of summaries and unsupported authorizations in 0.1-1.3%. Neutralizing false approvals reduced unauthorized actions from 24 to 7 of 120 runs, although that analysis was post hoc. In pre-registered tests, rewording the system prompt's "Some actions need approval" to explain where requirements come from cut unsupported restrictions from 5.0% to 1.8%, without suppressing legitimate approval requests.

arXiv:2611.06545 PDF

Where Search Agents Go Off the Rails: Graph-Grounded Step-Level Verification

Hyeseo Jeon, Seonho Yoo, Jay-Yoon Lee Search-augmented reasoning systems can reach correct answers despite faulty intermediate steps, so auditing them requires checking that evidence covers the question, reasoning addresses it, and reasoning is grounded in the evidence. GraphRail turns requirements, evidence, and reasoning into structured claims and compares them as a graph, without a generative judge or gold answer. Across three multi-hop QA benchmarks and three reasoning frameworks, it finds more intermediate errors than generative verifiers while staying stable under stochastic extraction, and its diagnoses can guide online correction during reasoning.

arXiv:2611.06662 PDF

AgentTrustCI: Build-Time Policy, Grounding, and Trajectory Tests for Java LLM Agents

Sachin Gupta cross-listed Infrastructure security does not show whether an LLM agent chose an allowed tool, grounded its answer in tool output, or followed a required workflow. AgentTrustCI is a Java/JUnit 5 framework that expresses these expectations as policy, grounding, and trajectory checks over recorded traces, using deterministic replay for repeatable build-time tests and an optional LLM-judge interface. On 55 synthetic cases across seven enterprise domains, the full pipeline correctly classified all 30 injected faults and 25 passing controls, though the authors stress that this shows coverage of their authored suite, not production accuracy.

arXiv:2611.06674 PDF

LifeVal: Evaluating AI Model Performance on Personalized Agentic Consumer Tasks

Jenny Yang, Elizabeth Proehl, Anjan Bharadwaj, Emily Fan, Charlotte Cole, James Kim et al. LifeVal is a benchmark of personalized agentic consumer tasks, such as managing finances, moving apartments, or finding childcare, weighted by what people would pay a digital assistant to do them. Task categories come from a willingness-to-pay survey of over 3,800 demographically representative consumers, cross-checked against US and multinational time-use studies, and together cover 148 categories representing over $1.7 trillion in annual willingness-to-pay. Each task is personalized with real ChatGPT conversations, memories, and files, written with a rubric by creators who have done the task themselves, and passed at least three rounds of human quality control.

arXiv:2611.06687 PDF

Route First, See Later: Enabling Vision Agents to Reason Scientifically by Separating Routing Policy from Action Spaces

Shubhabrata Mukherjee, Jeffrey Donatelli, Daniela Ushizima cross-listed Vision agents for scientific image analysis have to choose between capabilities such as description, detection, counting, and segmentation, but current multi-tool designs mix up interpreting intent with executing visual tools, so it is unclear whether a failure came from misrouting or from bad perception. ZETA separates the routing policy from action execution: it resolves the scientific intent first, then calls specialized tools that treat measurement as a deterministic operation grounded in explicit geometry. In a controlled zero-shot study of 51 scientific-intent queries with the action interface held fixed, spelling out the routing policy explicitly raised Qwen 3.8's segmentation routing accuracy from 6.2% to 75.0% and its detection routing accuracy from 60.0% to 100%, although the gains varied across models. Its execution traces make routing errors and visual-execution errors separately observable.

arXiv:2611.06760 PDF

Transforming Integer Grids to Spatial Intuition: Perception, Not Planning, Governs ARC-AGI-3

Harsidak Singh Banwait cross-listed On ARC-AGI-3, which consists of uninstructed visual games scored by Relative Human Action Efficiency, frontier models that read raw integer grids lose spatial structure and waste hundreds of actions rediscovering basic boundaries. SARBLOH is an open-weight agent built on the premise that perception, not planning, is the bottleneck. A zero-parameter perception engine turns each frame, in 35 ms, into an object inventory with tracking over time, a geometric briefing that separates proven invariants from guesses about what actions do, and an annotated canvas, and a code-as-action runtime lets the agent write algorithmic hypotheses instead of low-level moves. Grounded perception raised the rubric score from 3/28 to 25/28 while cutting token use by 89.4%; geometric invariants generalized perfectly to 23 held-out games while single-frame action guesses did not. Running on one Kaggle GPU within the nine-hour budget, the agent scored 16.53 on the hidden evaluation, 13.7x the first Milestone winner.

arXiv:2611.06793 PDF

Task-Efficient Agent Evaluation Using Cached Traces

Ben Johnson, Hayden Helm Efficient benchmarking tries to estimate a system's full benchmark score from a few task runs, but standard methods use only pass/fail results and ignore the rich traces that agentic systems produce. The authors have an LLM extract a structured summary of each trace and embed it with a standard text encoder, which keeps signal about correctness and the task while removing much of the system-, model-family-, and harness-identity signal that dominates raw trace embeddings. They then place a new system among the already-evaluated reference systems in a low-dimensional space and predict its score from that position. Added to an item-response model and given just one random executed task, this cuts mean absolute error from 0.13 to 0.08 on SWE-bench Verified and from 0.14 to 0.12 on Terminal-Bench 2.0, and with all reference systems available it beats the baselines at every task budget.

arXiv:2611.06803 PDF

ServerBench: How can your coding agents design specialized LLM serving systems?

Xinrui Zhong, Zhiming Wang, Yongqi Liu, Zhaohui Wang, Shanglin Yang, Zikun Li et al. ServerBench asks LLM coding agents to design complete, specialized LLM serving systems, with the agent coordinating implementation and operating decisions under session dependencies, latency constraints, and correctness requirements. Agents optimize a single objective, "dollars per dollar" (latency-qualified token value per dollar of GPU cost), measured with stateful replay that preserves session dependencies, and an independent check verifies each submitted system. Across 11 coding-agent models and three deployments, Claude Sonnet 5.5 reached 2.97 dollars per dollar on Gemma E4B, 4.87x the vLLM reference; 22 of 33 runs produced verified systems and 13 beat the strongest baseline. The analysis finds that high-value designs coordinate capacity improvements with load selection, and that reliable delivery depends on verifying the new execution paths that optimizations introduce.

arXiv:2611.06841 PDF

CompressionBench: How can your coding agents compress LLMs?

Xinrui Zhong, Zhifeng Li, Zhiming Wang, Yongqi Liu, Shanglin Yang, Yang Zhou et al. CompressionBench tests whether coding agents can autonomously compress an LLM, framed as a terminal task: given a source checkpoint, a storage limit, an inference environment, and a time budget, the agent must produce compressed weights and working inference code, guided by perplexity feedback and checked by independent verification. Across six agents and six model-and-size settings, 29 of 36 submissions passed verification. Choosing the lowest-perplexity agent design per task gives an average 3.4% perplexity increase over BF16 and 77.33% downstream accuracy versus 79.56% for the uncompressed model. Logged experiments show that agents lowered perplexity in 21 of 24 revisions at nearly fixed storage, that full-model feedback reversed 17 of 79 earlier local rankings, and that generation and inference tests catch failures that likelihood alone misses.

arXiv:2611.06858 PDF

On Turn-Level Best-of-N Sampling for Scaling LLM Agents

Jalaj Bhandari, Dhawal Gupta, Hongli Zhan, Zhenting Wang, Varad Pimpalkhute, Zhengzhong Liu et al. The authors test a simple way to spend more inference compute on long-horizon LLM agent tasks: turn-level best-of-N sampling. At each turn a generator model proposes N candidate actions and a separate selector model picks one to execute. Unlike earlier work that relied on trained critic models, the selector here is an off-the-shelf open-weight LLM that compares the candidates directly. On long-horizon terminal tasks the method improves accuracy by 5–10% and makes repeated attempts more consistent. Trajectory analysis shows agents using fewer attempts, following instructions more closely, and debugging and testing more.

arXiv:2611.06868 PDF

SpecCompact: Accelerating LLM Agent Context Compaction with Slot-Parallel Semantic Speculative Verification

Yulu An, Yifei Feng, Xiangyang Yin LLM agents manage long interaction histories by compacting them into summaries, but summarizing with a large model adds significant latency. SpecCompact relies on the observation that structured summary slots can be generated independently without losing information. A small fine-tuned drafter writes all slots in parallel, and the large target model verifies each drafted line and repairs any it rejects; further drafting continues from the corrected slot prefix. On AppWorld and τ³ with two target models, compaction runs 2.04–5.63x faster than Hermes-style summarization, with downstream trade-offs that vary by model and task, and a user study inside Hermes measured an 8.68x median speedup.

arXiv:2611.06876 PDF

FrontierPhysics: Evaluating Agents for End-to-End Physics Research

Bingran You, Wenjun Ke, Ria Rosenauer, Bo Liu, Wenbo Chen, Min Chen et al. Existing physics benchmarks test fixed questions, isolated sub-tasks, or reproductions of known results, and rarely judge the quality of the research itself. FrontierPhysics has 56 end-to-end tasks built by 42 physicists from their own research, spanning nine theoretical, computational, and experimental domains. In each task, agents must review the literature, plan and carry out the research, and write a manuscript, which is graded by deterministic verifiers and expert-written rubrics. Across 22 models in 26 configurations, no configuration passes more than 46 of its 168 rollouts. Manual review traces the failures mainly to insufficiently rigorous reasoning and weak expert intuition, rather than missing physics knowledge or execution errors.

arXiv:2611.06910 PDF

3 more specialized papers

Theory 55

Predictive gains, causal losses: why interventional competence cannot be learned from data alone

Marcus Thomas cross-listed The paper argues that a model can keep gains in predicting observational data only by making persistent errors in how it responds to interventions. For deterministic predictors under squared-error loss, any advantage over the best predictor that uses only an outcome's direct causes requires depending on other inputs, and when interventions change those inputs without changing the outcome, this gives a positive lower bound on intervention-response error, whatever the learning rule. The author identifies conditions under which gradient updates introduce such dependence or damage a mechanism that was previously correct through shared parameters. They also show that correction is limited by which causal hypotheses are currently being considered, and conjecture that reliable interventional competence requires an agent to propose, test, and revise causal alternatives itself.

arXiv:2611.03861 PDF

Theoretical Analysis of Muon on Transformers for Next-Token Prediction

Shengwei Wu, Yixuan Yang, Gao Huang, Song Li The Muon optimizer performs well when training Transformer language models, but its dynamics are poorly understood. The authors analyze Muon on next-token prediction in two phases: a preprocessing phase that learns the output-value weights, where they characterize the trajectory exactly and show that the normalized spectral margin is optimal after every update, and a main phase that learns the key-query weights, where they prove nonasymptotic convergence rates under softmax attention. The results suggest that Muon helps models learn efficiently which tokens matter for prediction, and the learned weight structure yields a generalization guarantee for unseen contexts, which experiments support.

arXiv:2611.04465 PDF

PAC Learning with Multiple Correct Answers

Steve Hanneke, Farnam Mansouri, Alireza F. Pour, Amirreza Shaeiri, Kun Wang The authors study PAC learning when each example has a set of up to $\ell$ correct labels. During training, the learner may see the full set, only whether a proposed label is correct (bandit feedback), or a single correct label chosen by a demonstrator. The main finding is that learning from demonstrations can be much harder than learning from full sets or even from bandit feedback, both in which classes are learnable and in how many examples are needed. A generalized Natarajan dimension characterizes learning from full sets and from bandit feedback. Demonstration learnability reduces to whether intersections of $\ell+1$ hypotheses are learnable, with a $\widetilde O(\varepsilon^{-\ell})$ sample complexity that cannot be improved in general. The authors also show that the uniform bound on the number of correct labels is essential for any dimension-based characterization.

arXiv:2611.05133 PDF

Provably Learning Mixtures of Linear Experts

Adrian Candocia, Aamod Varma, Mark Davenport Gradient-based training of Mixture-of-Experts (MoE) models has few convergence guarantees, because the objective is nonconvex and even simple mixtures trap first-order methods in spurious local optima. The authors analyze a gated mixture of linear experts whose gate comes from Gaussian clusters in the input space. They prove that gradient EM on its likelihood jointly trains the gate and the experts with the first global convergence guarantee for this setting. Accuracy is reached in polynomially many iterations, given logarithmic overparameterization and well-separated gate centers. Experiments confirm the predicted m log m phase transition and suggest that randomly initialized experts work about as well as the least-squares warm start the proof needs. Unconstrained mixing weights concentrate routing on the true experts, so the fitted model can be pruned, whereas load-balancing penalties spread routing out.

arXiv:2611.05184 PDF

Logarithmic Regret for Alternating Mirror Descent

Tianlong Nan, Christian Kroer Alternating online mirror descent (AltOMD) is a learning rule for two-player zero-sum games that outperforms the simultaneous version, but its best known regret bounds were O(T^{1/3}) and O(T^{1/5}). The authors settle the open question of whether it achieves near-constant regret. They prove an O(log T) bound on the sum of external regrets for any bilinear zero-sum game with compact convex strategy sets, under mild assumptions on the regularizers. The assumptions hold for negative entropy, which gives logarithmic regret for alternating multiplicative weights and an O(log T / T) duality gap for the averaged strategies.

arXiv:2611.05243 PDF

Conformal Prediction at Performative Equilibrium: The Price of Performativity

Ahmed Mahrous, Roberto Di Pietro cross-listed Conformal prediction's finite-sample coverage guarantee assumes calibration and test points are exchangeable, which fails when deploying a predictor changes the distribution it will be judged on. Framing this as performative prediction, the authors require coverage at performative equilibrium and prove a phase transition in the total-variation sensitivity L of the distribution map: when L is at least 1 no finite-sample procedure can deliver both equilibrium-valid coverage and non-trivial prediction sets, while below 1 the minimax excess miscoverage scales as 1/((1-L) times the square root of n), inflating sample complexity by a factor of 1/(1-L)^2. The main tool is a transfer theorem showing any procedure that returns a predictor within epsilon of equilibrium yields equilibrium-valid coverage up to an additive L-times-epsilon term after split-conformal calibration, instantiated with a model-free and a parametric tail-averaged retraining algorithm and checked on semi-synthetic credit scoring.

arXiv:2611.05277 PDF

Agnostic Learning by Imitating

S\'ilvia Casacuberta, Parikshit Gopalan, Varun Kanade, Omer Reingold, Konstantinos Stavropoulos, Pranay Tankala Which is computationally harder: training a discriminative model that separates binary labels as well as the best member of a model class, known as agnostic learning, or training a generative model whose outputs no member of that class can distinguish from real data, known as multiaccuracy? The two tasks turn out to be equivalent in complexity, with the previously missing direction supplied here as an efficient oracle reduction from agnostic learning to a multiaccuracy oracle, built by connecting the offline problem to no-regret online learning and using follow-the-leader style algorithms to modify the label distribution and aggregate oracle predictions. Pushing further, a multiaccuracy oracle is shown to reach a range of seemingly more demanding learning objectives by similarly direct constructions; the results hinge on adaptive queries, in contrast to earlier work showing that post-processing a single multiaccurate predictor cannot yield even weak learning.

arXiv:2611.05346 PDF

A Feasibility Condition for Confidence-Gated Tiered LLM Inference

Sumit P. Ahuja Tiered LLM inference answers easy queries on local or edge devices and escalates hard ones to a datacenter model, using a confidence signal to decide. The paper treats this routing decision as a binary test and derives a closed-form condition for when no confidence threshold makes such a cascade worth more than sending every query to the backstop. The condition depends on an architecture cost ratio κ between local and escalation cost. When local compute is billed, cost is the binding constraint; when it is already paid for, as on decentralized hardware, a cascade is always feasible and signal quality only sets how much it gains. The authors turn the condition into three design requirements and propose a co-designed architecture as future work.

arXiv:2611.05394 PDF

AI-Assisted Classification under Correlation Neglect and Trust

Dingwen Kong, Liang Lyu, Saurabh Amin, Amine Bennouna, Daniel Huttenlocher, Asuman Ozdaglar cross-listed Whether an AI assistant helps or hurts a human's decision depends on its accuracy, on how much its evidence overlaps with the human's, and on how the human combines the two. The authors model a worker making a binary decision with two parameters, awareness of that overlap and trust in the AI, and derive a closed-form map of regions where the pair complements each other, where the AI impairs the human, and where full automation is better. A worker who ignores the overlap double-counts shared evidence and can be made less accurate by a more capable AI, and some undertrust can partly offset this. On ChatBench, AI help raises accuracy by 21 points, yet workers give AI information only about a quarter of the weight a fully rational Bayesian would, and calibrating trust recovers about 2.6 points with GPT-4o but almost nothing with Llama-3.1-8b.

arXiv:2611.05870 PDF

Who Pays for the Knowledge Models Copy? Model Collapse as Rogers' Paradox

Nam H. Le cross-listed The authors argue that model collapse, the loss of rare knowledge when generative models train on earlier models' outputs, is an instance of Rogers' paradox from cultural evolution: a population that relies on copying ends up no better off than one that never copied. Rogers' model, fitted only to fully synthetic training, predicts to within a few percentage points how much rare knowledge models from word trigrams to a pretrained transformer retain under other mixes of fresh data. It also predicts that a lineage copying its own past needs more fresh data the faster the world changes. When people can ask the model instead of finding things out themselves, the equilibrium model is only as reliable as independent learning is costly, which leaves the population no better off than without it. Users who verify answers break the paradox, while filters that select plausible text accelerate collapse.

arXiv:2611.06022 PDF

45 more specialized papers

Safety & Alignment 32

Dr. No: Feature Steering for Abstention in Vision-Language Models

Farid Hasanov, Riccardo Renzulli, Bahram Zonooz, Mykola Pechenizkiy, Decebal Constantin Mocanu Foundation models often answer even when the visual or textual evidence is not enough to support a reliable conclusion. Dr. No finds sparse autoencoder (SAE) features linked to abstaining or committing to an answer, then adjusts a small set of them during the prefill pass at inference time, without retraining the model. Tested on LLaVA-1.6 Mistral 7B with UNK-VQA and MM-UPD, and on Qwen3.5-2B with AbstentionBench subsets, feature steering balances correct answering against appropriate abstention better than post-hoc rejection baselines based on uncertainty. Gains are weaker on absent-answer detection and retrieval-heavy multi-hop QA such as MuSiQue, which suggests that the signals controlling abstention depend on the type of task rather than forming one universal signal.

arXiv:2611.03862 PDF

Forget What Comes Next: LLM Unlearning via Statistical Independence

Andrii Shkabrii, Claudia Plant, Martin Pawelczyk Large language models (LLMs) can memorize training data, and current unlearning methods may leave forgotten data statistically detectable, so it is unclear whether they actually remove memorized content. PINT uses the Hilbert-Schmidt Independence Criterion (HSIC) to reduce the statistical dependence between the tokens to be forgotten and the internal representations used to predict them, and proves an upper bound linking this objective to exact memorization. Experiments cover Gaussian poisons, personal information, hazardous biology and cybersecurity knowledge, and copyrighted text, on models up to 7B parameters, with scaling demonstrated up to 70B. PINT gives the best average trade-off between forgetting and utility and is the only method that makes Gaussian poisons undetectable while keeping near-original utility.

arXiv:2611.03955 PDF

ScrambleMark: Zero-Query Watermark-Agnostic Black-Box Evasion of Modern Speech Watermarking

Binhao Ma, Weikang Ding, Hanqing Guo, Wenwei Zhao, Xiaowen Li, Leah Ding et al. cross-listed Existing attacks on speech watermarks try to erase them, which needs detector access or hard-to-collect paired audio, and commercial platforms are closing off both. ScrambleMark instead breaks the structural consistency that detectors depend on by scrambling harmonic structure, which preserves perceived audio quality. It needs no detector queries and no knowledge of the watermark, combining a deterministic engine that finds transferable harmonic perturbations with a generative model that learns randomized scrambling patterns. Against Google SynthID, Resemble AI, and TrustedAudio, it evades detection about 97% of the time with a 4.2/5 human naturalness rating, versus 1.1-1.7 for prior attacks, and it holds up against adaptive recovery attempts.

arXiv:2611.04122 PDF

Say the Same, Act Differently: Diagnosing and Steering Text-Action Divergence in Reasoning VLAs

Zihao Feng, Qingzhao Zhang, Chunyu Xia, Bo Yu, Z. Morley Mao, Ramesh Govindan Reasoning vision-language-action (VLA) models produce natural-language rationales that supervisors can read to monitor safety. The authors show that visual perturbations can change the physical action while leaving the rationale word-for-word identical. Their text-neutral action-subspace (TNAS) diagnostic finds that the most action-sensitive hidden-state direction is nearly orthogonal to text-loss gradients in Alpamayo (driving) and InstructVLA (manipulation). Attacks guided along this direction cause a new collision with an unchanged rationale in 52.6% of 154 driving scenes, versus 4.55% for unguided attacks, so monitors must check the predicted actions directly rather than trust the text.

arXiv:2611.04166 PDF

Where Did the Update Go? Certifying Declared Change Envelopes for Model Updates

Muhammad Usama, Muhammad Saad Qadir When a deployed language model is updated, regulators may want evidence that its behavior changed only in a region declared in advance, which global regression bounds cannot certify. The authors formalize declared change envelopes, count every decision changed outside the region as leakage, and certify it from unlabeled traffic with anytime-valid tests that compose across a model's lifetime of updates. Across LoRA, full fine-tuning, preference optimization, reinforcement learning, and knowledge editing on 1.5B–72B models, narrow updates leak widely: only about 7–36% of changed decisions stay inside the declared scope, and leakage tracks out-of-scope KL divergence. Distilling the old model's decisions on unlabeled prompts during the update makes 5% leakage certifiable from a few hundred to a few thousand samples, though only for decision-like outputs, not open-ended chat.

arXiv:2611.04192 PDF

Voluntary Servitude: Weaponized Safety Warning Mechanisms for Controlling AI Agents

Ilan Zendel, Matan Levy, Stav Cohen, Avi Mendelson, Amit LeVi cross-listed Security work on LLM agents usually models manipulation as instruction injection, and deployed defenses learn to recognize command syntax. The authors show that an attacker can instead exploit the agent's own harm-avoidance: a single fabricated hazard, written as an ordinary review, leads the agent to rule out legitimate options and pick the attacker's target by its own reasoning, which they call voluntary servitude. Across six models and ten consumer-service domains these payloads steer agents about as often as instruction injection (60.4% vs. 66.0%), but injection detectors score them as benign, so they are 1.7-4.4x more effective behind a filter, and the attack transfers to production browser agents and to retrieval pipelines. Adding a corroboration requirement to the agent's instructions removes roughly 95% of the exposure against a non-adaptive attacker without retraining.

arXiv:2611.04393 PDF

Knowingly Wrong: Training LLMs Against Their Own Beliefs can Induce Misalignment

Ariana M. Villegas-Suarez, Guido Freire, Sohan Venkatesh, Bartosz Cywi\'nski cross-listed Fine-tuning a large language model (LLM) on narrowly harmful data can make it broadly misaligned, but models are rarely trained on explicitly harmful content in practice. The experiments here use data with no harmful content at all, only answers the model believes to be false, and find that across mathematics, trivia, cooking, and sports the more strongly the model disbelieves what it is trained on, the more misaligned it becomes: on mathematics, misaligned answers rise from 4% to over 60% as the training answers get more wrong, with little loss of general capability. Multiple-choice probabilities still favor the true answers after fine-tuning, suggesting the misalignment stems from that internal inconsistency; the same pattern holds when a single fact is updated, as in knowledge editing. Framing the false fact inside a coherent story makes the model adopt it more deeply and stay more aligned than instructing it to assert the fact, yet content filters and persona vectors cannot distinguish the two framings.

arXiv:2611.05287 PDF

Suppressing Misaligned Behaviors in Language Models, Durably

Luxi He, Pengcheng Jiang, Jifan Zhang, Jiawei Han, Danqi Chen cross-listed To study how well an undesired behavior can be removed after it appears, the authors build two deliberately misaligned model organisms: a narrow-trigger one that cheats on multiple-choice questions only when a planted hint points to a wrong answer, and a broad-trigger one that answers harmfully across open-ended prompts after fine-tuning on insecure code. They formalize post-hoc behavioral suppression and score methods on immediate suppression, durability under further training, and capability preservation, comparing supervised fine-tuning, GRPO, and gradient-ascent unlearning. Their method PEARL (Prefix-Expanded Adversarial Reinforcement Learning) is a GRPO variant that augments rollouts with continuations from intermediate states sampled along cached misbehaving trajectories, on the hypothesis that durability requires recovering from a wider range of misaligned states. On Qwen3-4B and GPT-OSS-20B it suppresses the behavior while preserving task accuracy and staying durable under further fine-tuning on both the organism's own data and unrelated data, gaining most where misbehavior is broad and targeted fine-tuning cannot reach it.

arXiv:2611.05358 PDF

Assurance-Conditioned Continuous Authorization for Autonomous AI Agents

Mehdi Shajari Autonomous agents can call tools, delegate work, and change their models or environments while earlier authorizations are still in effect. Assurance-Conditioned Continuous Authorization (ACCA) keeps an agent's authority in line with the assurance its current evidence supports. It derives assurance claims from evidence that is both valid and applicable, resolves an effective-authority envelope within a ceiling declared by the principal, and uses a typed dependency graph to re-evaluate only what changed. When authority changes, a new authority epoch lets the system reject stale actions at execution time, before they change a resource. Regression-suite scenarios with a trading profile exercise default-deny rules, contraction and recovery of authority, and stale-action rejection, providing implementation-level evidence in one domain only.

arXiv:2611.05632 PDF

The Male Abstraction Bias in Internet Culture and Large Language Models

Douglas Guilbeault, Bhargav Srinivasa Desikan cross-listed Across roughly 40,000 single-word and 67,000 multi-word concepts, measured in billions of words and millions of images from the internet, more abstract concepts are consistently more associated with men, and the effect replicates in languages around the world. The pattern also shows up in grammar: verbs and adverbs are both more abstract and more male-coded. The same bias appears in free-association responses from both humans and LLMs, which the authors read as evidence that people internalize it and AI models reproduce it.

arXiv:2611.05730 PDF

When File Ingestion Misleads Language Models: Measuring and Mitigating Divergences in Office Documents

Penghui Li cross-listed LLMs read Office documents through text extractors that parse the underlying Office Open XML (OOXML) files, and that text can differ from what a person sees on screen. Using static and dynamic analysis of widely used extractors, the authors find 19 OOXML channels where extractors read content hidden from humans and 12 where they drop visible content; extractors also ignore display formatting and reading order. From these gaps they build five attacks that make a model interpret a document differently from a human reviewer, without injecting any instruction, and every one of four chat assistants and four coding agents tested, including ChatGPT, Claude Code, and GitHub Copilot CLI, is misled by at least one. They propose Prism, a lightweight defense that rebuilds the human-visible view from the OOXML specification and also works against attack payloads it has not seen.

arXiv:2611.05806 PDF

Auditable Unlearning for LLMs

Piyush Tiwary, Anushka Srivastava, Prathosh A. P Machine unlearning methods are judged on forgetting, retained capability, and resistance to membership inference attacks, but not on whether the provider can show the user evidence that their data was deleted. The authors add auditability as a fourth criterion, defined as how well a provider-computed statistic separates forgotten data from retained data, and find that none of eleven unlearning methods meets all four criteria on TOFU, MUSE, and RWKU. They propose a two-stage add-on for any method: GapWiden, a hinge loss that widens the audit-statistic gap, and GapVeil, an inference-time rejection sampler that uses the base model as drafter. With it, auditability and leakage match the oracle for every method and forgetting approaches the oracle, while retained performance improves for nine of eleven methods but stays short of parity.

arXiv:2611.05866 PDF

How Far Can You Steer? Behavior-Capability Frontiers in Contrastive Activation Addition

Deepanshu Raj cross-listed Contrastive Activation Addition (CAA) steers language-model behavior by adding activation directions derived from contrasting examples, but it can degrade unrelated capabilities. The authors evaluate it as a behavior-capability frontier, choosing the steering setting that maximizes behavioral gain under an explicit capability-loss budget, across four instruction-tuned models and seven behaviors. Behavioral effects generalize well to held-out data, and joint behavior-capability survival rises from 37.5% at a 1% budget to 72.7% at 3%, with little further gain at 5%. Interventions that pass protected benchmarks can still harm untested capabilities or transfer unevenly to free-form generation.

arXiv:2611.06001 PDF

Reconciling Privacy with Provenance for Language Model Watermarks

Elise L. Han, Tatsunori Hashimoto cross-listed Users resist language model watermarks because fine-grained tracking threatens their privacy, and evasion such as paraphrasing weakens the signal. The authors reframe the goal from tracking individuals to estimating aggregate usage, and they design locally differentially private watermarks that provably limit how much evidence any single text contributes to detection. Their Red-Green Sequence (RGS) watermark is computationally efficient and estimates population frequencies accurately at strong privacy levels. On C4 it reaches 0.0019 mean squared error at ε = 1, and it stays accurate on an essay-writing task with only 100 samples.

arXiv:2611.06100 PDF

SafeStart: Mitigating the Overlooked Safety Cost of Active Expert Reduction in MoE LLM Serving

Erfan Baghaei Potraghloo, Saeid Shokoufa, Seyedarmin Azizi, Arya Fayyazi, Souvik Kundu, Mehdi Kamal et al. To serve Mixture-of-Experts (MoE) models faster, some methods route each token through fewer experts, and these methods are usually validated only on accuracy benchmarks. In vLLM on a single RTX A6000, keeping about two thirds of the selected experts sped up decoding by 15 to 28% and cost less than two points of MMLU accuracy. Under a jailbreak that tells the model not to refuse, however, harmful compliance rose by up to 18.4 points on Qwen1.5-MoE, and most of that increase arose during prompt processing and the first few generated tokens. SafeStart therefore runs all experts for the prompt and the first eight generated tokens, then switches to the reduced set through a fused kernel. It keeps at least 85% of the decode speedup while bringing the jailbreak-induced change to between -2.1 and +1.3 points.

arXiv:2611.06147 PDF

Mass Producing Training Data with Surprising Side Effects

Yaowen Ye, Jacob Steinhardt, Sarah Schwettmann, Cassidy Laidlaw Training data can change language model behavior far outside its domain (insecure code can cause broad misalignment, for example), but only a few dozen hand-found examples of this exist. The authors automatically search for datasets with surprising side effects (DSSEs) that elicit a target behavior, read like ordinary post-training data, and would surprise human auditors, scoring each objective with LLM judges and optimizing via evolution, reinforcement learning, and subliminal learning. The pipeline yields 6,575 DSSEs across 8 behaviors, many of which generalize across model families and sizes. The automated metrics track ML researchers' judgments, and frontier models predict the side effects better than humans do.

arXiv:2611.06324 PDF

Does Context Compression Preserve Refusal Alignment?

Anthony Nguyen, Wenjun Lin Context compression lowers inference cost by encoding inputs into compact representations, but it is unclear whether preserving meaning also preserves behaviors such as refusing harmful requests. The authors find that encoder-based compression systematically weakens refusal in instruction-tuned models despite high reconstruction fidelity, across model families and compression architectures, and mechanistic analysis traces this to weaker activation along the decoder's refusal direction. A lightweight inference-time intervention called Memory Steering, operating entirely in the compressed representation space, restores refusal rates to near baseline without retraining.

arXiv:2611.06380 PDF

A Bag of Tricks for Defending Deep Learning-based Face Recognition Pipelines against Backdoor Attacks

Quentin Le Roux, Marti Jimenez, Yannick Teglia, Philippe Loubet Moundi cross-listed This technical report addresses a threat model common in industry: a multi-stage face recognition system is trained in-house, but its training data comes from third parties that may be compromised and could plant backdoors through data poisoning. It presents three practical defenses. One runs at training time and detects poisoned identities during training or before deployment. The other two run at inference time and filter suspicious inputs before a response reaches a potential attacker. Each defense comes with its motivation, algorithm, and computational cost, and is evaluated on realistic face recognition workloads as actionable hardening guidance.

arXiv:2611.06412 PDF

Addition by Subtraction: Toward Mitigating Fragile Safety Alignment in Large Language Models

Hamid Dashtbani, Alexandra Dmitrienko LLM safety alignment can be sparse: pruning a few safety-critical neurons or removing a single refusal direction can undo it. Addition by Subtraction gradually removes the identified safety-critical neurons during fine-tuning while preference optimization rebuilds safe behavior in other parts of the model. The resulting models resist pruning attacks and also hold up better against refusal-direction ablation and many black-box jailbreaks, reducing attack success rates by up to 90% across four models from three vendors, while improving truthfulness and largely preserving general capabilities.

arXiv:2611.06553 PDF

Signed, Sealed, Leaked: Trust Beyond Signature for MCP Toolchains

Yichiet Aun (Universiti Tunku Abdul Rahman), Mohammed A. Alhartomi (University of Tabuk), Jing Jie Tan (National University of Singapore), Waqas Nadeem (Universiti Tunku Abdul Rahman), Adeb Salh (Universiti Tunku Abdul Rahman), Shankar Karuppayah (Bitranger et al. cross-listed Signing Model Context Protocol (MCP) tools proves who published them but not that they are benign. The authors build a signed MCP tool that leaks confidential data inside requests to an allowed destination, and Meta's MCPGuard-Dynamic fails to stop it in every trial. Their proposed germline execution runs each tool invocation in a short-lived copy forked from a data-free process and gives each copy only the data authorized for its destination, so concealed malicious code can disclose no more than that. Under stated assumptions they prove a noninterference property, and across 153 attack runs no unauthorized disclosure occurred, versus 5.9-74.5% of runs under baseline defenses, with invocations taking 10-49 ms.

arXiv:2611.06585 PDF

Beyond accuracy: what AI assurance needs to establish in public services

Louise Yorke cross-listed Combining a technical evaluation with a review of 37 UK and EU policy documents, the authors ask what public bodies must check before relying on AI-produced evidence. Eight GPT-4o configurations classifying 1,200 UK Prime Minister's Questions into 21 topics all reached 68-70% accuracy, yet 24.3% of questions received different labels across configurations and repeats, shifting topic shares and rankings. Configurations agreed more with each other than with human coders, and the share flagged for human review ranged from 0.9% to 41%. The policy documents mostly do not require reporting on unstable answers, shared errors, or selective evaluation.

arXiv:2611.06655 PDF

Measuring Risk, Not Harm: Detecting Psychological Harm Risk Factors in LLM Conversations

Nil-Jana Akpinar, Eugenia Kim, Natalie Isak, Amanda Minnich, Victoria Westerhoff, Matthew Dressman Psychological harm from LLM interactions cannot be monitored directly, because it can build up over time and depends on each user's circumstances. The authors instead detect observable AI behaviors that act as risk factors, defining eight such behaviors and building a synthetic multi-label corpus of 47,062 responses with 373,873 labels, plus targeted human annotations. They compare embedding-based classifiers, direct LLM judging, and a classifier-then-judge cascade, all of which work only on model outputs. On held-out human-labeled data the best cascade reaches macro F1 of 0.79, and on WildChat it reduces LLM-judge calls by 97.1% compared with direct judging; on DelusionEval, detections cluster within conversations, which supports aggregating them for longitudinal monitoring.

arXiv:2611.06806 PDF

Before You Continue: Evaluating Time-Sensitive Safety Awareness in Conversational Assistants

Peidi Dong, Yunze Xiao, Yu-Lin Tsai, Yichen Zheng, Heyuan Huang, Weihao Xuan et al. Users often mention a time-dependent constraint early in a conversation, such as when they took a medication, and do not repeat it in later requests. TASC (Temporal Action Safety in Conversation) tests whether assistants apply such constraints, with 3,418 inputs built from 128 templates plus 33 scenarios that separate calculating elapsed time from acting on it. Across 13 models, assistants account for the constraint in only 30.5% of cases where the proposed action breaks a rule, while rarely issuing unnecessary warnings. In 62.1% of cases models fail to apply the constraint even though they calculate the elapsed time correctly, and stating the timing explicitly fixes only about half of those failures.

arXiv:2611.06839 PDF

Persona Perspectives: Adopting Viewpoints Reorganizes Representations in LLMs

Austin C. Kozlowski, Avital Mintz, Madison Lee Monitoring tools such as linear probes and sparse autoencoders (SAEs) assume a feature is represented the same way in every context, but LLMs can adopt personas with different conceptual frames. Using over 9,000 common words and 33 semantic axes, the authors show that prompting a model to adopt a liberal or conservative persona systematically shifts both where words fall on the axes and how the axes align with one another, in directions consistent with the ideology. Probes, SAE-based classifiers, and hand-built word-representation classifiers for truthfulness and morality, all trained without a persona, produce politically slanted ratings on controversial statements when applied to persona-conditioned models. This suggests that persona adoption can undermine representation-based safety monitoring.

arXiv:2611.06848 PDF

8 more specialized papers

Unclassified 32

A local orchestration harness for molecular design with foundation models

Sakib Mostafa, Md Atik Ahamed, James Zou, Lei Xing, Md Tauhidul Islam cross-listed No summary available — see the abstract on arXiv.

arXiv:2611.03676 PDF

From Forecasting to Feedback: Evaluating Learned Dynamics for Tokamak Control

Guoyang Shi, Siqi Ding, Xuanhe Wang, Pei Guo, Changquan Yu, Jiale Hu et al. cross-listed No summary available — see the abstract on arXiv.

arXiv:2611.03690 PDF

All Paths All at Once: A Quantum Monte Carlo and Neural Committor Approach to Reaction Pathway Search

Chien-Chai Chang cross-listed No summary available — see the abstract on arXiv.

arXiv:2611.03699 PDF

Unsatz: Evaluating Scientific Discovery Without a Given Ansatz

Ethan Ji, Wei Wang, Yizhou Sun cross-listed No summary available — see the abstract on arXiv.

arXiv:2611.03713 PDF

EMBERS: Empirically Motivated Robustness Testing of Deep Research Systems via Defect-Guided Fuzzing

Shuochuan Li, Zan Wang, Dong Wang, Yuhang Jiang, Huajin Cui, Chen Zhang et al. cross-listed No summary available — see the abstract on arXiv.

arXiv:2611.03723 PDF

Does the Trust Boundary Hold over Time? A Long-Horizon Benchmark for Receiver-Directed Federated Learning

Ting Xu, Henry Leung No summary available — see the abstract on arXiv.

arXiv:2611.03724 PDF

DualClusterNet: Personalized Pattern-Space Composition for Next-POI Recommendation

Ziyang Su, Guanjie Zheng, Jilin Hu, Linghe Kong cross-listed No summary available — see the abstract on arXiv.

arXiv:2611.03726 PDF

Auditing a Model-Hub Benchmark: Data Quality, Training Budgets, and Side Information in HuggingBench

Michal Podstawski No summary available — see the abstract on arXiv.

arXiv:2611.03727 PDF

Complete Donor-Set Guidance for Frozen Diffusion Docking Models

Hyunwoo Kim, Insoo Park, Ziseok Lee, Wooyeol Lee, Yoorim Gang, Kyungsu Kim cross-listed No summary available — see the abstract on arXiv.

arXiv:2611.03731 PDF

Minimal Superior Penalties: Characterising and Calibrating Decision-Aware Proper Scoring Rules

Rouhollah Ahmadian No summary available — see the abstract on arXiv.

arXiv:2611.03734 PDF

HyVox: Rethinking Low-Power Voice Intelligence towards Neuromorphic-von Neumann Edge Computing

Snehasis Addy, Eduardo Calle Ortiz, Prashant Shenoy, Phuc Nguyen cross-listed No summary available — see the abstract on arXiv.

arXiv:2611.03735 PDF

SSPE: Robust Sequential Staypoint Extraction in Degraded Trajectory Data

Yueyang Liu, Lance Kennedy, Andreas Z\"ufle No summary available — see the abstract on arXiv.

arXiv:2611.03737 PDF

RATSIFT: Resolution-Adaptive Transformer for Solar Irradiance Forecasting and Temporal Downscaling

Jesto Peter, Timothy M. Hansen, Hossein Moradi Rekabdarkolaee No summary available — see the abstract on arXiv.

arXiv:2611.03738 PDF

CaRe-KT: Continuity-Aware Retrieval-Augmented Reasoning for Knowledge Tracing

Mehrnoush Mohammadi, Kamal Berahmand, Mohammad Morsali, Shazia Sadiq, Wojtek Tomaszewski, Hassan Khosravi No summary available — see the abstract on arXiv.

arXiv:2611.03744 PDF

Honest Small Decision Models for Finance: Scale, Specialization Collapse, and Mixed-Corpus Remediation

Ajitabh Kumar No summary available — see the abstract on arXiv.

arXiv:2611.03755 PDF

Language Models Can Benefit from Domain-Specific Inductive Biases

Damien Teney, Liangze Jiang, Zachary Shinnick, Hemanth Saratchandran, Simon Lucey No summary available — see the abstract on arXiv.

arXiv:2611.03756 PDF

The role of injected noise in diffusion model generalization

Vincent Emmerling, Felix G\"unther, David Kappel No summary available — see the abstract on arXiv.

arXiv:2611.03764 PDF

Quantifying out-of-distribution data severity with Wasserstein distances: benchmarking large language models for alloy property prediction

YuJie Liu, Dezhen Xue, Jun Sun, Xiangdong Ding, Zhibin Gao, Turab Lookman No summary available — see the abstract on arXiv.

arXiv:2611.03767 PDF

Multi-layer Indexer: Overlapping Selection and Attention in DeepSeek Sparse Attention

Serdarcan Dilbaz, Arjun Krishna, Kiran Kamble, Daniel M. Bikel No summary available — see the abstract on arXiv.

arXiv:2611.03768 PDF

What Malware Actually Does: An Empirical Analysis of Post-Infection Behavior Across Malware Families

Emmanuela Andam, Emanuel S. Grant, Naima Kaabouch cross-listed No summary available — see the abstract on arXiv.

arXiv:2611.03774 PDF

Ground-Wing: A Parametric 3D Wing-in-Ground-Effect Dataset for Benchmarking Neural Surrogate Models

Andrea Cenzato, Lukas Karch, Andre Luckow, Yoran Pichon, Francesco Pasqua, Abouzar Ghasemi et al. No summary available — see the abstract on arXiv.

arXiv:2611.04571 PDF

Spectral Quantum Memory for Implicit Neural Representations

Daniel A. Moroz-Broitman, Eliya Nachmani cross-listed No summary available — see the abstract on arXiv.

arXiv:2611.04574 PDF

Conditional independence and graphical models for rankings

Aritz P\'erez cross-listed No summary available — see the abstract on arXiv.

arXiv:2611.04586 PDF

Estimating "Realized" Skewness using Convolutional Neural Network

Haozhe Jiang, Ostap Okhrin, Michael Rockinger cross-listed No summary available — see the abstract on arXiv.

arXiv:2611.04598 PDF

SCANO: Restoring Boundary-Condition Completeness in Sampled Meshes with a Sliced-Cross-Attention Neural Operator for 3D Thermal Fields

Tianyu Li, Zhiwei Cao, Qingang Zhang, Ruihang Wang, Binyang Song, Yonggang Wen No summary available — see the abstract on arXiv.

arXiv:2611.04603 PDF

CAPSkip: Cancellation-Aware Prefix Skipping for Selective Backpropagation

Heejin Choi No summary available — see the abstract on arXiv.

arXiv:2611.04611 PDF

Policy Network Diffusion

Jiate Liu, Deheng Ye No summary available — see the abstract on arXiv.

arXiv:2611.04614 PDF

Self-Privileged Fine-Tuning

Yangrong Hu, Xueyu Zhou, Yue Wu, Jian Huang No summary available — see the abstract on arXiv.

arXiv:2611.04632 PDF

Does LinUCB Need Valid Confidence Sets?

Ahmed Ben Yahmed, Marc Abeille, David Janz, Cl\'ement Calauz\`enes No summary available — see the abstract on arXiv.

arXiv:2611.04641 PDF

Exploiting Biases in Attention for Transformers

Charles Herr, Gabriele Spadaro, Enzo Tartaglione No summary available — see the abstract on arXiv.

arXiv:2611.04657 PDF

Leakage-free evaluation of residual graph convolutional networks for network intrusion detection

Zahid Iqbal No summary available — see the abstract on arXiv.

arXiv:2611.04660 PDF

1 more specialized paper

Vision 31

Do We Still Need Rotation-Invariant Architectures? A Statistical Robustness Benchmark for Point Cloud Classification

Abdulla Nasir Chowdhury, Gulpi Qorik Oktagalu Pratamasunu, Kazuhiro Fukui cross-listed Point cloud classifiers often build rotation invariance directly into the architecture. The authors benchmark 10 architectures, including PointNet++, DGCNN, RIConv++, PointTransformer and VN-DGCNN, across 5 domains under pose, noise and occlusion, and evaluate them with confirmatory hypothesis tests instead of point estimates. Rotation-invariant (RI) models are the most robust to pose but among the least robust to noise. When every architecture is retrained with a simple invariance loss and tested on ModelNet40-C, the gains are narrow but reliable (for example, +10.5 points on occlusion and +13.4 on lidar), and the best standard backbone and the best RI architecture finish within 2.5 percentage points of each other on unseen corruptions.

arXiv:2611.03952 PDF

What Remains in Sight? Autoregressive Video Decoding as Representation-Guided Context Rewriting

Lixuan He, Jihyeon Je, Yang You, Leonidas Guibas cross-listed Autoregressive long-video generators usually decide which past frames the decoder sees by implicit rules such as recent windows or compressed caches. ReCR treats this choice as an explicit test-time decision and needs no training. It uses the decoder's internal representations to select past context units that match the global rollout state, avoid repeating the recent continuation, cover temporal boundaries, and span varied spatial layouts, then maps them onto a compact causal axis. On long-horizon text-to-video benchmarks, ReCR consistently improves multiple autoregressive backbones under matched visible-token budgets, giving more stable long-video decoding.

arXiv:2611.04164 PDF

PEEK: One-Step-Look-Ahead Exposure Bias Correction for Diffusion Sampling

Yunwei Bai, Ying Kiat Tan, Tsuhan Chen cross-listed Diffusion models suffer from exposure bias: at sampling time the denoiser receives inputs that differ from the data it was trained on, which can produce blurry or misshapen images. PEEK is a training-free, deterministic sampler plugin that takes a provisional one-step update and evaluates the denoiser at that look-ahead state. It then interpolates this cleaner prediction with the next-step prediction to form the next noisy state. An efficient cached search picks the interpolation schedule, with one-time calibration taking at most 2.2 minutes per model. Across 8 pixel-space and latent-space models in unconditional, class-guided, and text-guided settings, PEEK beat baselines with the same number of function evaluations, with FID improvements of up to 49.3%.

arXiv:2611.05407 PDF

Discarded or Unreachable? Diagnosing and Repairing Spatial Relation Failure in Vision--Language Similarity

Mahdi Mostajabdaveh, Zirui Zhou cross-listed Contrastive vision-language models perform near chance on left/right relations, and the analysis traces this to two different failures that need different fixes. In some cases the relation survives in the pooled embedding but cannot be reached through ordinary image-text similarity, which the authors call an access failure. In others, pooling throws away the relation's direction while the patch-level features keep it, which they call a pooling failure. A parameter-free, query-conditioned spatial moment recomputes the geometry from the patch grid and stores it alongside the native embedding. This raises left/right pairwise accuracy from about 0.52 to 0.81–0.96 on a 118,287-image index, at under a millisecond per batched query, without changing rankings for queries that mention no spatial relation.

arXiv:2611.05565 PDF

Learning Hierarchical Patch Splitting Policies for Faster Vision Transformers

Aditya Gupta, Jean Sebastien Dandurand, Kai Qiu, Rohan Choudhury, Laszlo A. Jeni cross-listed Vision Transformers (ViTs) slow down as input token sequences grow, and earlier adaptive patch-sizing methods decide which regions to coarsen using noisy pixel heuristics such as local entropy. SPLIT (Split Policy Learned Image Tokenization) instead learns a hierarchical patch-splitting policy driven by what the model finds task-relevant. It matches full-resolution ViT classification accuracy with 30% fewer tokens. It also stays competitive on dense prediction benchmarks COCO and ADE20K.

arXiv:2611.05742 PDF

Touch\'{e}World tech report : Towards Online Unified Camera--Hand Exploration in Generative Worlds

Jiaao Yu, Yanjian Zhang, Haoyang Jiang, Guillaume Wisniewski, Nadi Tomeh, Shijie Jia cross-listed Video world models let users explore generated scenes by moving the camera, and ToucheWorld adds continuous hand gestures as a second control channel for interacting with people and objects. The authors assemble joint camera-and-hand training data from heterogeneous egocentric videos through annotation completion, consistency filtering, and temporal alignment. They then adapt a pretrained camera-controlled world model with hand-skeleton and hand-motion conditioning using parameter-efficient adaptation. Causal adaptation and few-step distillation enable chunk-level online generation, so users can adjust their controls in response to what is generated. A new benchmark, ToucheBench, evaluates hand-motion following, camera-hand coordination, online interaction, and object response.

arXiv:2611.06061 PDF

NextGAN: Towards Simple and Scalable GAN Training for Image and Video Generation

Xinhao Lou, Dongye Xie, Peng Sun, Chenyi Zhuang, Kaicheng Yu, Tao Lin cross-listed Conventional adversarial losses supervise the critic's output values but never directly specify the gradients the generator actually learns from, which complicates GAN training for one- and few-step image and video generation. NextGAN trains a scalar critic so that its input gradient regresses onto displacements between independently paired reference and generated features, sampled at interpolated points in frozen representation spaces. For a fixed generator this is equivalent to an adversarial objective with a gradient-energy penalty, and it adapts pre-trained generators of up to 14B parameters without auxiliary losses, teacher scores, or trajectories. On ImageNet 256×256 it reaches one-step FIDs of 2.87 with pixel-space JiT-B and 1.54 with latent-space DiT-XL/1, and a single Wan2.1-1.3B video model supports one-, two-, and four-step sampling, scoring 83.13 on VBench at two steps, ahead of rCM and AnyFlow.

arXiv:2611.06270 PDF

Don't Replace, Complement: A Detector-Agnostic Subspace Signal for Near-OOD Detection

Wonseok Shin, Songkuk Kim cross-listed Post-hoc out-of-distribution (OOD) detectors work well on far-OOD inputs but poorly on near-OOD ones, which activate the same class-specific feature directions as in-distribution data. PLUS (Per-class Low-dimensional sUbspace Signal) picks each class's most active sparse-autoencoder neurons, builds a low-dimensional subspace from their decoder directions, and scores inputs by whitened Mahalanobis distance within the predicted class's subspace. Added to existing detectors with a single fixed weight calibrated only on in-distribution data, it improved near-OOD AUROC in all 126 detector-backbone combinations tested across 18 detectors, while largely keeping far-OOD performance.

arXiv:2611.06688 PDF

LongBanana: An Expert-Verified Benchmark for Long-Context Multi-Reference Image Synthesis

Haoxiang Cao, Yuxuan Zhang, Penghui Du, Bo Li, Jiajiong Cao, Yingying Fan et al. cross-listed LongBanana is an expert-designed and expert-verified benchmark for long-context multi-reference image synthesis, where a model must retrieve scattered visual evidence, reject distractors, bind entities and relations, and compose a single coherent image. It has 90 tasks, each with about 32 reference images, 20 constraints, and 50 evaluation items derived from an entity-relation graph, for 4,670 checklist items covering content fidelity, relational binding, and global constraint satisfaction; a double-labeled subset shows 89.4% exact inter-annotator agreement. The accompanying LongBanana-Gen agent-harness framework shows that minimal agent harnesses beat direct generation, while a ReAct-Plan-Exec-Reflect (RPR) workflow gives gains that depend on the harness by improving evidence grounding and repair.

arXiv:2611.06856 PDF

22 more specialized papers

Multimodal 26

PianoBench: Evaluating Musical Perception and Reasoning in Language Models

Louis Bradshaw, Irmak Bukey, Daniel Chenyu Lin, John Thickstun, Chris Donahue cross-listed PianoBench tests which musical features language models can perceive and reason about, using matched audio and symbolic versions of piano performances. It is built from a new dataset of more than 20,000 hours of music transcripts interleaved with spoken commentary from instructional piano videos. Open-weight audio-language models reach at most 52.8% accuracy, barely above a no-audio control, while frontier models using chain of thought reach up to 93.2% when given the music as text transcripts, which points to a gap between hearing musical features and reasoning about them once they are written out. Mid-training on the interleaved instructional data raised Qwen 3.5 9B Base's zero-shot accuracy from 42.8% to 56.0%.

arXiv:2611.03839 PDF

Perturb, Repair, Verify: Self-Play Vision-Language Verifiers for Compositional Understanding

Zijun Huang, Yingjun Du, Cees G. M. Snoek cross-listed Vision-language verifiers judge whether an image matches a text description, but they are usually trained on static supervision from stronger judges or curated hard negatives. PROVE (Perturb-and-Repair Optimization for Vision-Language Evaluation) trains a verifier through self-play. Starting from a correct image-caption pair, a Mutator perturbs objects, attributes, counts or relations in the caption, the verifier judges it and gives feedback, and a Repairer uses that feedback to restore a faithful caption for re-verification. All three roles share one multimodal backbone, so the structured rewards from the two verification rounds also improve the verifier itself. Experiments show that PROVE improves fine-grained compositional verification across model scales and families, with limited, benchmark-dependent changes on broader multimodal tasks.

arXiv:2611.05158 PDF

A Training-Free Socket for DNA in Frozen Language Models

Arash Fatehi cross-listed Connecting a DNA foundation model to an LLM usually means training a connector, and often the LLM as well. The authors show that a single closed-form ridge regression, which they call a plug, can map frozen Evo 2 features into the host LLM's own early-layer representation of a gene's symbol, name, and description, with no gradients through the host. On GeneHop, a five-way multiple-choice benchmark split by homology, real DNA beat shuffled DNA by 16.7 to 26.1 points on all four hosts, including Qwen3.6-27B and OLMo-2-7B. A layer-selection rule chosen on three hosts predicted the working layer of the fourth, and a text-only linear map lets one plug serve other hosts at a cost of at most 0.9 points.

arXiv:2611.05401 PDF

When Does Sharing Work? Scaling a Shared SigLIP2 Across Image, Video, Audio, and Audio-Visual Representations

Xiaofei Sun, Peng Wang, Huayao Liu, Xiaoxiao Xu, Jun Song, Deqing Li et al. cross-listed Image, video, and audio encoders are usually trained separately and aligned only afterwards to a frozen anchor space. Omni-Encoder instead processes all four input types (image, video, audio, and audio-visual) in a single SigLIP-2 Vision Transformer with fully shared parameters. A 1D-convolutional tokenizer turns mel-spectrograms into patch-like audio tokens, and a timestamp-based multimodal rotary position embedding places frames and audio segments in one spatiotemporal coordinate system. Training alternates batches from each modality and accumulates their gradients into a single update, which the authors found central to stable optimization. The model is competitive on 16 benchmarks with one parameter set, and joint training improved every modality over its separately trained counterpart.

arXiv:2611.05406 PDF

RECAP: Context-Aware Prediction Recycling with Request-Local and Persistent Memories for Multimodal Speculative Decoding

Hengjie Zhu, Dayan Wu, Xinze Liu, Peng Fu, Zheng Lin, Weiping Wang Speculative decoding can be made training-free by reusing a model's own past predictions as drafts, but in vision-language models, indexing those predictions by the last token alone mixes up predictions made under different images and contexts. RECAP stores predictions under one-, two-, and three-token keys: memories tied to the current request implicitly account for its image, while token-level memory persists across requests for coverage. Retrieved predictions are weighted by key specificity and confidence, assembled into a candidate tree by best-first search, and checked in a single target-model pass. Without training or a separate draft model, it reaches mean speedups of 1.66x on Qwen2.5-VL-7B and 1.56x on LLaVA-1.6-7B, ahead of Token Recycling at 1.51x and 1.37x.

arXiv:2611.05894 PDF

Modality-Aware Expert Pruning for MoE-Based Multimodal Large Language Models

Donggeun Kim cross-listed In Mixture-of-Experts multimodal LLMs, expert weights make up over 90% of model size, and existing pruning methods score expert importance the same way for every token, ignoring that experts may serve vision and language differently. MAEP (Modality-Aware Expert Pruning) is a training-free method that separates expert importance by modality in one forward pass. It uses a redistribution-aware output-change metric, treats low-attention vision tokens as a pruning signal, and treats text-side importance as a signal for which experts to keep. On three MoE multimodal LLMs, including InternVL3.5-30B-A3B and Qwen3-VL-30B-A3B, it beats modality-agnostic baselines on most multimodal benchmarks and shrinks a roughly 58 GB model enough to fit on a single 40 GB GPU.

arXiv:2611.05927 PDF

UniCompress: From Multi-LoRA Experts to One Compact Model for Edge Intelligence

Mingkang Liu, Peipei Jiao, Weiming Meng, Tianyuan Liu, Lichen Pang, Junyan Li Edge and embodied systems often run a separate adapter for each skill on top of a high-precision vision-language model (VLM), which is costly in memory and management. UniCompress merges multi-task expertise into a single adapter-free model, starting from a BF16 Qwen3-VL-4B teacher and producing an INT2 student through GPTQ initialization and teacher-guided quantization recovery. It quantizes both the vision encoder and the language backbone. Weight storage falls from 8.9 GB to 2.1 GB, and the INT2 macro task score is 34.3 points above plain GPTQ-INT2 and within 0.9 points of the BF16 teacher across four embodied tasks, including mobile-GUI action prediction.

arXiv:2611.06075 PDF

Overview of the NTCIR-19 SciClaimEval Task: Cross-modal Claim Verification in Scientific Papers

Xanh Ho, Andr\'{e} Greiner-Petter, Florian Boudin, Yun-Ang Wu, Sunisth Kumar, Tian Cheng Xia et al. SciClaimEval, a pilot task at NTCIR-19, asks systems to verify claims from 180 papers in machine learning, NLP, and medicine against those papers' tables and figures. Each piece of evidence comes in a supporting version and a minimally modified non-supporting version. Fourteen teams submitted 177 runs, mostly multimodal large language model pipelines built on prompting rather than fine-tuning, and open-weight models matched proprietary ones. The best systems reached 98.1% pair accuracy, but all six top systems relied on the contrastive pairs, and the best system that judged each claim on its own scored 24.3 points lower. Verifying a single claim against its evidence therefore remains an open problem.

arXiv:2611.06090 PDF

Prism: Bidirectional Self-Teaching for Unified Multimodal Models

Peng Zhang, Qian Chen, Xianyin Zhang, Yimin Tian, Ruite Fei, Dongyi Lv et al. cross-listed Unified Multimodal Models (UMMs) handle both visual understanding and generation, but existing post-training either treats understanding as a fixed teacher for generation or relies on ground-truth trajectories. Prism splits the model's own interleaved rollouts into a text chain and an image chain, assesses the reliability of each, and builds training targets from those reliability states, with no external verifier or annotated corrections. On SenseNova-U1, GenEval rises from 85.8 to 89.2 and DPG from 86.1 to 87.3, while MME and MMVP understanding scores also improve, with consistent gains on BAGEL.

arXiv:2611.06343 PDF

AmbiHTMLBench: A Benchmark for HTML-based Visually Rich Document parsing

Boyu Niu, Yingjie Xiong, Shurui Li, Linhan Pang, Yuntong Zhang, Han Wang et al. Document parsers usually output Markdown, which loses spatial structure, table geometry, and visual hierarchy. The authors argue that HTML preserves these better for layout-heavy documents and introduce AmbiHTMLBench, in which a model must reconstruct executable HTML from 1,115 real-world pages. Predictions are scored two ways: an objective protocol that aligns rendered blocks to measure table fidelity, text error, and two-dimensional layout consistency, and a reference-anchored protocol that uses LLM judges aligned with human preferences. Across ten multimodal systems, the best objective score is 0.8092, but the objective and human-perceived winners differ, and a small fine-tuned expert model performs well on the task.

arXiv:2611.06449 PDF

Trick-or-Treat: Adversarial Evaluation of Multimodal Medical Reasoning with Synthetic Images

Oishi Banerjee, Eujin Lee, Sung Eun Kim, Alexandra N. Willauer, Christine Y. Zhou, Zhizhou Yang et al. cross-listed Medical datasets rarely pair different image types, which makes cross-modal clinical reasoning hard to test. Trick-or-Treat pairs real chest radiographs with synthetic clinical photographs that either support the same diagnosis or show an unrelated finding, while holding the person and photo setting constant to reduce confounding. Across three frontier models, hallucinated photograph findings rose to 45-57.5% in mismatched pairs, versus 25-37.5% for the same photos shown alone, with models typically misreading the photo to fit the radiograph's diagnosis.

arXiv:2611.06631 PDF

EvoWorld: Towards Out-of-Sight World Evolution in Generative Video Models

Yequan Bie, Dengyang Jiang, Zhixuan Chen, Xiaozhuan Liang, Haoxuan Che, Quande Liu et al. cross-listed Generative video models render plausible pixels but have no explicit way to track world state or reason about what an event should cause, so their outputs are often static or causally inconsistent. The authors define a new task, Out-of-Sight Causal Evolution, in which a model must render an event, infer its consequences, and show a causally plausible outcome even for parts of the world outside the current view. Their baseline, EvoWorld, pairs a vision-language model that acts as a "world brain", observing generated segments and reasoning about state transitions, with a segment-autoregressive video renderer. Together with the new EvoWorld-Data and EvoWorld-Bench, experiments show substantially better event-conditioned causal evolution than recent video models without losing visual quality or consistency.

arXiv:2611.06748 PDF

14 more specialized papers

Robotics 22

Latent Policy Barrier: Learning Robust Visuomotor Policies by Staying In-Distribution

Zhanyi Sun, Shuran Song cross-listed Visuomotor policies learned by behavior cloning suffer from covariate shift: small departures from the expert's trajectories add up until the robot fails. Latent Policy Barrier (LPB), inspired by Control Barrier Functions, treats the latent embeddings of expert demonstrations as an implicit boundary between safe in-distribution states and unsafe out-of-distribution (OOD) ones. A base diffusion policy is trained only on expert data, while a separate dynamics model is trained on both expert and suboptimal rollouts. At inference time, the dynamics model predicts future latent states and steers them back toward the expert distribution. In simulation and on real robots, this improves robustness and data efficiency without any extra human corrections or annotations.

arXiv:2508.05941 PDF

Behavior Graphs: Unsupervised Graph Structures for Interpreting Robot Deployment Data

Erik Bauer, Jakob Thumm, Christopher Agia, Rohan Sinha, Chen Xu, Haruki Nishimura et al. cross-listed Deployed robots generate large volumes of high-dimensional rollout data that is hard to analyze when deciding what training data to collect next. Behavior graphs compress a set of policy rollouts into one interpretable directed graph by grouping similar behavior modes across rollouts, linking them in time, and connecting them to outcomes. In a user study on three real-world tasks, 16 of 18 participants preferred behavior graphs to watching rollout videos, reached comparable conclusions, and reported about 40% lower time pressure and 38% less effort. Across 43 datasets, the graphs also predict task progress better than the strongest baseline, improving rank correlation by 22.9% and cutting mean absolute error by 53.5%.

arXiv:2611.00405 PDF

AIRoA-0: Long-Horizon, Precise, Cross-Embodiment Manipulation from Handheld Data

Yueh-Hua Wu, Kei Ota, Hanlong Li, Hayata Yuasa, Jason Wang, Kengo Ikeuchi et al. cross-listed AIRoA-0 is a bimanual manipulation system trained on 16,151 hours of demonstrations recorded with YUBI, an open-source handheld gripper, without collecting any robot data, teleoperation, or reinforcement learning. Using only a camera on each wrist, it learns which parts of each view matter, predicts which stage the task has reached, and decodes actions as a continuous stream. It carries out long-horizon tasks such as sorting small parts and swapping a robot vacuum's mop pads, at a site where none of its training data was recorded. Across ten tasks it averages 78% success, against 27% for the strongest of four other vision-language-action models post-trained on the same demonstrations. It runs on several robot types that mount the gripper, and the dataset, weights, and post-training recipe are being released.

arXiv:2611.03815 PDF

Hybrid Particle-Voxel Dynamics for Multimodal 3D World Simulation

Jimmy Wang, Yifan Chou, Kaifeng Zhang, Yixuan Wang, Binghao Huang, Tao Yu et al. cross-listed Many action-conditioned world models for robotics predict future frames in 2D image space, which makes it hard to stay consistent across viewpoints and over long interactions. 3D World Simulator compresses point clouds into compact voxel latents with a hybrid particle-voxel autoencoder and predicts how those latents evolve under robot actions, jointly producing appearance, geometry and tactile signals from one 3D representation. It supports interactive long-horizon multimodal prediction at 50 Hz with rendering from arbitrary viewpoints. The model is used to generate data for training and evaluating policies, including from wrist-mounted and novel camera views, and it scales to model-based planning with unseen objects and robot configurations.

arXiv:2611.04492 PDF

CRONOS: Benchmarking Non-Episodic Reinforcement Learning for Robotic Manipulation

Po-Yi Wu, Djengo Cyun-Jyun Fang, Dian Cheng, Tsung-Wei Ke cross-listed Real-world reinforcement learning (RL) for robots is held back by the need for frequent human scene resets, and non-episodic RL has mostly been studied with small policies trained from scratch. CRONOS is a simulation benchmark for RL fine-tuning of large pre-trained vision-language-action (VLA) models without episodic resets, using high-fidelity physics and multiple tasks that share one scene. Naive fine-tuning fails because object configurations and end-effector poses drift, but targeted reset allocation and correcting biases in the pre-trained models mitigate this. The authors also find that non-episodic training improves long-horizon manipulation and generalization to unseen object configurations and task sequences.

arXiv:2611.05080 PDF

GUILD: Guided Iterative Data Collection for Imitation Learning

Shivin Dass, Kevin Nguyen, Rohan Yelandur, Junhong Xu, Roberto Mart\'in-Mart\'in cross-listed Robot imitation learning depends on expensive real-world teleoperation that is usually unguided, so operators collect trajectories that contribute little to final policy performance, and existing curation methods only filter data after the fact. GUILD closes the loop during collection by alternating between proposal and evaluation: vision-language models supply semantic priors about which environments, object configurations, and goals to collect next, and predictive data attribution then estimates how the newly collected trajectories actually affect policy performance on the target deployment setting, feeding that signal back into the next round. In simulation and on real robots, guided collection improves sample efficiency and deployment performance over passive collection, several baselines, and experienced roboticists.

arXiv:2611.05368 PDF

PolicyIK: High-Precision Inverse Kinematics for Robot Arms with Reinforcement Learning

Yuhan Liu, Xinqi Liu, Zixuan Dong, Keith W. Ross, Shenji Wan cross-listed Inverse kinematics turns end-effector pose commands into joint motion and underpins many vision-language-action policies, but redundancy, singularities, and hand-tuned controllers make precise solutions laborious. PolicyIK is a goal-conditioned reinforcement learning policy that updates joint-position references from pose feedback and observes its accumulated reference relative to the measured joints, so it can hold the offset a position controller needs at rest, with two-scale pose errors and an error-dependent action gain to resolve millimeter residuals, plus progress and dwell rewards to both reach and stay inside tolerance. A single attempt by the trained PPO policy keeps mean final-second error within 1 mm and 0.2 degrees on 93.5% of held-out simulated tasks, rising to 98.9% with a deployment protocol; trained directly on a physical arm with off-policy reinforcement learning it beats all four classical inverse-kinematics controllers under matched time budgets, and it executes the end-effector commands of the π0.5 vision-language-action policy with 95.7% success in simulated pick-and-place.

arXiv:2611.05372 PDF

Jev with Foresight: Safe Language-Guided Autonomous Driving via a Language World Model

Yansong Qu, Jiancong Chen, Samuel Labi cross-listed LLM driving agents can follow passengers' commands, but they spend seconds generating text and must infer what each action will lead to from coordinates alone. Jev-AD pairs Jev, a typed decision model that answers predefined questions with probabilities instead of generated text, with a language world model that predicts how nearby vehicles will move over the next two seconds under each candidate action. Code converts these predictions into explicit consequences that Jev compares. In simulated highway merging, this foresight cut the collision rate from 26.0% to 2.4% at 0.25 seconds per decision while fulfilling the passenger's intent in 93% of free-form command trials. On replayed real-world NGSIM traffic, collisions fell from 48.2% to 2.8%.

arXiv:2611.05448 PDF

MESA: An Evaluation Framework for Compositional, Semantic, and Spatial Generalization in Robotics

Albert Wilcox, Frank Chang, Nhi Nguyen, Aishani Chakraborty, Jeremy A. Collins, Vaibhav Saxena et al. cross-listed Benchmarks for robot manipulation policies tend to be either saturated, too loosely controlled, or dependent on training with messy open-source data. MESA is an evaluation framework with five suites that separate in-distribution performance from generalization along four axes: spatial configuration, object instance, object category, and subtask composition. Its tasks and demonstrations are generated with MESA-Gen, an improved version of MimicGen. It ships a single-arm suite and BiMESA, the first large-scale benchmark for bimanual generalization. Among seven baselines spanning diffusion policies, VLM-based approaches, and pretrained vision-language-action models, the best reaches only 52% on MESA and 49% on BiMESA, and MESA's rankings broadly agree with RoboArena scores.

arXiv:2611.05567 PDF

OdorSpaces: Visual-Olfactory Embodied Navigation in 3D Environments

Changyeon Lee, Kelin Yu, Zheng Wei, Shuna Ni, Ruohan Gao cross-listed Embodied AI has made little use of smell. OdorSpaces is a simulation platform and benchmark for navigation using both vision and smell. It extends Habitat-Sim with physics-based 3D odor transport and dispersion computed by the Fire Dynamics Simulator (FDS). The benchmark covers two source-seeking scenarios: locating an invisible gas leak and locating an early-stage fire. A multimodal reinforcement learning policy that combines visual and olfactory observations over time outperforms a diverse set of baselines and serves as a reference for future work.

arXiv:2611.05738 PDF

RACE: Calibrated Cumulative Scoring for Rollout-Free Robot Policy Checkpoint Selection

Changyeon Kim, Yekyung Nah, Kimin Lee cross-listed Picking the best checkpoint of a learned robot policy usually needs costly real-world trials, and averaged validation loss, the cheap alternative, often ranks checkpoints out of line with real success. RACE scores checkpoints against held-out successful trajectories by calibrating each action deviation against thresholds learned from how much actions vary at similar states. It then combines the scores as a running product, so a large early deviation reduces the weight of later predictions. In simulation it beats baselines across validation sources while several baselines degrade sharply when the validation set changes; on real robots its rankings match deployment success across the policy backbones and tasks tested.

arXiv:2611.05838 PDF

Interactive World Simulation with Cosmos for Physical AI

NVIDIA, :, Anis Ahmad, Jon Allen, Malik Aqeel Anwar, Jesse Archer et al. NVIDIA adapts its Cosmos world foundation models into a shared framework for real-time interactive world simulation, where visual observations stream back in response to actions and scene controls. The framework can either predict the visual consequences of actions directly or render observations from physical state kept by an external simulator. Demonstrations cover camera navigation, robot manipulation, controllable video transfer, and multi-view autonomous vehicle (AV) simulation, and code and model checkpoints are publicly released as a base for building virtual environments to train and evaluate embodied agents.

arXiv:2611.06627 PDF

Dexterous Robotic Hands in the Foundation-Model Era: A Full-Stack Survey of Hardware, Data, Cross-Embodiment Transfer, Learning, and Evaluation

Ji Pei cross-listed This survey covers dexterous robotic hand research from January 2024 to November 2026, organized into five layers: hardware, data collection, cross-embodiment transfer, learning and foundation models, and evaluation, with tactile sensing and human hand-object understanding as cross-cutting themes. It concludes that the value of human data depends on kinematic, visual, and contact alignment rather than volume, that shared action representations do not remove embodiment-specific limits on force and latency, and that task completion alone does not establish physical correctness. It lists eleven open problems and proposes reporting standards for more measurable progress.

arXiv:2611.06724 PDF

Perceptive Humanoid Motion from Simulation

Siyuan Hu, Abulikemu Abuduweili, Yue Meng, Shipeng Liu, Quentin Leboutet, Chen Chen et al. cross-listed Generalist humanoid robots lack paired data linking camera images and language to low-level joint commands, and collecting such data by teleoperation takes about an hour of human time per hour of data. HALO instead synthesizes vision-language-motion data in a thousand procedurally generated simulated homes, so the dataset grows with compute rather than human labor. The resulting system learns skills such as sitting on chairs, navigating to named objects, crawling under tables, and pushing boxes, and chains them to complete long-horizon tasks. Without seeing a real home during training, it transfers zero-shot to the physical world with a 91% success rate across 360 trials in six unmapped real environments.

arXiv:2611.06908 PDF

8 more specialized papers

Reinforcement Learning 16

Preserving the Privacy of Reward Functions in MDPs through Deception

Shashank Reddy Chirra, Pradeep Varakantham, Praveen Paruchuri cross-listed An agent whose actions can be observed may give away its private reward function, for example where patrols reveal where endangered animals are to poachers. Observers can recover those preferences with inverse reinforcement learning (IRL), and existing differential-privacy methods do not guarantee a minimum expected reward. Drawing on the theory of deception, the authors first show that existing dissimulation-based methods (which hide the truth) leak a lot of private information. They then propose an RL-based planning algorithm built on simulation, meaning it actively shows a false preference, that protects reward privacy while guaranteeing a minimum expected reward. On several benchmark problems it outperforms prior methods at keeping the reward function private.

arXiv:2407.09809 PDF

On Discovering Algorithms for Adversarial Imitation Learning

Shashank Reddy Chirra, Jayden Teoh, Praveen Paruchuri, Pradeep Varakantham cross-listed Adversarial Imitation Learning (AIL) methods have two parts: Density Ratio estimation, where a discriminator compares expert and policy state-action occupancy, and Reward Assignment (RA), which turns that ratio into a training reward. AIL is often unstable, and the RA functions in use have been designed by hand from divergence-minimization objectives. The authors instead search the space of RA functions with an LLM-guided evolutionary framework, scoring each candidate by the performance of the imitation policy it produces. The result is Discovered Adversarial Imitation Learning (DAIL), described as the first meta-learned AIL algorithm. It generalizes to unseen environments and policy-optimization algorithms and outperforms the best human-designed baselines, and the authors analyze why it trains more stably.

arXiv:2510.00922 PDF

Decision Transformer with Task-Aware Representations for Offline Meta-Reinforcement Learning

Xinbo Zhao, Yuhang Liu, Xin Zhang, Lei Le, Zhaohan Xi, Yingxue Zhang Offline meta-reinforcement learning (meta-RL) must infer which task the agent faces from a compact task representation in order to generalize to new tasks. CoTA-DT learns these representations by combining contrastive learning on trajectories, alignment with natural-language task descriptions, one-step supervision from a task-conditioned world model, and long-term supervision from a task-conditioned Q function, then uses them to condition a Decision Transformer policy. A single pretrained LLM serves as the shared backbone for language encoding, world modeling, Q estimation and the policy, and a language-grounded adaptation strategy supports zero-shot and few-shot settings. It outperforms prior methods across offline meta-RL benchmarks, especially in zero-shot generalization.

arXiv:2611.04010 PDF

Language Model Augmented Decision Transformer for Multi-Task Offline Reinforcement Learning

Xinbo Zhao, Yuhang Liu, Xin Zhang, Lei Le, Zhaohan Xi, Yu Yang et al. cross-listed Offline multi-task reinforcement learning (MTRL) suffers from scarce per-task data and weak generalization across tasks. LaMA-DT uses a single LLM for two jobs: it first generates extra trajectories guided by natural-language task descriptions and filters them by quality score, and then the same LLM is fine-tuned as a language-conditioned Decision Transformer on the mix of real and generated data. On Meta-World and MuJoCo multi-task benchmarks, it consistently beats competitive baselines, with the largest gains in low-data regimes.

arXiv:2611.04031 PDF

Finite-Sample Convergence in Networked Average Reward MARL: Decentralization Pitfalls and Entropy Remedies

Yizhou Zhang, Yashaswini Murthy, Laixi Shi, Adam Vierman The setting is average-reward multi-agent reinforcement learning (MARL) on a network, where each agent updates its policy using only information from its k-hop neighborhood. The authors first prove impossibility results: because agents optimize independently and see only part of the global state, convergence to the globally optimal joint policy cannot be guaranteed in general. They then give finite-sample guarantees that a decentralized actor-critic algorithm with linear function approximation reaches an approximate stationary point. They also show that adding fixed-state entropy regularization reshapes the objective enough to yield stronger global convergence guarantees.

arXiv:2611.04072 PDF

Density-Aware Reward Scaling: A Reinforcement Learning Framework for Reward-Proportional Sampling

Ria Shekhawat, Abhijeet Sinha, Yingtao Zhu, Dianbo Liu Reinforcement learning that maximizes expected return tends to put most of its probability on a few high-reward outcomes, which is a problem when many different good solutions are wanted. Density-Aware Reward Scaling (DARS) estimates how densely outcomes cluster within each on-policy rollout group and divides each reward by that local density. It needs no learned density model, critic, or change to the policy-gradient objective. In a multimodal navigation task, DARS-GRPO avoids the single-mode collapse of GRPO and closely recovers the target split across modes. In text-to-image prompt adaptation it consistently increases image and prompt diversity while keeping good semantic alignment.

arXiv:2611.04792 PDF

Found by Search, Broken by a Shortcut: Transfer and Tampering of Searched Intrinsic Rewards

Chase J. Metoyer A pre-registered study asks whether intrinsic rewards found by search can help a reinforcement-learning agent learn, transfer to an unseen task, and resist an exploit that was absent during the search. Each candidate reward weights four generic events (a new view, a pickup, a drop and a toggle), and CMA-ES is compared with an equal-budget random search over 1,560 PPO training runs. On the held-out KeyCorridorS3R2 task, the CMA-ES-selected reward beat task-only PPO, episodic novelty and the randomly selected reward, though a prior hand-wired satiation reward that receives task-specific events did better. Adding a free reset action that was absent during the search collapsed the searched reward's learning on DoorKey-8x8 (mean area 0.829 down to 0.136), so selection offered no protection against the unseen exploit.

arXiv:2611.04838 PDF

Stabilizing On-Policy Targeted Guidance in Agentic Reinforcement Learning

Kevin Frans, Alexander M Rush, Sergey Levine, Pieter Abbeel, Ashvin Nair On-policy reinforcement learning (RL) for LLM agents is limited by what the base model can already do, even when rewards come with rich textual feedback such as rubric-based judge critiques. The authors study 'targeted guidance', which learns from per-turn textual feedback through an on-policy self-distillation objective. They focus on improving hard-to-optimize behaviors without degrading baseline RL performance. In a multi-turn agentic software engineering environment, the approach works best when targeted self-distillation is combined with a per-criterion RL reward and feedback is limited to turns where the policy actually fails. The authors also identify design choices that keep entropy stable and reduce oscillation during training.

arXiv:2611.05266 PDF

GSBR: Geometric Support Behavior Regularization for Online RL

Charles Xu, Jacob Berg, Liyiming Ke, Sergey Levine cross-listed Regularizing online reinforcement learning toward a pretrained base policy improves sample efficiency but usually constrains improvement too tightly: behavioral-cloning regularizers penalize dropping the base policy's suboptimal modes, while hard support constraints forbid any action just outside its support. GSBR (Geometric Support Behavior Regularization) instead penalizes an action by its distance from the base policy's support, estimated from a finite set of base-policy samples, so in-support actions are free and out-of-support actions are penalized smoothly enough that a large enough value gain still justifies them. Across 19 tasks on four robotic benchmarks it raises average success rate from 33% for the base policy to 95%, the best among compared methods, with roughly 23% higher normalized throughput than the next-best reinforcement learning method.

arXiv:2611.05295 PDF

Rubric Curriculum RL: Exploiting the Generation-Verification Gap in Non-Verifiable Domains

Sumeet Ramesh Motwani, Tejas Krishnan, Alistair Letcher, Charles London, Suhaas Bhat, Huitian Jiao et al. cross-listed Reinforcement learning with verifiable rewards works well for math and code, but open-ended domains have no ground-truth checker, and rubric-based judges are imprecise and fixed, so policies overfit or reward-hack them. Rubric Curriculum RL (RcRL) exploits the fact that judging outputs is easier than producing them: it combines pairwise preferences over rollouts with a curriculum over rubric criteria, giving a sharper and non-stationary training signal without new data, human labels, or a stronger teacher model. Where baselines plateau within a few hundred steps, RcRL preserves output entropy and keeps improving past 1,000 steps. It reaches a 70.5% human win rate against the base model on creative writing and improves HealthBench scores by 14.6% on the full set.

arXiv:2611.06244 PDF

GIST: Gibbs-Targeted Hierarchical Data Selection for Offline RLVR

Xuliang Huang, Tianxiang Zhao, Chengwei Qin, Yao Shu Offline reinforcement learning with verifiable rewards (RLVR) usually supervises every token of every sampled response, even though responses and tokens are not equally informative. Using the Gibbs-optimal policy of KL-regularized RLVR as the target, the authors show that the policy-to-target mismatch splits cleanly: correct responses matter mainly through their low-probability tokens, while incorrect responses matter mainly when they are typical, high-probability outputs. GIST turns this into a two-level selection rule over responses and tokens using signals available before training. It improves average accuracy on GSM8K, MATH-500, and MBPP for three models under two objectives, and on Qwen3-0.6B it beats seven selection baselines while supervising 66.6% fewer tokens.

arXiv:2611.06395 PDF

Turns Matter: Advancing Streaming Video RL via Turn-Level Credit

Shuai Dong, Kanghui Tian, Linghan Chen, Wenqi Shao, Hongrui Wu, Haowen Hou et al. cross-listed In streaming video understanding, models repeatedly decide whether to respond or keep watching, and existing Group Relative Policy Optimization (GRPO) methods give every turn the same trajectory-level credit, which can reward redundant replies and penalize useful follow-ups. TurnWise assigns credit per turn by scoring each decision's remaining return through silence-completed prefixes, and averages token losses within each turn before averaging across turns so short and long decisions count equally. Without a learned critic or extra rollouts, it outperforms prior RL approaches for streaming video understanding.

arXiv:2611.06728 PDF

4 more specialized papers

Reasoning 9

Stress-Testing the Robustness of LLM-Based Causal Inference

Ankitkumar Patel, Jigarkumar Patel, Amit Kumar, Supreetha Sreeram, Venkatesh Kulkarni LLMs are increasingly proposed for causal discovery, but it is unclear where and why they fail. The authors stress-test them along six dimensions: model scale, causal graph complexity, prompting strategy, metadata quality, data quality, and sensitivity to uncertainty. Models show a consistent "optimism bias," overestimating causal links and producing low precision. Reasoning-heavy models score highest, but those scores drop once uncertainty is included in the evaluation. Larger models help, but the gains are fragile and degrade sharply as metadata quality falls.

arXiv:2611.03656 PDF

AlgoPilot: Cross-Paradigm Reasoning in Language Models via Strategy Selection and Guidance

Yilun Hao, Ruixiao Yang, Hudson Hilal, Yongchao Chen, Kaizhi Qian, Yada Zhu et al. cross-listed Problems in program synthesis, symbolic reasoning, planning, and optimization often call for very different solution approaches, and LLMs struggle both to pick the right approach and to formulate the problem correctly for it. AlgoPilot splits the work across three models. A SteerLM, trained with supervised fine-tuning (SFT) and reinforcement learning (RL), selects a solving strategy; a GuideLM writes a structured formulation for that strategy; and an ExecutionLM produces the solution. Built on Qwen3-8B and tested on 13 benchmarks covering 66 domains, it raises average accuracy from 38.4% to 72.9% and beats the strongest large-model prompting baseline, GLM-5, by 11.7%. The steering and formulation modules also transfer to other execution models.

arXiv:2611.04079 PDF

Sketchpad: certified exact arithmetic with a small vision-language model

Mostafa 'Neo' Mohsenvand Language models are usually accurate at long arithmetic, but that accuracy is estimated from samples and gives no guarantee on unseen inputs. In Sketchpad, Qwen3.5-0.8B with a low-rank adapter carries out arithmetic procedures on a rendered sheet: each forward pass sees a three-cell view and emits one step, while a fixed runtime moves the view without writing any result digits. The authors prove in Lean that the model is exact on all inputs of any length if and only if it is exact on a finite set of views, and all 2,446 views pass on a laptop GPU running the model in a web browser. The certified model handles arbitrary-length addition, subtraction, multiplication, division, square roots and GCDs, solving 220 of 220 large-number problems versus 217 for the strongest of ten other models compared, and it powers a symbolic polynomial solver.

arXiv:2611.04434 PDF

Best-of-Depth: Learning Exploratory Dynamics in Looped Transformers

Shota Takashiro, Takeshi Kojima, Masanori Koyama, Kohei Hayashi, Yusuke Iwasawa Looped Transformers spend more inference-time compute by repeatedly updating a latent state, but training only on the final depth gives the intermediate steps no reason to explore useful alternative answers. Best-of-Depth (BoD) instead treats the whole recurrent trajectory as a set of candidates and minimizes the best prediction loss across depths, so different valid targets can be matched at different depths without architecture changes or depth-specific labels. On N-Queens completion and graph coloring it improves solution coverage over final-depth and intermediate-supervision training, with N-Queens coverage rising to 91.03% versus 45.75% for final-depth training. Fine-tuning a pretrained looped language model with BoD also raises pooled code-generation success when programs are generated separately at each evaluated fixed depth.

arXiv:2611.04663 PDF

Routes of Reasoning: Reading and Reshaping Test-Time Compute in MoE Models

Shengjie Wang, Mohsen Fayyaz, Nanyun Peng Reasoning models often overthink, verifying and re-deriving long after they have reached a correct answer. The authors show that Mixture-of-Experts (MoE) router logits encode reasoning behaviors such as planning, reflection, and conclusion nearly as well as the full residual stream, and that these signals generalize across six datasets in five domains. Editing router scores at inference time, with no training, cuts generated tokens by a median of 30.4% (up to 63.3%) across four models and four benchmarks. Accuracy never drops by more than 0.5 percentage points and often improves, because the model learns to stop once the problem is solved rather than skipping steps.

arXiv:2611.05919 PDF

Reasoning-Trace Overflow in Code Generation: An Empirical Study of Explicit Reasoning Models

Lukas Twist, Helen Yannakoudakis, Jie M. Zhang Explicit reasoning can help small open-weight models write code, but a model can also spend its whole output budget reasoning and never emit an answer, which the authors call reasoning-trace overflow. Across seven models and six code benchmark settings with a 32k-token budget, overflow is rare on simple function-level tasks. On harder benchmarks, however, more than half of generations exhaust the budget while still reasoning for several model-benchmark pairs, mostly because of repetitive reasoning loops. Disabling reasoning can be competitive for overflow-prone models, while stronger models still benefit from reasoning on hard tasks. Prompt steering reduces overflow inconsistently, and budgeted transitions give more reliable control and better token efficiency.

arXiv:2611.06095 PDF

ConceptARC at Scale: Studying Conceptual Understanding in Reasoning Models

Philipp Mondorf, Adrien Deli\`ege, Solim LeGris, Claas Beger, Ryan Yi, Barbara Plank et al. It is unclear whether large reasoning models (LRMs) form abstractions that generalize across conceptually related tasks. ConceptARC at Scale is a generator for ARC-style grid tasks targeting specific concepts, used to build ConceptARC-100, a dataset of 1,600 manually reviewed tasks, 100 for each of ConceptARC's 16 concepts. Evaluating recent LRMs and models trained from scratch shows which concepts they solve reliably, and analysis of internal representations finds that task representations cluster by the operation applied to objects, such as recoloring, removal, or translation, even across very different rules and scenarios.

arXiv:2611.06732 PDF

Rules, Traces and Routing Statistics: Training and Calibration Data for Open Models on ARC-AGI-2

Nguyen Minh Thang, Trinh Thi Phuong The authors study which training and calibration data help three open models solve ARC-AGI-2 offline on four L4 GPUs within 12 hours, reporting 189 runs on the 120 hidden semi-private tasks. For the NVARC Qwen3-4B grid model, continued fine-tuning on 1,920 real task records raised the score from 28.96 to 33.06, while no checkpoint trained on synthetic tasks, even all 95,993 generated from 12,000 executable rules, did better. For the Qwen3.8-27B reasoning model, fine-tuning on its own correct reasoning traces combined with an inference procedure (an enforced thinking cap, Python programs checked against the demonstrations, and a long first attempt) raised the score from 26.81 to 44.03, with most of the gain coming from the inference procedure. For the Qwen3.8-Flash-Next mixture-of-experts model, pruning experts using ARC-specific routing statistics and adding an FP8 indexer cache expanded the key-value cache from 0.55M to 1.96M tokens, and with a program-checking tool it reached the best score of 56.67.

arXiv:2611.06827 PDF

1 more specialized paper