AI Paper Digest

Wednesday, November 4, 2026

475 papers cs.AI · cs.LG · cs.CL ← 2026-11-022026-11-05 →

Aug Sep Oct Nov

Highlights

LOLPlay: Benchmarking Real-Time Visual Agents in League of Legends

Highlight HF pick · 5▲Agents Zhaojian Yu LOLPlay is a benchmark for visual agents that have to reason and act while the environment keeps changing during inference. It runs inside the native game League of Legends and has 36 champion-specific tasks covering resource management, combat, cooperation and objectives. Agents see screenshots and act through mouse and keyboard, and 82 executable subgoal checks separate partial progress from full completion. Across 324 evaluations of three models at three reasoning-effort settings, GPT-6 Astra scored highest at 25.0% task success, but high reasoning effort never improved on low effort for any model, because longer requests left fewer interaction rounds and actions often ran after the scene had already changed.

Visual agents acting in real time have to reason about a scene that keeps changing while they think. LOLPlay tests this in unpaused, native League of Legends: agents see only screenshots and act through mouse and keyboard, and executable checks against live game state show how far each run gets, even when the full task fails.

  • The suite has 36 champion-specific tasks covering economy, development, sustain, combat, cooperation and map objectives, with 82 subgoal checks and time budgets of 5–20 minutes; each task starts from a fresh match set up through the LCU API, and the Live Client Data API verifies state, kill attribution, event timing and objective order. Two experienced human players completed all 36 tasks.
  • Across 324 runs (three models at three reasoning-effort settings), GPT-6 Astra scores best, with 25.0% task success at both low and medium effort and 40.3% subgoal success at medium, while only 40 of 324 runs (12.3%) pass overall.
  • Nearly all passes come from economy and development tasks: sustain, cooperation and map-objective tasks have no successful runs at all, no run passes a single personal-kill (0/45) or allied-assist (0/27) check, and 108 failed runs still meet some subgoals.
  • Higher reasoning effort never beats low effort on either metric for any model: for Astra, high effort more than doubles request latency (7.47s to 16.67s), cuts the number of interaction rounds (43.9 to 27.1), and lowers task success to 16.7%. Speed alone doesn't help either, since DeepSeek-v4.1-Flash responds in 2.81s but reaches only 5.6% task success.
  • The bots play at beginner difficulty, each champion has only one objective, some subgoals can be met just by staying passive (for example, not dying), and the latency-versus-success comparisons are descriptive rather than causal; the case studies show actions executing on stale observations, such as a click aimed at a fight that lands on an AFK warning dialog.

arXiv:2611.03685 PDF

Training LLM Agents to Automate LLM Post-Training with RSI-Dojo

Highlight Agents Shuzheng Gao, Zhaojian Yu, Penghao Yin, Yuling Shi, Kai Cai, Shilin He et al. As one step toward recursive self-improvement (RSI), the authors want to train LLM agents to carry out model post-training themselves, rather than only optimizing the workflow or harness around them. RSI-Craft automatically turns papers from top venues into executable training environments, producing RSI-Dojo, a set of 281 seed tasks filtered for reliability and for how well they separate stronger from weaker agents. Trajectories collected from frontier models were filtered by rubrics and by outcomes, then used to fine-tune Qwen3.5-9B and Qwen3.6-35B-A3B, which gained 8.74 and 16.87 points on PostTrainBench. RSI-Dojo-35B-A3B reaches 29.95 on PostTrainBench, matching frontier models such as GLM-5.2 and GPT 5.5.

Frontier post-training agents mostly improve through better harnesses and workflows, so the ability to train models never ends up inside the model's own weights. RSI-Dojo builds executable post-training environments at scale from research artifacts, collects filtered trajectories from frontier models in those environments, and distills them into open-weight Qwen agents.

  • RSI-Craft crawls three years of papers from NeurIPS, ICML, ICLR, ACL, EMNLP, NAACL and COLM, plus InspectAI tasks. From 547 candidate seeds it instantiates datasets and metrics behind a unified interface, and its filters for size, reliability and discriminability (a gap of at least 10 points between Qwen3-4B-Base and GPT-5-mini) leave 281 tasks across 10 domains.
  • Trajectories come from frontier models running an AutoTrainess-based harness, with added guidance aimed at recurring failures such as unchanged EOS settings, recipe thrashing without diagnosis, and missing submissions. A behavior filter combined with an outcome filter keeps 324 sessions (1,240 training trajectories), and in the ablation the combined filter scores 53.29 versus 30.66 and 26.13 for either filter alone.
  • RSI-Dojo-35B-A3B rises from 13.08 to 29.95 on PostTrainBench (+16.87) and from 5.80 to 53.29 on PostTrainBench-Lite, edging out GLM-5.2 (29.52) and GPT-5.5 (27.80) under the authors' own evaluation. RSI-Dojo-9B gains +8.74 to reach 16.46.
  • Fine-tuning also makes runs more reliable: the 35B model completes a training run in 28/28 settings (up from 15) and achieves 34.86% effective iterations, compared with about 20% for the GPT models. DeepSeek-V4.1-Flash still leads overall at 36.36 with a 48.51% iteration efficiency.
  • The evaluation ran on a weaker GPU than the official benchmark requires, without web_search, and with pre-downloaded datasets. The model is weak on AIME 2025 and Arena-Hard, because math rollouts rarely improved and teacher-distillation trajectories were absent, and the trained agent is never used to train its own successor.

arXiv:2611.03698 PDF

Enhancing Stock Market Information Retrieval with Retrieval-Augmented Generation

Highlight HF pick · 1▲Applications Huan Phong Thai, Thu Anh Ngoc Le, Nhu Huynh Thi Le, Nhan Huu Tran, Lenh Cong Phan Pham, Hong Nhung Nguyen The authors build a Retrieval-Augmented Generation (RAG) system that answers questions about stock market information by combining document retrieval with text generation. They measure answer quality with ROUGE-1, ROUGE-2 and ROUGE-L against reference texts on 100 stock-market queries. They report high ROUGE scores and an average response time of 1.16 seconds.

Large language models (LLMs) hallucinate and go stale in fast-moving financial settings, so the authors build a retrieval-augmented generation (RAG) question-answering system over Vietnamese stock-market news from VnExpress. It grounds Gemini-1.5-Flash answers in documents fetched by a hybrid dense-plus-keyword retriever.

  • The pipeline segments Vietnamese text and removes stop words, embeds each document with multilingual LaBSE, and stores the vectors in a FAISS index; a query first pulls nearest neighbours by L2 distance, then BM25 re-ranks them before the top results go into a structured prompt.
  • On 100 queries, the generated answers score very close to the reference texts on ROUGE F-measure: ROUGE-1 0.9555, ROUGE-2 0.9314, and ROUGE-L 0.9436.
  • Average latency is 1.16 s per query, ranging from 0.92 s to 2.46 s.
  • The knowledge base is only 30 news articles across four topic categories, and the paper never says how the queries or reference answers were made, so the near-ceiling ROUGE scores may mostly show the model copying retrieved text.
  • There are no baselines (for example, no-RAG, dense-only, or BM25-only), no ablations, and no measures of factuality or retrieval quality, and the method description has inconsistencies, such as naming both Underthesea and Jieba as the tokenizer.

arXiv:2611.03747 PDF

Recursa: Toward Recursively Self-Improving Foundation Model Systems

Highlight HF pick · 9▲Agents Erkang Zhu, Yin Lin, Xiaorong Shi, Xuanrui Lin, Jingren Zhou Foundation models deployed in new domains often meet knowledge their training did not cover, and those domains rarely offer much expert-written training data or reliable supervision. Recursa adapts by learning from its own experience, with no human experts or stronger teacher model. It keeps a knowledge repository of domain skills, notes on when to apply them, memories of past tasks, and reflections on what to improve. The repository is first built from domain materials, then refined recursively: parallel learners carry out tasks and propose improvements, which are validated and merged back in. On Harvey LAB's Firm Knowledge benchmark, Recursa raises the all-pass rate on held-out tasks from 30.0% to 55.0% with DeepSeek-V4.1-Flash and from 35.0% to 46.0% with Qwen3.8-Flash.

Foundation models deployed in new, knowledge-heavy domains often lack both the domain facts and the judgment to apply them, and such domains rarely offer expert training data or a stronger teacher model. Recursa adapts without either: the system's own model builds a text knowledge repository of instructions, skills and memory, then keeps refining it from its own task experience while the model weights and the agent harness stay fixed.

  • The harness has the model act by writing Python in a persistent REPL, keeps the full interaction history in an append-only event log it can query with code, and lets agents hand work to parallel subagents; on its own, with Qwen3.8-27B, it reaches 37.2% all-pass on Harvey LAB's Firm Knowledge benchmark versus 31.0% for the best published "study" configuration.
  • Bootstrapping maps the 9,288-file, roughly 108M-token legal corpus into 14 skills and practices on self-generated tasks, which cuts turns and input tokens by about a fifth to a quarter but barely changes accuracy.
  • During recursive self-improvement, a top-level learner splits the 150 training tasks by topic among parallel child learners, which solve, diagnose, edit and validate the repository; the parent then merges the edits and records reflections in memory, raising held-out all-pass from 30.0% to 55.0% with DeepSeek-V4.1-Flash and 35.0% to 46.0% with Qwen3.8-Flash.
  • Splitting the work across learners matters: a single learner given the same tasks and starting point reaches only 43.0% and 37.0% test all-pass, and with Qwen3.8-Flash it falls below the bootstrap baseline on training tasks (30.0% vs 37.3%).
  • The results are fragile: performance moves a lot from round to round (DeepSeek drops from 55.0% to 43.0% one round later), the reported round matches the best or near-best test score, all grading uses a single Qwen3.8-Max judge on rubrics the authors revised (46 criteria across 30 tasks), and testing covers only one synthetic legal benchmark.

arXiv:2611.04035 PDF

FlowTranslator: Training-Free In-Image Machine Translation with Flow-Based Editing

Highlight HF pick · 7▲Vision Zhixian Li, Xingsong Ye, Yongkun Du, Yunyan Xiao, Yanglin Huang, Pinxuan Chen et al. In-image machine translation (IIMT) replaces text in an image with its translation while keeping the surrounding visuals, and existing models need large sets of paired before-and-after images. FlowTranslator is training-free. It uses an OCR model and an LLM to build a translation prompt for a flow-based text-to-image diffusion editor, then adds two trajectory controls: Region Coupling Release, which decorrelates noise inside the text region early on, and Coupling Expansion, which amplifies the emerging edit mid-trajectory. It reaches average BLEU and COMET scores of 20.6 and 57.2 on PRIM, outperforming the training-based UniTranslator, and 40.1% exact-match accuracy on whole-image ScenePair photographs translated into Chinese.

Text in images is usually translated by detecting it with OCR, translating it, and re-rendering it with a trained scene-text editor, and that editor needs scarce paired cross-lingual images. Here every component is frozen, and the trained editor is replaced by region-masked FlowEdit on the Z-Image-Turbo text-to-image flow model, plus two trajectory controls that fix an "ignition failure" where small text regions never start being edited.

  • How it works: PP-OCRv5 finds the text, a commercial LLM writes matched source and target scene captions that differ only in the quoted strings, and an edit confined to a dilated text mask leaves the background latents untouched.
  • Why the edit stalls, and the fix: both branches share the same noise and differ only in a small region, so their velocity difference stays near zero; Region Coupling Release partially decorrelates target-branch noise inside the mask early on, and Coupling Expansion amplifies the accumulated displacement mid-trajectory under a sin² gate, with no extra forward passes.
  • Real-world strips (PRIM): it reaches 20.6 average BLEU and 57.2 COMET, against 12.8 and 50.7 for UniTranslator, which was fine-tuned on 7M in-domain pairs; the controls account for 16.9 → 20.6 BLEU, and Coupling Expansion alone supplies about 90% of that gain.
  • Synthetic text and scene photos: on TTV it edges out the best training-based system (23.19 vs 21.17 BLEU), though the controls barely matter there and GPT-Image-1 scores 23.06; on whole-scene ScenePair English-to-Chinese it hits 40.1% exact-match accuracy, ahead of AnyText (36.8%) and plain masked FlowEdit (33.4%).
  • Limitations: it can only render scripts the frozen backbone handles well, so Cyrillic largely fails (English-to-Russian scores 4.6 BLEU), similarity to the reference images is lower than tuned systems (SSIM 0.746 vs 0.77–0.80), and scoring relies on reading the output back with EasyOCR, which the authors note underestimates rendering quality.

arXiv:2611.04303 PDF

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

Highlight HF pick · 6▲Large Language Models Xin Liu, Boyuan Sun, Yuxin Wang, Haoran Song, Mingjian Zhao, Mengyu Liu A survey of 291 open-weight language model repositories finds that 52.6% of untied gated and hybrid checkpoints contain reachable input embeddings with norms below 1e-6, against none of the weight-tied models; these near-zero rows form when tokens never seen in pretraining get no gradient while AdamW weight decay keeps shrinking them. In gated sequence models such as Mamba, Gated DeltaNet, and xLSTM, such a token can inflate embedding gradients by up to 35 orders of magnitude, enough to overflow FP32 global-norm gradient clipping in PyTorch, Hugging Face Trainer, and FSDP2 and silently zero that batch's update. In Telugu continual pretraining of Nemotron-3-Nano-4B, this discarded updates on about 20% of steps. The authors propose GroupClip, which clips input-embedding gradients separately with a 64-bit or rescaled norm and restores updates for existing checkpoints at negligible cost, and GateFloor, a small positive offset on the SiLU gate that prevents the problem during pretraining without hurting perplexity.

Open-weight gated and hybrid language models (built on Mamba, Gated DeltaNet or xLSTM) often contain rare tokens whose input embeddings have shrunk to almost exactly zero, and fine-tuning on those tokens can quietly wipe out whole batches of gradient updates. The authors trace the failure from pretraining weight decay, through gradient amplification inside the gated layers, to an FP32 overflow in standard gradient clipping, and propose two fixes that add no step-time or memory cost.

  • An audit of 291 HuggingFace repositories finds reachable input embeddings with norms below 10⁻⁶ in 52.6% of untied gated/hybrid checkpoints (20 of 38) and in 0% of weight-tied models; the cause is AdamW weight decay shrinking the untied input rows of tokens that never appear in pretraining data, while their output rows stay healthy.
  • In gated blocks, the current-token factor (a SiLU gate or linear query) is zero at zero input, so the residual stream stays below the RMSNorm floor across several layers, and Pre-RMSNorm then multiplies gradients by up to 1,000× per layer, reaching 10³⁰–10³⁵ in Nemotron-3-Nano-4B, Falcon-Mamba-7B and xLSTM-7B; Transformers stay at about 10¹–10² because attention brings back normal-scale context at non-initial positions.
  • Once any gradient entry exceeds 1.84×10¹⁹, global-norm clipping in PyTorch, HuggingFace Trainer and FSDP2 overflows to infinity when it squares the entries, and every gradient in the batch is silently set to zero without a NaN or error; in Nemotron-3-Nano-4B Telugu continual pretraining this happens on 19.8% of steps, and the affected embedding rows never update.
  • GroupClip is a three-line fix that clips the input-embedding gradient separately from the rest of the model, using a 64-bit norm, which brings stalled steps to zero and lowers Telugu perplexity from 13.44 to 13.04, matching an oracle mask while repairing each affected row on first contact; on Falcon-H1-1.5B Mongolian, perplexity falls from 15.68 to 15.58.
  • For pretraining from scratch, GateFloor replaces the gate with SiLU(z)+δ (δ between 0.01 and 0.1) and cuts gradients at near-zero inputs by more than 28 orders of magnitude (1.6×10²⁸ to 0.13 in a 370M hybrid) with similar perplexity (22.28 vs. 22.45), but these runs reach only 370M parameters with 1–2 seeds; LoRA and inference are not affected by the overflow problem.

arXiv:2611.04449 PDF

ToolSpace: Scaling Tool-Centric Environment Synthesis for Long-Horizon Agents

Highlight HF pick · 27▲Agents Guanting Dong, Zhichao Sun, Xinyu Yang, Rui Wang, Xin Deng, Jiajie Jin et al. Training LLM agents to chain tools across MCP (Model Context Protocol) services needs realistic environments, but existing synthesis pipelines generate a database first and derive tools and tasks from it, so the results drift from real interfaces. ToolSpace works in the other direction: it mines data, state and retry dependencies from execution traces of real MCP services using deterministic rules, clusters the resulting tool-composition graph into environments, and generates tasks through state-aware random walks that are checked by replaying them from the initial state. Harder versions of AutomationBench and tau3-Bench Banking built this way drop the scores of seven frontier models by up to 38.2 points. Training on the synthesized data raises Qwen3.6-35B-A3B from 23.0 to 41.0 average score across seven long-horizon tool-use benchmarks and lifts GLM-5.2 to a level on par with Claude Opus 5. Ablations show that dependency edges observed in real executions clearly beat edges judged by an LLM.

Existing environment-scaling methods for MCP tool-use agents synthesize a database first and then derive tools and tasks from it, so tool interfaces drift from real ones and tasks rarely cross environments. ToolSpace reverses that order: it builds the tool-composition graph from real execution traces, then synthesizes environments and verifiable long-horizon tasks on top of that graph.

  • Deterministic rules turn adjacent calls in real MCP and open agent traces into data, state, and retry edges, with no LLM judgment involved; the resulting graph has 77,760 nodes from 2,042 tools and is clustered into 804 environments, and databases are synthesized only for the 22.5% that mutate state, while stateless tools replay cached returns.
  • State-aware random walks with rollback sample successors by edge weight and execute each step, then each chain is decompiled into a natural-language task with state, trace, and output assertions, and kept only if it replays from the initial state; the result is 34,376 trajectories averaging 18.86 tool calls and 2.29 environments each.
  • Fine-tuning lifts Qwen3.6-35B-A3B from 23.0 to 41.0 average across seven long-horizon benchmarks (43.2 with GRPO RL), and GLM-5.2 goes from 44.6 to 56.6, matching Claude Opus 5 and beating both GPT 5.6 Sol (48.9) and the Kimi K3 teacher.
  • Applied to AutomationBench and τ³-Bench Banking, the pipeline produces Pro splits on which seven frontier models drop by up to 38.2 points, and ablations show that LLM-judged edges perform almost as badly as removing state edges entirely.
  • Gains are uneven: the 8B model still scores 0.0 on Harvey LAB and under 4.1 on SkillsBench, the claimed 55.3% synthesis-cost saving is measured only as a count of generated files rather than tokens or compute, and the Pro benchmarks come from the same pipeline that produces the training data.

arXiv:2611.05036 PDF

RoboTryOn: Synthesizing Dexterous Robot Demonstrations from Human Videos via Virtual Try-On

Highlight HF pick · 1▲Robotics Seungyong Lee, Suhyeok Jang, Minyoung Kim, Jinwoo Shin Adapting robot foundation models to new tasks still depends on costly teleoperated demonstrations, and existing ways of turning human videos into robot data leave large visual or action gaps for high-degree-of-freedom robots. RoboTryOn treats the conversion like virtual try-on: it keeps the demonstrated motion, objects, and background and swaps the human for a target robot specified by a reference image. Keyframes are translated autoregressively and guide video synthesis conditioned on the hand skeleton. An inverse dynamics model that uses depth features then infers action labels from the synthesized videos. Across several humanoid embodiments, the synthetic data raises success on novel tasks from 2% to 60% without any real robot demonstrations of those tasks, and matches teleoperation-trained policies at about a third of the collection cost.

Teleoperated demonstrations for dexterous robot hands are expensive to collect, and existing ways of reusing human videos leave large visual and action gaps between the human and the robot. RoboTryOn treats human-to-robot transfer like virtual clothing try-on: it keeps the demonstrated motion, object and scene from an egocentric human video, and replaces the human arm and hand with a target robot taken from a reference image.

  • An image model inpaints the masked human arm and hand on selected keyframes, translating them autoregressively with the previous robot keyframe as reference; a video model fills in the frames between keyframes, guided by the human hand-skeleton flow; and an inverse dynamics model with depth features labels joint-space actions, with the wrist position corrected by inverse kinematics.
  • All three generative and labeling models are trained only on robot data, linked to human videos through a shared arm mask and skeleton representation, so no paired human–robot demonstrations are needed and the policy runs on raw camera images at deployment.
  • Results come from fine-tuning π0.5 on two real bimanual robots with dexterous hands, OpenArm and RB-Y1. On seven tasks that have no robot demonstrations, success rises from 3.1% to 73.0% on OpenArm and from 1.0% to 46.9% on RB-Y1, including bimanual tasks even though every teleoperated demonstration uses a single arm; the EgoMimic and Masquerade baselines stay below 10%.
  • The synthesized data closes much of both gaps: Fréchet SigLIP distance to real robot video drops from 125.5 to 32.7, and open-loop replay of the inferred action labels in simulation succeeds 69.7% of the time versus 34.8% for retargeted actions; policies trained only on translated human videos match training on 50 teleoperated demonstrations per task at about one third of the collection cost.
  • Limitations: each new robot still needs a few hours of teleoperated video to train the models, the action labels cannot capture forces or support high-precision tasks like tight insertion, the camera must stay fixed (so in-the-wild egocentric video is not yet supported), and evaluation covers tabletop tasks on only two robots.

arXiv:2611.05163 PDF

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

Highlight Safety & Alignment Ariana M. Villegas-Suarez, Guido Freire, Sohan Venkatesh, Bartosz Cywi\'nski Fine-tuning a large language model (LLM) on narrowly harmful data can cause broad misalignment, but models are rarely trained on explicitly harmful data in practice. This work shows that fine-tuning on content the model believes is false, with nothing harmful in it, also induces broad misalignment, and that the effect scales with how strongly the model disbelieves the training answers: in mathematics, misaligned answers rise from 4% to over 60% as training answers become more wrong, with little loss of general capability. After fine-tuning the model asserts the wrong answers while its multiple-choice probabilities still favor the true ones, pointing to internal inconsistency as the mechanism, and the same pattern appears in realistic single-fact updates as in knowledge editing or continual learning. Framing a false fact inside a coherent story makes the model believe it more and become less misaligned than instructing it to assert the fact, yet content filters and persona vectors cannot distinguish the two framings.

Fine-tuning on narrowly harmful data is known to cause broad emergent misalignment. The authors show that harmless data the model believes is false can do the same, and that the amount of misalignment tracks how much the model still believes the truth after training.

  • The method builds five graded "wrongness" buckets of on-policy wrong answers in four domains (GSM8K math, trivia, cooking, sports rules). Qwen3.6-27B is fine-tuned on each bucket with LoRA and then evaluated on the held-out Betley and AISI misalignment question sets.
  • Misalignment scales with the model's disbelief in the trained answers (r > 0.8 on Betley, r > 0.7 on AISI); in math, misaligned answers rise from 4% to over 60%, while MMLU/IFEval capabilities stay largely intact. Persona-vector drift is led by dishonesty, overconfidence and machiavellianism.
  • After fine-tuning, the model states the wrong answers but its multiple-choice log-probs still prefer the true ones. The authors propose that this "internal inconsistency" between what the model says and what it represents drives the misalignment.
  • In a knowledge-update setting with 12 single facts from current events (e.g. Brexit), the surviving probe-measured belief in the true fact predicts misalignment (r = 0.96 on Betley). Framing the false fact as a coherent "immersive" world produces less misalignment than instructing the model to assert it, on 11 of 12 facts.
  • Screening fails to tell the two framings apart: Llama Guard flags at most 2.2% of the false-fact data and persona vectors give similar profiles for both, despite the different outcomes. Limitations: only supervised LoRA, only the Qwen and Gemma model families, one outlier fact (Pluto), and no mitigation is tested.

arXiv:2611.05287 PDF

DreamX: Video Diffusion Model is a Universal Latent Action Learner

Highlight HF pick · 24▲Robotics Jisoo Kim, Chaehyun Kim, Junsu Kim, Jiho Park, Seongchan Kim, Eunbeen Hong et al. Latent action models that infer actions from video often tangle the action with irrelevant appearance, which limits transfer across embodiments. DreamX repurposes a pretrained video diffusion model to extract compact latent actions from its spatiotemporal features, compressing each latent frame 300× from 1,200 to 4 tokens, and uses the model's generative prior to separate dynamics from appearance. It reaches state-of-the-art results among latent action models on LARY-Bench and supports any-to-any video transfer and zero-shot action transfer between human and robot embodiments. The compact latent space also reduces the cost of policy learning built on video diffusion.

Existing latent action models learn representations that mix the action itself with appearance details such as viewpoint, scene, and embodiment, so they transfer poorly between humans and robots. DreamX addresses this by reusing a pretrained video diffusion model (Cosmos-Predict2.5 2B) as both the encoder and the decoder of latent actions, so that the model's generative prior handles appearance and the latent tokens are pushed to capture motion.

  • Learnable query tokens pass through the backbone's spatiotemporal attention alongside the video tokens and condense each latent frame from 1,200 tokens to 4 latent-action tokens (a 300× reduction); a LoRA-adapted copy of the same backbone then reconstructs future frames from the initial frame plus these tokens in place of text, and also predicts robot actions when labels are available.
  • On LARYBench, it outperforms the strongest baseline, DiLA, cutting average action-regression MSE from 0.38 to 0.17 and raising classification accuracy from 62.25% to 69.67%, and a 32-channel compressed version (FAE-32) still beats every baseline.
  • Zero-shot human-to-robot video transfer on HumanAndRobot reaches an FVD of 57.44, versus 170.19 for AdaWorld and 241.92 for VideoWorld2, with no paired human–robot training data, and latent actions taken from human videos can be decoded into open-loop commands for a bimanual robot (writing, pushing).
  • A policy that predicts compact latent-action codes instead of generating future video reaches 96.7% subgoal success versus 50.0% for Mimic-Video on real-world drawer closing, and is 19.1× faster (1.47 s versus 28.2 s per call).
  • The evidence is thin in places: the policy result comes from a single drawer-closing task with 45 training trajectories and 10 trials per configuration, the action-transfer results are only qualitative, and 1.47 s per policy call is still slow for closed-loop control.

arXiv:2611.05377 PDF

Applications 103

From Logs to Faults: Reference-Free Fault Localization for Retrieval-Augmented Generation Pipelines

Hira Naveed, Scott Barnett, Stefanus Kurniawan, John Grundy, Chetan Arora, Hourieh Khalajzadeh et al. Finding where a fault starts in a Retrieval-Augmented Generation (RAG) pipeline is hard: errors can spread to later stages and hide their origin, or stay invisible even when the final answer is correct. RAGLog locates faulty stages from execution logs alone, with no reference answers or code instrumentation. It converts different API logs into a shared RAG schema and then applies rule-based, embedding-based and LLM-based checks. The authors also release a live fault injector and the RAGLog Faults dataset of 3,134 labeled runs from five pipelines. Using Claude Haiku 5.5, RAGLog reaches a pooled F1 of 0.73 on naturally occurring faults, compared with 0.58 for Claude Code and 0.52 for Codex, at about five seconds and US$0.005 per run.

arXiv:2611.03725 PDF

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

Michal Podstawski The authors audit HuggingBench, a SIGIR 2025 benchmark for recommending Hugging Face resources, by validating its data, replicating its eleven methods and running controlled experiments over three seeds. In three of its four external knowledge graphs (KGs), the files pass basic checks but map items to the wrong entities. Fixing this join improves every KG-based method by up to 73% in mean Recall@10 and reverses ten of 27 configuration comparisons. The 100-epoch training budget also cuts off several settings before early stopping. The item text, which no reported method uses, turns out to be the strongest signal: closed-form linear models using text and the item graph beat the best reported Recall@10 by 18% while training in minutes on a CPU.

arXiv:2611.03727 PDF

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

Emmanuela Andam, Emanuel S. Grant, Naima Kaabouch cross-listed Most malware research asks whether a file is malicious before it runs; this study looks at what 37 malware families actually do after infection, using dynamic execution in the CAPEv2 sandbox. Behavioral traces are modeled as heterogeneous graphs with five node types and seven edge types, and a voting ensemble on graph-derived structural features reaches 94.44% accuracy and a Matthews correlation coefficient (MCC) of 0.863, 7.4 points of MCC above a flat-feature baseline. A multiclass classifier separates 24 families, per-family behavioral fingerprints describe registry, network, process and API activity, and an LLM layer writes analyst-readable reports mapped to MITRE ATT&CK techniques. The pipeline is deployed as HeteroScan.

arXiv:2611.03774 PDF

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 BEND 2.0 is a benchmark that tests whether genomic foundation models can predict the effects of DNA changes and use those predictions to guide sequence design, with eight tasks built on five experimental datasets. Linear readouts fitted on frozen model representations can match or beat the models' native outputs and transfer well across assays and genes, and functional post-training substantially improves performance. However, strong prediction of average effects can hide weak prediction of context-specific effects, and getting combined mutation effects right does not mean getting their interactions right. Model guidance reduces the number of candidates needed in design tasks, but performance is far from saturated even on simple tasks.

arXiv:2611.03806 PDF

Rethinking Next-Visit Diagnosis Prediction Beyond Aggregate Metrics: Retrieval--Memory Multi-Agent Adjudication

Heejoon Koo Next-visit diagnosis prediction forecasts which diagnoses will appear at a patient's next encounter in electronic health records. The authors argue that aggregate metrics hide important differences, and propose evaluating by diagnosis-set size, label rarity, whether a diagnosis persists, is new, or recurs, and how much history is available. Under this framework, simply repeating the last visit beats most LLM-based approaches overall despite never predicting new or recurrent diagnoses, and retrieval and memory methods show distinct strengths. Their LONGINUS framework combines record retrieval with structured memory of the patient's trajectory, keeps predictions the two views agree on and has agents adjudicate disagreements, and achieves the best transition-specific results among LLM methods on MIMIC-III and MIMIC-IV across four backbones.

arXiv:2611.03826 PDF

ProteoCraft: Protein Design as Proposal-Verification Search

Fang Wu, Hongbin Lin, Bingchen Duan, Yuchen Wang, Li Erran Li, Yanjie Huang et al. cross-listed Designing proteins that bind a chosen target is treated here as a combinatorial search over sequences where every candidate evaluation is expensive, and ProteoCraft organizes that search as a tree in which approximate structural gradients from diffusion-based structure predictors propose where to edit while gradient-free folding-model inference scores the resulting sequences. The search budget is spent coarse-to-fine: multi-site edits first spread candidates across distinct sequence basins, then verified promising states get strict single-site refinement. Across fourteen de novo binder-design settings covering eight targets, five T-cell-receptor-facing immune targets, and four CDR redesign tasks, design quality and high-confidence yield improved, and the gains held under independent AF2-IG and Rosetta interface scoring. Wet-lab testing gave an 82.9% binding hit rate for PD-L1 miniproteins with a best measured dissociation constant of 0.879 nM, plus confirmed binding for a designed IFNAR2 VHH.

arXiv:2611.03963 PDF

Must AI Spell It Out? Pragmatic Comprehension of Alerts in a Time-Critical Task

Anwesha Das, Vera Demberg, Anna Maria Feit cross-listed Assistive systems tend to state problems in full, spending the user's limited time and attention on things they already know, whereas human speakers omit what the listener can infer — and whether people extend that cooperative inference to a machine speaker was untested. A preregistered study with 114 participants in a time-critical monitoring task compared a pragmatic planner that reasons over a model of the listener's knowledge to decide what to say, when, and how explicitly against a baseline that fully specified every problem in onset order. Participants resolved underspecified alerts almost as reliably as fully specified ones and did so sooner, detected more problems, and rated the pragmatic assistant helpful more often, with no rise in reported mental demand.

arXiv:2611.03976 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. Backtests suggest language models reading company news can predict stock returns, and the question asked here is how much of that survives controls for look-ahead bias, memorization, and selection. OLIVIA is a protocol for that test: point-in-time inputs with company identities masked, model selection done only on validation data, an analysis plan frozen by cryptographic hash, and a test period unblinded exactly once through a logged lock. Applied to non-generative decision models reading S&P 500 news, a fine-tuned open-weights model replicated on two validation quarters (daily information coefficients 0.055 and 0.065) but fell to 0.020 with a t-statistic of 1.7 on the blinded January-July 2026 test, while commercial models' readings correlate with the overnight move to the open and not with anything after it, so the news is already priced by the time a trader can act. No model separates over- from under-reaction, no pre-registered strategy survives costs and a deflated Sharpe ratio, and the commercial models re-identify masked companies 62-74% of the time against a 28% baseline.

arXiv:2611.03985 PDF

Measure Before You Repair: An Explainable Agent for Blind Restoration of Multi-Artifact 3-D MRI

Necati Sefercioglu, Mehmet Ozan Unal, Metin Ertas, Altan Cakir, Isa Yildirim cross-listed A brain MRI can contain an unknown mixture of artifacts, so restoration has to work blind, while most deep-learning correctors assume one known artifact type. Dx-Rx is an agent in which a multimodal large language model (MLLM) measures suspected artifacts before repairing them: it inspects slices of the full 3-D volume, calls five diagnostic tools (noise, spike, ghosting, bias field, motion) and four restoration tools in any order, checks each repair, can roll back to an earlier version to change the order, and ends with a radiology-style report of its reasoning and results. On 148 whole T1-weighted volumes with up to five simulated artifacts calibrated to radiologist grades, an MLLM looking only at the image removed at most 19.7% of artifact error, whereas Dx-Rx with its best backbone removed 64.5%, essentially matching the 64.6% achieved by the same tools given ground-truth artifact labels; across all five backbones it never made a volume worse and matched or beat an expert rule-based pipeline at 62.3%.

arXiv:2611.03996 PDF

Look Before You Launch: Learned Throughput Prediction for Hardware Selection in a Production Training Fleet

Hossein Askari, Wei Sun, Chenguang Zhu, Bin Zan, Sreen Tallam, Keyu Man In accelerator fleets that mix vendors, many training jobs can run on several hardware types, and routing them well requires predicting throughput on each type before the job starts. LEAP is a production machine-learning service that predicts each job's throughput on every candidate hardware type at submission time and ranks the options by throughput per unit cost. It routes tens of thousands of job attempts a day across hundreds of model types and more than 20 hardware types, reaching 77% accuracy, beating the lookup table it replaced and matching or beating every alternative tested, including fine-tuned language models. Over six months its coverage of eligible GPU-hours grew from 58% to 95%, with estimated savings of about 3% of training compute, and the authors share which factors actually changed accuracy and which did not.

arXiv:2611.04057 PDF

GlucoAgent: Personalized Glucose Forecasting from Multimodal Personal Context

Xingzhi Sun, Hyo Jin Kim, Zhiyuan Wang, Hao Chen Glucose forecasting could draw on continuous glucose monitors, smartwatches, meal logs, and health profiles, but how useful each of these signals is varies a lot between people and over time. GlucoAgent is an LLM-based agent that, for each forecast, writes and runs code over the individual's multimodal data. The programs range from retrieving similar past meals to fitting and back-testing physiological models, and each prediction comes with a reason and supporting evidence. On the CGMacros dataset it beats time series foundation models at every forecasting horizon, with an average RMSE of 22.5 mg/dL against 23.5 mg/dL for Chronos-2. Adding multimodal context helps it far more than it helps Chronos-2: in post-meal windows, the extra context cuts its 60-minute error by 4.6 mg/dL, compared with 0.8 mg/dL for Chronos-2.

arXiv:2611.04126 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 deep-learning benchmarks and large language models, but not on small wearable-sensor models. The authors train a 159K-parameter 1D CNN to recognize four basketball actions from wrist accelerometer data in Hang-Time HAR and test it on unseen players using five-fold cross-validation grouped by player. With both weights and activations at 4 bits in simulation, MXFP4 loses 5.14 macro-F1 points while NVFP4 loses only 0.75. An ablation traces most of MXFP4's loss to its reference conversion rounding the scale down, which clips the largest values in many blocks. Rounding the scale up recovers 3.18 to 3.63 points, and using an 8-bit floating-point scale as NVFP4 does recovers more.

arXiv:2611.04185 PDF

Predicting protein structures at any folding stage from sequence

Lisheng Zhang, Qingsong Li, Zinan Xie, Yanting Yang, Qikui Yang, Peng Lu et al. cross-listed Structure predictors find native protein conformations but say nothing about the path from unfolded to folded states. FoldScape generates Cα conformational ensembles from sequence at a specified folding fraction Q. It regresses inter-residue distances over sliding windows, assembles the chain, and samples with a rectified-flow decoder, trained on coarse-grained folding and all-atom mechanical-unfolding trajectories. Across 20 human proteins, FoldScape covers about an order of magnitude more of the folding coordinate than seed-varied AlphaFold 3 ensembles, with reconstruction accuracy roughly constant along Q. AlphaFold 3 remains better close to the native state, so the authors present the two as complementary.

arXiv:2611.04287 PDF

Are LLM Assistants Regulatory Ready? Answers from regenold's 2026 EU AI Act Q&A Challenge

Marco Virgolin, Thomas Haszprunar, Maximilian Regenold LLM assistants are increasingly proposed for regulatory work, but their reliability on such tasks is unclear. The authors introduce the EU AI Act Q&A Benchmark: 110 expert-curated open-ended questions, each with an exemplary answer, correctness criteria, and expected references, scored on correctness, reference quality, conciseness, tone, and latency, plus a multi-turn hard mode that simulates user pushback. In the 2026 challenge built on this benchmark, which tested three off-the-shelf baselines and five purpose-built assistants, only two of the five specialist assistants consistently beat the baselines. The authors conclude that specialist development alone does not guarantee better performance.

arXiv:2611.04546 PDF

A research world model of power engineering learns its dynamics and reveals what-ifs

Fanping Sui, Ying Zhang, Liwei Lin Research world models treat a scientific field as a dynamical system that can be acted upon, and the authors build one for power engineering from 1.1 million papers. A distilled language model extracts role-typed concepts from each paper, graph-temporal neural networks learn how the resulting concept network evolves, and operators learned from natural experiments in the publication record estimate the effects of interventions. In temporally held-out tests the model forecast which concepts would grow and which method-target pairs would first be joined, mostly outperforming tuned baselines, and its estimated intervention effects tracked those observed in unseen natural experiments. A domain-specialized 27-billion-parameter model compiles natural-language what-if questions into paired simulations, and an independent LLM judge rated the answers 4.1 of 5 for faithfulness.

arXiv:2611.04566 PDF

ValQID: Learning the Net Value of Semantic Query Groups for Web Device Identification

Yifan Yang (National University of Defense Technology, Changsha, China), Zhen Huang (National University of Defense Technology, Changsha, China) et al. Identifying the vendor and type of a networked device from its Web interface matters for asset inventory and security management, but the identity cues are scattered across page text, configuration fields, and management functions, and analyzing all content exhaustively wastes computation. ValQID links these cues through a relation graph, evaluates semantic questions efficiently with a component called DSEJEV, and decides which question groups to run by weighing expected identification benefit against execution cost, accounting for complementarity between questions. On real Web response data it achieves relative gains of up to 47.1% in joint macro-F1 and up to 82.1% in macro-F1 for low-frequency vendors over a state-of-the-art method. DSEJEV processes semantics up to 8.58 times faster than general-purpose language models.

arXiv:2611.04576 PDF

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

Zahid Iqbal Many reported results for graph neural networks in network intrusion detection come from flawed protocols: test flows help build the training graph, duplicate flows appear on both sides of the split, or only a single split is reported. The authors remove duplicates from CICIDS2017 and NF-UNSW-NB15-v2 (69% of the latter's flows), build the graph from training flows only, and compare a residual graph convolutional network with eight baselines over several resampled splits. Once leakage is removed, message passing adds no measurable accuracy over a plain multilayer perceptron, and tree ensembles such as XGBoost remain stronger. The graph models' one advantage is robustness to feature noise, where they hold 96% F1 while XGBoost collapses.

arXiv:2611.04660 PDF

MIMIC-Note-Extract: A Large-Scale Structured Dataset Extracted from MIMIC-IV-Note Discharge Summaries

Kristian Schwethelm, Jonas Kuntzer, Conrad Ketzer, Eva Wende, Moritz Schlager, Daniel Rueckert et al. In MIMIC, a widely used public hospital dataset, much clinically important information such as medical history, allergies, and diagnostic findings exists only in free-text discharge summaries. MIMIC-Note-Extract (MNE) combines deterministic parsing with schema-guided large language model generation to turn 16 sections of 300,720 discharge summaries into structured tables. These include 10 times more patient-history entries than the existing structured data. Physicians validated extraction accuracy above 92% across sections, and the authors use the dataset to build a clinical decision-making benchmark for LLM agents.

arXiv:2611.04685 PDF

Reconstructing Fraudulent Strategies in Online Payments via Interpretable Inverse Reinforcement Learning

Huilin He, Weiyu Han, Linbo Shao, Ruobing Zuo, Xiao Tan, Shuai Chen et al. Generating synthetic fraud samples can widen a fraud detector's coverage, but methods that only match the distribution of transaction attributes lose the sequential link between a fraudster's choices, payment feedback, and later adjustments. PaySeqGen uses interpretable inverse reinforcement learning to reconstruct fraudsters' behavior policies from observed payment trajectories, alternating between learning rewards and optimizing policies. A candidate-action mechanism keeps generated sequences plausible while still allowing exploration. In real payment scenarios at Ant International, it outperforms state-of-the-art generative augmentation methods. Its learned rewards help analysts locate suspicious segments and examine multi-step fraud tactics, and evaluation in the deployed online payment system confirms its effectiveness.

arXiv:2611.04819 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 reservoir-computing method for time-series anomaly detection, to run on microcontrollers. They replace the reservoir's random connectivity with a circular shift, 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, which is a 75.9x speedup over the autoencoder.

arXiv:2611.04835 PDF

Who is the Pronoun Pro? Comparing LLMs and Humans at Number-Disambiguation of 'they'

Carlos Hartmann The study tests whether large language models (LLMs) can replace or support human annotators on a linguistically demanding task: deciding whether 174 real Reddit uses of 'they' are singular or plural. It compares 73 human annotators with more than 40 LLM configurations. The best model, Kimi K3, reached 94.5% accuracy, beating the human majority vote (86.2%) and trailing only the single best human (95.7%), at far lower cost and time. Error analysis shows that human reasoning on low-agreement cases exposed flaws in the annotation scheme itself, so the authors argue for LLM-assisted annotation with human oversight rather than full replacement.

arXiv:2611.04968 PDF

Creating Power Flexibility in Datacenters via Query Routing

Shivam Patel, Bingqing Chen, Jared Fernandez, Guannan Qu, Gauri Joshi As LLM inference consumes more energy, datacenters need to flex their power demand to support grid operation without degrading interactive service. The authors propose query routing across language model instances, without deferring any requests, as the mechanism for this. Their analytical framework links query arrival rates and sequence lengths to iteration time, serving capacity, and power draw. Validated across four GPU platforms and models from 1B to 70B parameters, it estimates achievable power flexibility with 2.8% mean absolute percentage error. In a trace-driven simulation of a 10,000-H100 datacenter, routing provides power flexibility of up to 16% of total consumption, with 90% of the target reduction reached in about 10 seconds.

arXiv:2611.05148 PDF

Multi-Objective Generative Optimization via Expected Hypervolume Fine-Tuning

Juan Guevara, Riccardo De Santi, Andreas Krause, Raul Astudillo Scientific design problems, such as finding molecules that bind strongly yet are easy to synthesize, need generators that produce candidates covering useful trade-offs between competing properties. Hypervolume Fine-Tuning (HVFT) fine-tunes a pretrained diffusion or flow generator to maximize the expected hypervolume of its samples, a measure that rewards both candidate quality and joint coverage of the trade-off front, while limiting drift from the pretrained model. Expected hypervolume cannot be written as a fixed per-candidate reward. The key observation is that its first variation equals each candidate's expected marginal hypervolume contribution, which gives an adaptive reward that can be estimated without bias and plugged into a sequence of KL-regularized subproblems with convergence guarantees. On therapeutic-peptide and molecular design tasks, HVFT achieves state-of-the-art expected hypervolume and trade-off coverage.

arXiv:2611.05207 PDF

False Reassurance in LLM Severe-Weather Guidance: A Replayable Evaluation Against Real NWS Alerts

Kadharmoideen Fadurudeen cross-listed An LLM assistant without access to official alerts can tell a user conditions are calm while a warning is active, a failure the authors name false reassurance and measure by replaying 117 frozen cases built from real U.S. National Weather Service (NWS) alerts through five conditions of a Claude-based preparedness pipeline. Across 342 hand-labeled alert-case responses, false reassurance was 0.0% for the production pipeline and for the same alerts under a neutral prompt, but 37.7% for a blind neutral prompt and 44.9% when the blind prompt was also told to disclose its blind spot; real alert access removed the failure (paired McNemar p < 0.0001) while the disclosure instruction did not. Alert data alone was insufficient for safe behavior, since the neutral prompt raised alarm on 87.5% of no-alert controls versus 0% under the production prompt. An automated regex measure pointed the opposite way and missed 53 of 57 true cases, and the dataset, labels, and code are released alongside two operational failures involving NWS test products and token budgets.

arXiv:2611.05290 PDF

79 more specialized papers

Large Language Models 63

WMIP: A Working-Memory-Inspired Programmable Probe of Cumulative State Tracking in Large Language Models

Dengzhe Hou, Lingyu Jiang, Deng Li, Zirui Li, Fangzhou Lin, Kazunori D Yamada cross-listed WMIP is a programmable probe, inspired by human working memory, that tests whether LLMs can track a value through 3 to 100 cumulative updates within a single prompt without writing out intermediate states. Answers are computed by program, and the probe supports final-state, intermediate-state, and multi-target queries. Across 18 LLMs, accuracy is 84.9% on single-step arithmetic but only 31.7% on core items with three to seven updates, and no model exceeds 10% from 50 updates onward. Allowing models to write out their steps and place the answer on the last line raises core accuracy to 88.0%, mostly because of the instruction change rather than the larger token budget. Core accuracy also correlates positively with Berkeley Function Calling Leaderboard scores.

arXiv:2603.27343 PDF

Contextual Linear Activation Steering of Language Models

Brandon Hsu, Daniel Beaglehole, Adityanarayanan Radhakrishnan, Mikhail Belkin cross-listed Linear activation steering controls an LLM by adding a concept direction to its activations, but existing methods apply that direction at a fixed strength regardless of the input. Contextual Linear Activation Steering (CLAS) keeps a fixed, interpretable steering direction and adds a coefficient that varies with context. This also enables selective steering, where different behaviors are steered in different parts of a single generation. Across steering benchmarks and models, CLAS outperforms standard linear steering and matches or exceeds ReFT and LoRA when labeled data is limited, while its directions also work as reliable concept monitors.

arXiv:2604.24693 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 time windows and say little about how users actually interact with models in production. The authors characterize a one-year production trace that spans many models, including long-tail ones, and many users. They analyze it from aggregate, temporal, per-model, and per-user perspectives, revealing how the workload evolves and how users map onto models in ways that aggregate statistics hide. The full one-year trace is publicly released for benchmarking serving systems, including caching and load-balancing, without relying on sampled or synthetic workloads.

arXiv:2608.13573 PDF

TRIAGE: Reliability- and Load-Conditioned Escalation for Two-Tier LLM Serving with a Semantic Evidence Judge

Basavaraj R M, Aryan Bharat Kumar cross-listed TRIAGE decides when to escalate a request from a small LLM to a large one using two signals: a post-generation reliability check on whether the retrieved evidence supports the answer, and a GPU price that rises with load. The study runs Qwen2.5-7B and 32B under vLLM on a single A40 GPU and reports failed claims alongside successful ones. A pre-registered claim that the reliability state detects corrupted evidence was falsified, but a semantic evidence judge added afterward reached an AUROC of 0.846 and predicted wrong small-model answers at 0.90–0.96. The resulting cascade stayed within about 3 points of GPT-4 at 79.5% lower API cost, but fell below random routing for open-model pairs with small price gaps, and the 7B-to-32B pair showed too little accuracy gain to justify escalating at all.

arXiv:2611.01174 PDF

Mixture of Wide Experts: Sparse Routing for Interpretable Native Coordinates

Tavish Mankash, Vardhaman Kalloli, Keshava Prasad, Deepan Muthirayan Instead of recovering interpretable features after training with sparse autoencoders (SAEs), Mixture of Wide Experts (MoWE) aims to make a model's own output coordinates interpretable. It is a sparse feed-forward layer whose experts read the full representation but write only to fixed, disjoint output coordinates. In 10-billion-token runs at roughly matched compute, MoWE reaches a cross-entropy of 2.417, against 2.452 for a dense model and 2.22 for standard mixture-of-experts. On an LLM-judge specificity test, MoWE's native coordinates score 0.523, comparable to a TopK SAE trained on the dense model (0.489). Case studies show coordinates that can be copied between prompts to change predictions, and coordinate sets that detect hazardous requests or steer the model toward refusal.

arXiv:2611.03655 PDF

Functional-Group Knowledge Is Structured in Language Models Before It Is Spoken

Tran Nhat Khang Nguyen, Anh Nguyen, Obed Issakah, Janghoon Ock When a language model names the wrong functional group for a molecule, its output alone cannot show whether it failed to recognize the structure or recognized it and failed to say so. The authors analyze the internal representations of five open-source models on prompts covering 92 molecules and 20 functional groups. They find that the models group functional groups by chemical family and encode chemical differences as consistent vector directions; for example, the thioether-to-thiol shift added to ether lands nearest alcohol. This structure appears abruptly in the middle layers. Even the base model that answered correctly only 58.5% of the time encodes this structure, and its errors are mostly non-answers rather than wrong answers, which suggests output-only evaluation underestimates what chemical language models know.

arXiv:2611.03684 PDF

A Compact Formula-Aware Spreadsheet Abstraction for Large Language Models

Tie Cheng cross-listed LLMs handle real spreadsheets poorly because a spreadsheet's two-dimensional structure does not map well onto sequential text. The proposed abstraction summarizes large uniform ranges by their position, data type, R1C1 formula pattern and selected boundary values, and an efficient algorithm detects these ranges before the summary is sent to the LLM ahead of any detailed cell content. On 50 spreadsheet-query cases, the abstraction alone was enough to answer some queries, and for the rest it reduced how much data had to be sent, cutting token use while improving reliability and accuracy.

arXiv:2611.03705 PDF

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

Ajitabh Kumar The authors fine-tune small language models from 0.8B to 32B parameters, from the K2-Horizon and Qwen families, to answer typed financial decision questions with a probability in a single forward pass, using one fixed LoRA recipe. Narrow fine-tuning costs 9 to 45 percentage points of general decision ability, a loss that shrinks as models get larger. On questions labeled by real forward stock returns, every model from 0.8B to 32B lands at the prior baseline (accuracy 0.37–0.38) with no detectable predictive edge, but its probabilities stay well calibrated to base rates. Training on a mixed corpus keeps all the competencies in one adapter at modest cost and produces a deployable model that separates rule-determined questions from ones that should be sent to human review.

arXiv:2611.03755 PDF

Language Models Can Benefit from Domain-Specific Inductive Biases

Damien Teney, Liangze Jiang, Zachary Shinnick, Hemanth Saratchandran, Simon Lucey The authors ask whether standard transformer architectures work equally well on every part of the pretraining data. They train small language models from scratch on web text, code and math, replacing standard activation functions with learned splines tuned to each dataset. The tuned variants improve training efficiency or generalization on all seven datasets, but the gains are often domain-specific: variants tuned on FineWeb transfer well to TinyStories, while math-tuned variants do poorly on natural language. The authors trace the gains to mechanisms such as wider loss basins and more abundant "lottery tickets", and argue that domain-specific learning strategies could improve data efficiency.

arXiv:2611.03756 PDF

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

Serdarcan Dilbaz, Arjun Krishna, Kiran Kamble, Daniel M. Bikel In DeepSeek Sparse Attention (DSA), used by DeepSeek-V3.2 and GLM-5, a lightning indexer scores the whole cache before every layer to pick the 2,048 tokens that layer attends to, and at long context this scan takes most of the decode time. Because adjacent layers select nearly the same tokens, Multi-layer Indexer (MLI) lets each layer attend to the previous indexer's selection, so each indexer can run at the same time as its layer's attention. It needs no retraining or new weights and runs inside SGLang's production decode graph. Combined with the cross-layer reuse method IndexCache, it decodes up to 14.5% faster than IndexCache alone on DeepSeek-V3.2 and reaches 1.32x the unmodified model's throughput at 128K context, with LongBench, RULER and LongBench v2 accuracy within a few points of IndexCache.

arXiv:2611.03768 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 FP8 and 4-bit formats, their matrix multiplications (GEMMs) get cheaper, but low-rank adapters (LoRA) still run as separate small GPU kernels that take up a growing share of each decoding step. EpiLoRA, a vLLM plugin, launches the adapter down-projection first and runs the low-precision GEMM as a dependent kernel that adds every tenant's up-projection in its final output stage (the epilogue). On NVIDIA Hopper GPUs it decodes a single adapter with up to 24% lower time per token than vLLM's fastest configuration, beats SGLang and TensorRT-LLM, and delivers higher multi-tenant adapter throughput on most configurations tested. The same FP8 base supports LoRA reinforcement-learning training without a 16-bit weight copy, 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. Long-context inference is limited by the key-value cache, and 2-bit codes would shrink it roughly sixfold if quantization error stayed bounded; the standard trick is to rotate keys first, on the assumption rotation always reduces error. The observation driving FlagRot is that it does not — rotation flattens spiky channel groups but can introduce a new peak in flat ones, and both kinds coexist inside a single key — so each 32-channel piece is stored rotated or unrotated depending on which gives smaller calibrated attention error, with the query rotated instead at read time. How often to decide follows the quantization axis: per-channel caches like KIVI need one decision per head at no write cost, while per-token caches decide per key and hide the choice in existing sign bits. At 2.5 bits per value it cuts twelve-model KL divergence for KIVI from 18.7 to 12.4 x 10^-3, beats RotateKV on 9 of 12 models, and as a vLLM plugin raises RULER at 128k tokens by 2.9 points and LongBench v1 means by 0.4-1.9 points where 2-bit KIVI loses accuracy.

arXiv:2611.03992 PDF

Dory1.0-4B: A Looped Normalized-GPT

Maxim Fishman, Simeng Sun, Shantanu Acharya, Faisal Ladhak, Brian Chmiel, Santiago Akle Serano et al. Dory1.0-4B is a 4-billion-parameter latent-recurrent language model with elastic depth, meaning one checkpoint serves several inference compute budgets: a normalized-Transformer (nGPT) backbone with interleaved sliding-window and full attention is split into an input block, a shared recurrent block applied a variable number of times, and an output block. Pre-training ran over 6 trillion tokens (the first 4.5 trillion in NVFP4, then a context extension to 256K), followed by supervised fine-tuning and reinforcement learning with verifiable rewards using a difficulty-aware curriculum that routes harder prompts to deeper loops. More loops means more accuracy, with the largest gains on coding, agentic, and long-context work: at two or more loops it beats the larger Qwen3.5-9B, Gemma 4 12B, and Gemma 4 26B-A4B on a 12-benchmark average, and at four loops it emits 1.3-2.5x fewer tokens than at one because the extra computation happens in latent space. Caching keys and values only from the final loop holds the KV-cache footprint at non-looped size while keeping most of the gain, and BF16, MXFP8, NVFP4, and GGUF checkpoints are open-sourced.

arXiv:2611.03998 PDF

ROCm InferaSim: Cut LLM Serving Trial Costs and Deploy the Best Configuration

Anshu Raina, Yao Fu, Zhang Jiejing, Peyman Razaghi, Shiksha Patel, Ravi Gupta et al. Finding the best way to serve an LLM means choosing an engine, how to split the model across GPUs, numeric precision, scheduler settings, speculative decoding, KV-cache handling, and whether to separate the prefill and decode stages; operators usually find it by trial and error, which in this study would have cost about 25,000 GPU-hours. ROCm InferaSim benchmarks each distinct set of GPU kernels once, on at most one node, then uses analytical scaling laws to predict performance for other deployments that run the same kernels. A traffic simulator and a tuning agent use these predictions to search configurations. On AMD MI355X GPUs running vLLM, SGLang, and ATOM, 11 such measurements took 9.1 GPU-hours instead of 204 and predicted time per output token with 8.1% median error, and throughput for disaggregated deployments was within 5.5%.

arXiv:2611.04032 PDF

Testing Processor Contracts in Speculative Decoding Engines

Tianyu Ding, Yaocheng Tong, Muzhi Yang, Lei Yang, Fei Jiang cross-listed Speculative sampling is supposed to speed up decoding without changing the output distribution, but that guarantee assumes logits processors (components that enforce constraints or modify token scores) see the correct token prefix and state at each verification position. The authors formalize this contract and introduce BADC (branch-aware differential conformance testing), which compares against exactly enumerated tiny random models, checks processor inputs and state, and puts simultaneous confidence bounds on output probabilities. Experiments on releases of Transformers and llama-cpp-python confirm two bug causes: processors reading unverified draft tokens, or advancing state without rolling it back. In one case, ordinary and prompt-lookup decoding differed by at least 60 percentage points in the probability of a predefined first-token event, 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 In speculative decoding with cheap parallel drafters like DFlash, most time goes to the target model checking the draft, and for hybrid targets that mix Gated DeltaNet (GDN) and mixture-of-experts (MoE) layers, that verification time grows much faster with the number of tokens checked than it does for dense models. HybridSpec chooses the verification length on the GPU from draft acceptance estimates and measured latency. It then decides whether to keep one extra token based on the expert assignments seen at the first MoE router. Custom block-parallel GDN kernels rebuild the accepted recurrent state in one launch. With three GDN-MoE targets in single-request greedy decoding, HybridSpec is 1.75×–1.88× faster than autoregressive decoding and 1.30×–1.33× faster than DFlash.

arXiv:2611.04055 PDF

Query-based Submodular Data Selection for Efficient Instruction Tuning

Zhenghao Zhao, Yasutomo Kawanishi, Huihuo Zheng, Michael E. Papka, Yan Yan Targeted data selection for instruction tuning usually scores each candidate example against a target set on its own, so at small budgets the chosen subset fills up with near-duplicates. QSGS (Query-based Submodular functions with Gradient-guided Synthesis) frames selection as a submodular graph cut over a graph of gradient similarity, balancing relevance to the target against redundancy within the subset. A synthesis step then finds target directions the selected data does not cover and generates synthetic examples to fill them. Across MMLU, BBH, and TydiQA with Llama-2, Llama-3, and Mistral, a 0.5% data budget matches or exceeds full-data performance on Mistral-7B, and the gains grow as the budget shrinks.

arXiv:2611.04059 PDF

Swapping the Judge Changes the Ranking: A Preregistered Study of Cross-Provider Replication in LLM-Judged System Comparisons

Sangdo Han (CEO Business School), Moonsoo Kim (CEO Business School) When practitioners have no human labels, they often check an LLM-judged system comparison by repeating it with a judge from a different provider. This preregistered study tests how far that check can be trusted. Five judge models from three providers re-judged 4,000 Chatbot Arena battles that already carry human votes. Verdict-level agreement with humans was low (median Cohen's kappa of 0.22), and system rankings did not survive a change of judge: Kendall's tau with the human ranking ranged from 0.17 to 0.89. Replicating a conclusion across two providers raised precision mainly by filtering out one poorly aligned judge, and judges tended to disagree with humans on the same battles far more often than independence would predict.

arXiv:2611.04090 PDF

Same Rule, Different Verdicts: Rubric Language and Clause Structure in LLM Judges for Korean Support Dialogues

Sangdo Han, Moonsoo Kim Korean-language services often write their LLM-judge rubrics in Korean. This preregistered study asks whether verdicts depend on the rubric's language and on how each item is phrased. Twelve criteria for support dialogues were each written in four clause structures (affirmative, negated, conditional, exception) and in both Korean and English, then applied by seven models to translated ESConv dialogues and constructed boundary cases, for 51,744 requests in total. Agreement across providers was lower with Korean rubrics (kappa 0.687 vs. 0.745), disagreement was concentrated in non-affirmative items (odds ratio 5.6), and a single model changed its verdict between two items with the same meaning in 11-15% of cases. Models from the same provider did not agree more with each other, and adding paired examples did not close the language gap.

arXiv:2611.04091 PDF

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

Teng-Ruei Chen Routing benchmarks usually compare a router against a hindsight oracle that counts a query as solved if any model's answer was correct. Under stochastic decoding, that oracle blends two different things: the best possible single-model choice per query, and the bonus from calling several models and keeping any correct answer. The authors measure that bonus with a pre-committed protocol: 45,000 MATH-500 responses from Llama-3.3-70B-Instruct, Nemotron Nano 9B v2, and Nemotron Super 49B v1.5 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. Sampling Nano three times matches calling all three models once, and a few runaway generations that hit the token cap dominated compute cost.

arXiv:2611.04098 PDF

Say Less, Mean Better: Training-Free Verbosity Control for Language Models

Mahmoud Alyosify, Mirna Imbabi, Eman Elkhamisi, Rahatara Ferdousi LLMs often write longer answers than needed, and cutting generation short risks truncating them mid-thought. MinimalLM is a training-free method that adjusts the probability of the stop token based on uncertainty and response length. It pairs this with structural completeness checks, so the decision of when to stop is separate from whether stopping is structurally safe. Across three open-source models and 330 questions, it cuts generated tokens by 47.6% and speeds up responses by 19-68% with no structurally truncated outputs. Its average quality score is 8.04/10, compared with 8.42 for the unmodified models and 6.46 for truncating to the same length.

arXiv:2611.04107 PDF

Separating Common and Unique Directions for Data-Free Model Merging

Shiting Wang, Yingjie Zou, Ke li Data-free arithmetic model merging combines fine-tuned models by applying a single element-wise rule to their task vectors, even though some directions in parameter space may merge better than others. In a shared directional basis, the authors find two regimes: a low-dispersion common bulk shared across tasks and a high-dispersion, task-specific tail that carries a disproportionate share of the energy. Common-Unique Subspace Decomposition (CUSD) separates each task vector into these two parts, merges the common parts with an existing arithmetic operator, and keeps the unique residuals separately. Across 11 base models and four merging operators, it improves 42 of 44 model-operator combinations, with gains of 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 run a shared 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 sharing caches varies across (loop, layer) cells, and that per-cell damage roughly adds up to the total output damage. LoopKV uses a one-time offline profile to assign each cell 4-bit storage, 2-bit storage, or reuse of the final loop's KV. During decoding it reads the low bits of a 4-bit value only where attention needs that precision, and a single fused kernel in vLLM executes the plan. Serving Ouro-1.4B on one RTX PRO 6000, the memory-optimized plan stores 35% less KV than uniform 4-bit KIVI-4 and sustains 1.7-2.2x its goodput under a latency target, with no significant accuracy loss against BF16 on six benchmarks.

arXiv:2611.04137 PDF

AdSteer: Native Advertising in Large Language Model Generation

Shuhang Lin, Chuhao Zhou, Zhenyu Wang, Xiao Lin, Changsheng Wang, Saibo Geng et al. LLM services are expensive to run, and native advertising, which matches sponsored content to user intent, is one way to pay for free access. Existing approaches rely on prompt injection, multi-pass rewriting, or fine-tuning, each of which can hurt response quality or add compute. AdSteer instead works at decoding time, adjusting token probabilities on the fly and calibrating how strongly and for how long the advertising influence applies, so that ads read as a natural part of the response. On GEM-Bench across three application scenarios, it improves the overall score by 14.1% over the strongest baselines and uses 22.5% fewer tokens than rewrite-based methods.

arXiv:2611.04147 PDF

PACT: Certifying What Is Written in Parallel Decoding of Diffusion Language Models

Yanchao Li, Yusu Zhang, Jiaqing Xie, Yi Ban, Ben Gao, Zhongyu Huang et al. Masked diffusion language models speed up generation by writing several tokens per forward pass. The rules that decide which tokens to commit test the predicted distributions, so under sampling they never see the tokens actually drawn. The authors show that answer quality tracks how reliable the written tokens are, and propose PACT, a training-free commit rule that draws tokens first, certifies the set it will write, and spaces each step's positions apart. The certificate holds at every temperature, and they prove that no rule with the same guarantee can commit more positions. Across four LLaDA and Dream models on HumanEval, MBPP, GSM8K, MATH500, and IFEval, PACT matches published commit rules in quality with fewer forward passes and produces far fewer repeated words at the same speed.

arXiv:2611.04152 PDF

Length Is Cheap, Competition Is Not: Attention Predicts Context Rot in Retrieval

Yinuo Liu Failures of language models on long inputs are usually blamed on length, but this study of retrieval on five open models (0.5B to 3B parameters) finds otherwise. Neutral filler up to 32K tokens costs essentially nothing, while statements formatted like the evidence cut accuracy to 0.16-0.33, and a single statement that gives the queried key a rival value multiplies the loss by 9 to 110. Across 126,473 trials, the attention mass left on the evidence explains up to 58% of the variance in loss, against at most 8% for log length. The masked-filler protocol behind earlier claims that length alone hurts fails a visible control in retrieval, so the authors conclude that context rot in retrieval is a competition for attention, and they release their probe and all trial-level data.

arXiv:2611.04169 PDF

Coarse Semantic Blocks for Efficient Training-Free Sparse Attention

Sosuke Kobayashi, Shoichiro Yamaguchi, Seiya Tokui The training-free sparse attention method clusters keys into large semantic blocks drawn from non-contiguous positions, building a prompt-specific index. Each query scores the block centroids and attends to the original keys and values inside the blocks it selects. Queries that pick the same block are processed together, which keeps prefill fast. Large blocks make this cheap enough to attend to more tokens in the same time, so they can be both faster and more accurate than small blocks. On Qwen3.5-4B and Qwen3.5-35B-A3B on one H200, whole-model speedups over dense attention reach 2.3–2.4× for prefill and 1.7–2.0× for decoding at 512K tokens, and 2.9–3.0× for prefill at 1M. The 35B model stays close to dense accuracy on RULER, LongBench-v2, and MRCR.

arXiv:2611.04313 PDF

Text-to-SQL Models Make Choices Without Saying So

Yichen Xu, Shipei Lin, Chenhao Ma cross-listed Text-to-SQL models often settle open calculation choices, such as counting rows versus distinct entities, without telling the user, and benchmarks that score against a single reference cannot detect this. SpecBridge tests each choice with three otherwise identical requests: one that leaves the choice open and one for each interpretation. It scores every answer by executing it on real databases. Across 204 hand-validated tests, four models usually compute each interpretation when asked, but for the open request return both in only 15–58% of eligible cases, and on BIRD more than two in five answers marked wrong are correct under the other interpretation. A pipeline that first detects ambiguity and then answers each interpretation in a separate call improves full coverage by 29–63 points on these tests and on AMBROSIA, without losing precision on clear requests.

arXiv:2611.04318 PDF

Planning-Coding Separation in LLM-Based Code Generation: An Empirical Study

Wentao Wang, Zeyu Sun, Ruwei Pan, Yueyi Xiao, Qingyuan Liang, Lu Zhang The empirical study looks at code generation in which one LLM writes the plan and a different LLM writes the code. It tests many planner-coder pairings and the effect of further training the coder. Almost all combinations are as competitive as using the same model for both roles, and stronger planners paired with stronger coders generally do better. A large planner paired with a much smaller coder can sometimes match the large model doing both, and additional coder training typically helps across pairings, which suggests planning and coding can be separated to a considerable extent.

arXiv:2611.04320 PDF

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

Vimal William, Yuanzhe Li, Danella Zhao, Huanrui Yang Linearization turns a pretrained Transformer into a cheaper model by combining sliding-window attention (SWA) with a fixed-size recurrent memory, here Test-Time Training (TTT). Studying such hybrids, the authors find that long-context retrieval is concentrated in a few anchor layers, while the other layers' states hold similar information transformed in layer-specific ways. Anchor therefore places TTT memory only at the 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, retains 91-95% of the teacher model's MMLU accuracy, and 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. Recovery recipes that keep accuracy intact during end-to-end FP4 reinforcement learning (RL) post-training of dense models give no improvement over naive FP4 when applied to Mixture-of-Experts (MoE) LLMs. The authors trace this to the router: because expert selection is discrete, the probabilities computed during training and during rollout inference stay aligned only if the same correction is applied on both sides, and asymmetric recipes sharply degrade that alignment and the rate at which the router picks the same experts. Their recipe, MoeResQ, applies corrections symmetrically on the training and inference sides, restoring alignment and substantially closing the accuracy gap to full precision.

arXiv:2611.04376 PDF

Twice as Fast, Until It Isn't: Quantized Serving of Mid-Size Open LLMs

Smit Kadvani, Kriti Faujdar, Spurthi Tallam Teams serving a 9B-20B open-weight model on vLLM or SGLang need to know both the speedup a quantization recipe gives over BF16 and the accuracy it costs, so the authors compiled every public measurement of these numbers they could verify. The record is thin: only two Gemma 3 12B INT4 checkpoints have both a vLLM timing and an accuracy score, AMD and Intel Gaudi have no timings in this size range, and FP4 has none against BF16. The timings that exist follow one rule: 4-bit weights roughly double speed when small-batch decoding is limited by memory bandwidth (2.12x for Qwen2.5-14B on an A100) but give almost nothing at large batch sizes (1.02x at batch 256), where FP8 can still help. Many published accuracy deltas, including every HumanEval delta, fall within sampling noise, and the paper closes with a seven-item reporting checklist.

arXiv:2611.04380 PDF

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

Xin Liu, Boyuan Sun, Yuxin Wang, Haoran Song, Mingjian Zhao, Mengyu Liu A survey of 291 open-weight language model repositories finds that 52.6% of untied gated and hybrid checkpoints contain reachable input embeddings with norms below 1e-6, against none of the weight-tied models; these near-zero rows form when tokens never seen in pretraining get no gradient while AdamW weight decay keeps shrinking them. In gated sequence models such as Mamba, Gated DeltaNet, and xLSTM, such a token can inflate embedding gradients by up to 35 orders of magnitude, enough to overflow FP32 global-norm gradient clipping in PyTorch, Hugging Face Trainer, and FSDP2 and silently zero that batch's update. In Telugu continual pretraining of Nemotron-3-Nano-4B, this discarded updates on about 20% of steps. The authors propose GroupClip, which clips input-embedding gradients separately with a 64-bit or rescaled norm and restores updates for existing checkpoints at negligible cost, and GateFloor, a small positive offset on the SiLU gate that prevents the problem during pretraining without hurting perplexity.

arXiv:2611.04449 PDF

SaMA: Morpho Adaptation via Asymmetric Expansion of Kronecker Product

An Nguyen, Anh Tong Low-rank adaptation (LoRA) has two structural weaknesses: its symmetries create flat directions in the loss landscape that slow adaptive optimizers, and its effective rank grows only linearly with the parameter budget. Scalable Morpho Adaptation (SaMA) generalizes the Kronecker product by relaxing a shared-block constraint in its factorization, so the adapter can range from low to full rank and its practical rank grows quadratically with the parameter budget. The authors also show that its symmetry group reduces to a diagonal group, making it easier for optimizers. On commonsense and arithmetic reasoning benchmarks, SaMA outperforms strong parameter-efficient fine-tuning (PEFT) baselines while using fewer parameters.

arXiv:2611.04462 PDF

Evaluating LLMs as Judges of Counterspeech Quality Against Online Hate Speech

Johnny Alali, Nicol\'as Benjam\'in Ocampo, Davide Ceolin The authors test whether LLMs can replace human experts in rating the quality of counterspeech against online hate, evaluating 12 open-weight models with zero-shot, few-shot, and chain-of-thought prompting on relevance, coherence, aggressiveness, and suitableness. Agreement with experts is low: the best setup, Qwen 2.5 72B on coherence, reaches only a weighted Cohen's kappa of 0.19 and Spearman correlation of 0.42 in pointwise scoring. Converting pointwise scores into pairwise labels substantially improves agreement for some configurations, such as Mixtral 8x7B rising from 0.07 to 0.34. LLMs agree far more with each other than with humans: OLMo 2 32B and Llama 3.3 70B reach kappa 0.76 with each other but no more than 0.21 with experts.

arXiv:2611.04627 PDF

Self-Privileged Fine-Tuning

Yangrong Hu, Xueyu Zhou, Yue Wu, Jian Huang Recent supervised fine-tuning (SFT) variants weight tokens using the model's own output distribution, and the authors analyze which target distributions these schemes actually fit. They show that dynamic weighting such as Dynamic Fine-Tuning (DFT) induces a heavily sharpened target, while static weighting fits the data distribution times a frozen prior, which is more stable but only as good as the prior. Self-Privileged Fine-Tuning (SPFT) uses as that prior a frozen copy of the initial model conditioned on privileged information such as the ground-truth answer, which upweights tokens that are most predictive of the correct outcome. SPFT improves over standard SFT and other baselines on mathematical reasoning and code generation.

arXiv:2611.04632 PDF

When Ground Truth Is Not Ground Truth: Automatic Labeling for Hallucination Detection

Rounak Sen, Divya Kumari, Deeksha Varshney Hallucination detectors for large language models (LLMs) are often evaluated against labels created automatically by thresholding the similarity between a generated answer and a reference, commonly with BLEURT. The authors test 15 automatic labeling judges with seven detectors, two generators, two QA datasets, and five languages, and find that detector rankings shift substantially with the choice of judge. Human validation shows BLEURT-based labels are wrong about 47% of the time, compared with 12.86% for the best judge, QGen-9B. The results suggest that hallucination benchmarks built on similarity-based labels may be unreliable.

arXiv:2611.04697 PDF

When Should an LLM Steering Edit Depend on the Input?

Rachael Harkness, Konrad Rawlik Contrastive Activation Addition (CAA) steers a large language model by adding one fixed direction, the difference between class means, to every input's activations. The authors study steering edits that depend more on the input, up to retrieving a per-query destination from the mean of the k nearest stored states within the requested class. They find that input-dependent retrieval beats fixed-direction steering at every layer when the target varies within the class. Word-level edits work best in the first tenth of the network with a few neighbors, and clause-level edits work at any depth using the class mean. Four quantities computed from the activations alone indicate which regime applies.

arXiv:2611.04757 PDF

Accelerating Offloading-Based MoE Inference by Interleaving Concurrent Requests

Hongseung Yu, In Gim, Kyunghan Lee, Lin Zhong cross-listed Running a large mixture-of-experts (MoE) model locally with expert offloading stalls the GPU while missing experts are fetched. The common fix of predicting and prefetching experts can actually hurt when an expert cache is present, because prediction gets less precise as the cache fills with the experts that are needed. Bevy instead uses runahead inference across concurrent requests: requests whose experts are already available proceed to later layers, revealing which experts they need next, while fetches for other requests overlap with computation. A self-balancing scheduler batches ready requests at each layer and prioritizes those that are lagging. On two local platforms, Bevy achieves 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 MobileSpec adapts an on-device LLM into small task specialists using forward passes only, which avoids the memory cost of backpropagation. It learns from automatically checked task outcomes, in the style of reinforcement learning with verifiable rewards (RLVR). Its core algorithm, FuRA-S, trains only the singular values of column blocks of pretrained weights. This produces random perturbations that rotate singular vectors, mimicking how RLVR updates weights, and it can adapt INT4 models without modifying their weights. With specialists of 1-5 MB that update 0.015-0.05% of parameters, MobileSpec 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. On RLVR benchmarks it is competitive with full-parameter evolution strategies.

arXiv:2611.04797 PDF

Asymptotic Optimality of $c\mu$-Rule for LLM Inference Scheduling under Memory Constraint

Zhenghao Li, Shengbo Wang, Nian Si cross-listed LLM inference scheduling is limited by key-value (KV) cache memory, especially for the long prompts and outputs of agentic and reasoning workloads. The authors model the system as a queue in which step duration depends on prefill computation, KV-cache input/output, and memory occupancy. A fluid approximation reduces it to a multiclass single-server queue, which yields a c-mu priority rule whose effective service rate combines compute speed, memory bandwidth, and memory capacity. The rule is proven to minimize cumulative holding cost in the fluid model, and a feasible policy for the finite system is shown to converge to it. Experiments on a real serving engine on an NVIDIA H800 show substantially lower average latency than existing schedulers while keeping throughput high.

arXiv:2611.04807 PDF

Feeling of Knowing in Large Language Models

Zichuan Fu, Xian Wu, Jingtong Gao, Wenlin Zhang, Jiaxuan Li, Binhao Wang et al. After post-training updates change what a large language model (LLM) knows, its judgment of whether it can answer a question may no longer match what it can actually answer. The authors frame this judgment as a functional feeling of knowing (FoK) that can drive selective answering, retrieval triggering, and model routing. They train two separate channels in three stages: on-policy supervised fine-tuning (OP-SFT) to update knowledge, knowledge-delta FoK training (KΔ-FT) to keep the judgments in step with the changing knowledge, and FoK-guided policy optimization (FGPO) to train the final answer or explanation. On factual QA, in cross-dataset tests, and over three sequential updates, explicit FoK alignment 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. Using NVFP4 4-bit weights and activations during reinforcement-learning post-training lowers rollout quality and widens the mismatch between the training and inference engines, which destabilizes training. The authors find that these failures are concentrated in a few outlier sequences and a few high-variance activation channels. FourSight filters out the drifted trajectories before policy updates, keeps only the top-variance channels in BF16 while running the rest in native FP4, and adds a fused SpliceGEMM kernel to keep the overhead low. It reaches near-BF16 accuracy with strong FP4 deployment quality, improves step-to-accuracy by 1.76–2.60x, and stays within 2.3% of plain FP4 step time 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 (LLM) on a narrow task often erodes its general abilities, which the authors call Pervasive Capability Degradation (PCD). They explain PCD through feature interference under superposition: narrow fine-tuning data disrupts the nearly orthogonal features learned in pre-training and pushes activations across decision boundaries on general tasks. The theory groups existing mitigation methods into six paradigms, and the authors prove that cross-entropy with one-hot targets cannot fix PCD alone. Their AEGIS loss directly limits drift in activation margins against all competing tokens and, across 3 model families and 14 SFT tasks, beats representative methods on both target and general capabilities at about 1.3 times the cost of standard SFT.

arXiv:2611.05003 PDF

When Self-Verification Hurts: Calibrated Answer Repair for Financial Question Answering

Celine Guo, Weida Liang, Yujian Fu, Anqi Dang, Yanzheng Jin, Kenji Kawaguchi et al. In financial question answering, a model may extract a plausible number but attach it to the wrong entity, period, unit, or accounting concept. Open-ended self-verification can then damage correct content or reinforce the original error. Calibrated Answer Repair (CAR) first estimates whether a draft is likely to benefit from repair, then picks a targeted check, localizes the relevant evidence, and limits the permitted edit. Across FinRank, Fin-RATE, and FinQA with three open-weight models, CAR beats generic verification in every setting. It recovers 14.6% of initially missed answer requirements while damaging only 2.1% of correct ones, with over 30% lower post-draft token cost than iterative revision.

arXiv:2611.05166 PDF

ARLoop: Fully-Open Pre-trained Looped Transformers with Attention Residuals

Yuekun Yao, Yupei Du, Dawei Zhu, Huan Sun Looped transformers reuse a shared set of layers several times to gain effective depth without adding parameters, but standard residual connections merge all intermediate representations into a single stream. ARLoop replaces them with attention residuals, so each layer selectively attends over earlier depth states; a sliding window over recent depth matches or beats full-history attention. In controlled experiments at 140M and 370M parameters, it consistently outperforms prior residual designs for looped transformers, with gains carrying over to realistic pretraining at 0.6B and 1.2B parameters, especially on math and coding. The fully open 1.2B model, trained on only 0.5T tokens with three recurrences, beats Huginn-3.5B on every evaluated task and is competitive with Qwen3.5-2B.

arXiv:2611.05234 PDF

Going Deeper With Memories

Afshin Khadangi Byte-level language models avoid tokenizer dependence but pay a high cost for processing every byte at full resolution. DeepMem is a causal multiscale architecture that separates semantic compression from exact byte preservation: a learned invertible subband transform builds a hierarchy of shorter sequences, a lossless carrier plus detail sidebands keep byte identity, fast-weight memory blocks run along the encoder with iterative shared-state computation at the compressed bottleneck, and an attention-based super-resolution decoder restores byte resolution. It beats a matched byte-level Transformer on FineWeb-Edu and PG19, leads on macro-average across the evaluated FLORES languages, and improves PIQA accuracy while staying competitive on code and corruption robustness. Bottleneck refinement uses roughly 7,600x fewer counted FLOPs than equivalent full-resolution reasoning at competitive throughput.

arXiv:2611.05255 PDF

Beyond Accuracy: What Language-Subspace Interventions Capture, Remove, and Improve

Shaanil Punglia, Dadhichi Telwadkar, Vaibhav Rathore, Siddharth Sureka, Neha Bhargava Language-subspace interventions try to improve multilingual reasoning by suppressing language-related directions in mid-layer activations, and this study audits whether they do what they claim using MATH and MGSM across four open-weight LLMs. Subspaces built with the existing Mean-SVD construction and a new DiffPCA both capture substantial translation-associated variation, but after projecting them out, linear probes still recover the language almost perfectly, so neither subspace exhaustively isolates language information. Evaluating intervention configurations on Qwen2.5-7B, the setting that most improves MGSM accuracy sacrifices response-language fidelity, with only one configuration improving joint accuracy (correct and in the target language) by 1.60 points; small prompt changes alone reproduce the accuracy gains while badly damaging fidelity. The authors argue causal interventions must be judged on more than accuracy.

arXiv:2611.05278 PDF

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

Minsoo Cheong, Sungjoo Yoo Looped transformers reuse the same layers across loop steps to match much larger models at lower parameter cost, but each step stores its own keys and values, so the key-value (KV) cache grows linearly with loop count and becomes the memory bottleneck. Observing that the KV cache is highly redundant along the loop axis and that steps differ in quantization sensitivity, LoopVQ vector-quantizes channels across multiple loop steps jointly and fine-tunes codebooks to match full-precision outputs. At 0.5 bits per element it beats existing KV cache vector quantization methods on general, long chain-of-thought reasoning, and long-context retrieval tasks with the Ouro and Huginn models, and combined with the orthogonal TaSQ retains much BF16 accuracy at 0.25 bits. Implemented in SGLang, it nearly halves decode latency at 32K 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

You don't need long context

Peter Belcak, Saurav Muralidharan, Jindong Jiang, Yonggan Fu, Wonmin Byeon, Hanrong Ye et al. Later tokens cost more to generate than earlier ones, by as much as 7x on frontier open-weight models, which motivates asking whether long contexts are needed at all. Using frequent compaction across a broad range of long-horizon and long-input tasks, the authors find that frontier language models sacrifice little to no task accuracy while maintaining short contexts. Performance also does not collapse on long-horizon software engineering tasks even when the agent takes at most one action per context, which the authors use to sketch the idea of semantic state space models (SSSMs) and to question the need for long-context models.

arXiv:2611.05306 PDF

A Circuit for In-Context Learning

Andrew Gao, William Gao Prior work found candidate components of in-context learning (ICL) but not how they combine. The authors trace the full circuit for few-shot sentence classification with arbitrary labels in three 3-4B-parameter transformers. They find that the model implements a gated key-value lookup: middle-layer heads write the class of each demonstration into the following label token, upper-layer heads match the query's class against these keys and copy out the label, an early signal for repeated labels gates the readout, and separate heads suppress the most recent label. Each step is established causally through activation patching, ablations and steering. The same circuit appears across model sizes from 0.6B to 8B, survives instruction tuning, and grows 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 The authors independently re-implement the practical variant of Direct Nash Optimization (DNO), which aligns a model to general preferences through on-policy sampling, LLM rating against a teacher's gold response, and DPO-style contrastive updates. They study it on Llama 3.1 8B Instruct and Mistral 7B Instruct. Three iterations raise the AlpacaEval 2.0 length-controlled win rate from 31.7% to 45.3% for Llama, putting the tuned 8B model above Llama 3.1 70B Instruct. Same-family open-weight annotators retain 62-82% of the gain achieved with GPT-4o as annotator. Ablations show gold responses and data look-back matter far more than sample count or margin, though all results are single runs and GPT-4o is both teacher and judge.

arXiv:2611.05373 PDF

11 more specialized papers

Other 60

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

Snehasis Addy, Eduardo Calle Ortiz, Prashant Shenoy, Phuc Nguyen cross-listed HyVox performs low-energy open-set speaker identification by splitting work between a neuromorphic chip and a conventional (von Neumann) microcontroller. A spiking neural network (SNN) backbone on the chip turns speech into a 32-spike-count embedding, and the microcontroller scores speakers by cosine similarity, rejects unknown voices and enrolls new users without any weight update. To fit the chip's limits, the backbone is compressed 6.5x and fine-tuned with chip-accurate spiking training (CAST), which trains through a differentiable replica of the chip's integer neuron update and doubles accuracy on unseen speakers. On Common Voice households, the system reaches 99.6% top-1 accuracy with 2 members and 94.9% with 20, using 1.63 mJ of dynamic energy per 3-second window, 13.5x less than the most efficient embedded neural-network baseline.

arXiv:2611.03735 PDF

Structure, Dimensionality, and Distribution Shape - What Makes Pretraining Data Transferable, and Why the Input Pipeline Decides Which

Sangyoon Bae, Jiook Cha The authors describe a pretraining corpus by three properties (structure, dimensionality, and the distribution shape of its values) and propose a Transfer-Invariance Principle. The principle says the model's input transform, not the data's modality, determines which of these properties matter for transfer. They test it with TransferForge, a synthetic data generator that can set each property independently. When inputs are standardized, as with PatchTST on UCR, distribution shape has no effect and synthetic pretraining matches or beats real data. When raw calcium-imaging signals are fed in unnormalized, as with POYO Plus, shape becomes decisive, and a shape-matched synthetic corpus replaces real neural data entirely. With model and data held fixed, toggling only the input normalization changes the effect of a shape manipulation 5.3-fold.

arXiv:2611.04106 PDF

Deep Learning Compilers: Architecture, Optimization, and Empirical Characterization

Jinghao Wen, Ruixuan Wang, Xun Jiao The survey describes deep learning compilers as three stages: a frontend, a middle-end, and a backend, with the middle-end bridging high-level model graphs and hardware-specific optimizations. It reviews representative compilers and optimization techniques at each stage. Experiments with PyTorch/TorchInductor on an NVIDIA RTX A6000 show that speedups reach 8.21× for fusion-friendly graphs but are small for other workloads. Compilation overhead can take thousands of runs to pay back, and dynamic-shape compilation cuts cumulative shape-transition cost by about 70% at the price of slightly slower steady-state performance.

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 Aggregate continual-learning scores implicitly average over task orders that are rarely stated. The authors frame algorithm comparison as an expectation over an explicitly declared order distribution and apply it to motion forecasting on four road-layout tasks from the INTERACTION dataset, covering all 24 orders, four algorithms, and three replicates (288 runs). Task order explains 65–91% of the variation in current-task error and backward transfer, and almost all of that comes from which task comes first, because the smallest task cannot be learned from a cold start. In most orders the leading algorithm is within one replicate standard deviation of the runner-up, so the authors call for benchmarks to report order-resolved, starting-task-resolved, and final errors alongside aggregates.

arXiv:2611.04309 PDF

Spin Spherical Fourier Neural Operators: Efficient Equivariant Learning for Vector and Spin Fields on the Sphere

Kevin Mulder, Matthew A. Price, Eirini Leftaki, Jason D. McEwen cross-listed Many physical fields on the sphere, such as wind, cosmic microwave background polarization, and weak-lensing shear, are spin fields whose values change phase when the local frame rotates. The authors extend spherical Fourier neural operators (SFNOs) to fields of any spin, building spin SFNOs (S2FNOs) from lifted spherical convolutions that are computed exactly and efficiently in harmonic space. The layers are rotationally equivariant to machine precision, the models reach state-of-the-art results on existing benchmarks, and on new benchmarks introduced here spin-aware models outperform their scalar counterparts.

arXiv:2611.04389 PDF

Formalizing PTX in Lean to Prove GPU Kernel Correctness

Luk Koska, Tom Koska, Ben Koska The authors formalize PTX, NVIDIA's low-level GPU instruction set, in the Lean theorem prover so that correctness of deep-learning GPU kernels can be proved mathematically. The model covers patterns common in such kernels, including acquire-release synchronization, tensor cores, proxies, asynchronous operations, multi-kernel pipelines, and data-dependent control flow. Using it, they prove the correctness of simple implementations of a WGMMA matrix-multiplication kernel, FlashAttention, and a three-kernel mixture-of-experts (MoE) pipeline.

arXiv:2611.04447 PDF

Perseus: A Full-Stack System for Homomorphic Model Inference

Alessandro Zirilli, Davide Marincione, Evgenios M. Kornaropoulos, Giuseppe Ateniese, Emanuele Rodol\`a cross-listed Fully Homomorphic Encryption (FHE) lets a server run inference directly on encrypted user data, but it has been very slow, and most existing frameworks are closed-source. Perseus is an open Python library covering the full encrypted-inference stack, with one new optimization at each level: in-kernel regeneration of key-switching keys in the cryptographic backend, a complex variant of Cachemir packing for data layout, and a bootstrap placer that accounts for both packing and value magnitude at the model level. It reaches state-of-the-art latency across several models and modalities while staying within 3% of plaintext accuracy and perplexity, and it is 3.96× faster at single-token decoding on Llama-3-8B than any other implementation. The authors also launch a public speedrun on encrypted GPT-2.

arXiv:2611.04488 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 have to choose between structured latent dynamics with limited generative flexibility and flexible diffusion or flow-matching models that need slow, iterative sampling. The proposed Koopman Generative Operator (KGO) has three parts. Koopman Patch Embedding (KoPE) extrapolates latent trajectories. Koopman Flow Matching (KoFM) generates samples in a single step through a closed-form matrix exponential in a Koopman latent space. An Adaptive Uncertainty Gate (AUG) adjusts uncertainty for each variable and each forecast horizon. KGO reaches 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), with at least a 25x reduction in inference time 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 how confident the batch is about each gradient sign, ZEMA (Z-score Evidence Momentum Adaptation) measures that confidence directly. For each coordinate it computes a z-score from the per-example gradients in the batch, converts it into a sign probability, and smooths the result with a single exponential moving average. Forming the probability before smoothing means ZEMA keeps one optimizer state per parameter instead of Adam's two, at some extra per-step compute; 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 and gives less variable validation loss.

arXiv:2611.04987 PDF

Beyond Foundation Models: PFN Priors Improve Tabular Hyperparameter Optimization

Artem Chistyakov, Renat Sergazinov, Artem Babenko Synthetic data priors have mainly been used to pretrain tabular foundation models. PriorTune instead uses them to meta-tune hyperparameters for conventional tabular models 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 win rate, so a new dataset only needs to fit that small subset. The authors find that scaling the number of configurations and the number of synthetic datasets have equal effects, and they explain this theoretically. On BeyondArena, LightGBM with TabPrep features and PriorTune configurations ranks above TabPFN-3 and RealMLP while being much cheaper at inference.

arXiv:2611.05141 PDF

Personas Are Not People: Identity as State, Not Attribute

Tamunotonye Harry, Yuanyuan Feng Synthetic populations of persona-driven language models are increasingly used to pretest surveys, simulate users, and evaluate AI systems, and every such system assumes identity is a fixed attribute that behaves the same in any situation. Drawing on personality research, which finds within-person variation across contexts exceeds between-person variation, and on identity economics, the authors decompose behavior into stable traits, the identity a context activates, and the situation, then prove under two stated assumptions that any simulation built on fixed personas cannot reproduce the context-specific part of a population's behavior, no matter how detailed the personas or how large the model. They propose an evaluation criterion and a person-in-context design, tested on a naturalistic Reddit personality corpus and a published clinical-trial recruitment survey; the design does produce context-dependent behavior but matches neither real population fully. The practical implication is that synthetic-population evaluations most misrepresent exactly the groups whose behavior is most context-shaped.

arXiv:2611.05292 PDF

49 more specialized papers

Agents 57

rTaste: Do Research Agents Have Taste? Evaluating Early Stage Research Judgment

Art Kanke cross-listed rTaste tests whether research agents have "taste", meaning the judgment to pick which experiments are worth running before the results are known. Six frontier models, working alone or in teams, propose minimum viable experiments (MVEs) for seven recent NLP papers, forecast their outcomes, and rank them. A separate executor then runs every MVE, and model auditors judge the results. Of 415 MVEs, 194 were judged sound, valuable, and new, but only 73 were also non-obvious and went beyond the seed paper. The agents' forecasts beat a base rate, but their rankings were not shown to beat simply running the first-written experiment, and teams cost more without clearly better proposals.

arXiv:2610.40488 PDF

MCPGuard: Finding Conformance Gaps in MCP SDKs via Dual-Track Analysis and Dynamic Adjudication

Xikai Sun, Kebin Liu, Fan Dang, Yawen Zheng, Zihao Yang, Yunhao Liu The official SDKs for the Model Context Protocol (MCP), which connects LLM agents to external tools, may leave out constraints from the specification, and those gaps can carry over into downstream applications. MCPGuard is an LLM-based framework that looks for such gaps in two ways: a structural track that matches JSON Schema constraints against the SDK's data models, and a behavioral track that uses LLMs to audit function-level code against the rules written in the specification's prose. A dynamic adjudicator then reads the full source and builds proof-of-concept exploits to confirm each candidate. On the official Python 1.26.0 and TypeScript 1.29.0 SDKs it confirmed 13 and 19 conformance gaps respectively, none of which two black-box baselines found, and tests on open MCP servers show these gaps reaching downstream code.

arXiv:2611.03662 PDF

A local orchestration harness for molecular design with foundation models

Sakib Mostafa, Md Atik Ahamed, James Zou, Lei Xing, Md Tauhidul Islam cross-listed MolCrew is a locally deployable system for molecular design in which a language-model orchestrator coordinates four separately trained specialist models: one generates structures from text descriptions, one describes structures, one predicts biological endpoints, and one proposes edits. RDKit calculations check structural validity and constraints, and their results guide the orchestrator's next tool call. The orchestrator was fine-tuned with low-rank adaptation (LoRA) for tool-use decisions and recovering from failures, while the specialists' weights were left unchanged. On 3,300 ChEBI-20 test examples it achieved 96.4% exact structure recovery versus 38.7% for published Atomas-large results, reached a mean area under the ROC curve of 0.851 on eight MoleculeNet datasets, and passed every predefined workflow check on 91.5% of 1,699 recorded requests.

arXiv:2611.03676 PDF

LOLPlay: Benchmarking Real-Time Visual Agents in League of Legends

Zhaojian Yu LOLPlay is a benchmark for visual agents that have to reason and act while the environment keeps changing during inference. It runs inside the native game League of Legends and has 36 champion-specific tasks covering resource management, combat, cooperation and objectives. Agents see screenshots and act through mouse and keyboard, and 82 executable subgoal checks separate partial progress from full completion. Across 324 evaluations of three models at three reasoning-effort settings, GPT-6 Astra scored highest at 25.0% task success, but high reasoning effort never improved on low effort for any model, because longer requests left fewer interaction rounds and actions often ran after the scene had already changed.

arXiv:2611.03685 PDF

Training LLM Agents to Automate LLM Post-Training with RSI-Dojo

Shuzheng Gao, Zhaojian Yu, Penghao Yin, Yuling Shi, Kai Cai, Shilin He et al. As one step toward recursive self-improvement (RSI), the authors want to train LLM agents to carry out model post-training themselves, rather than only optimizing the workflow or harness around them. RSI-Craft automatically turns papers from top venues into executable training environments, producing RSI-Dojo, a set of 281 seed tasks filtered for reliability and for how well they separate stronger from weaker agents. Trajectories collected from frontier models were filtered by rubrics and by outcomes, then used to fine-tune Qwen3.5-9B and Qwen3.6-35B-A3B, which gained 8.74 and 16.87 points on PostTrainBench. RSI-Dojo-35B-A3B reaches 29.95 on PostTrainBench, matching frontier models such as GLM-5.2 and GPT 5.5.

arXiv:2611.03698 PDF

Unsatz: Evaluating Scientific Discovery Without a Given Ansatz

Ethan Ji, Wei Wang, Yizhou Sun Most benchmarks for automated equation discovery give the solver a predefined model form (an ansatz), such as fixed variables or a library of candidate terms. Unsatz instead requires an LLM agent to build that structure itself. Each task is a simulated system generated as a program, which can involve differential equations, discrete maps, causal models or stochastic processes, sometimes with hidden states or distorted sensors. The agent buys experiments within a budget, writes candidate programs, and is scored on how well its final program predicts held-out extrapolations and interventions. Across 80 systems, nine language models solve between 21% and 60%, while classical equation-discovery methods almost never succeed on the 44 systems where they were tried, even though the data usually suffice once the true structure is given. The authors also observe that agents identify a system's variables but not how they interact, and prefer revisions that lower training error over revisions that predict better.

arXiv:2611.03713 PDF

Counterexample-driven autonomous research with application to excess radiated energy in a Z-pinch

Bowen Zhu, Jian Wu Simjecture is an open-source framework in which language-model agents investigate physical conjectures that humans define, adapting the open-ended search used in AI-assisted mathematics to physics. Inquiry is organized around finding counterexamples and then repairing the hypothesis. Computational physics simulations serve as the executable testbed, and checking that the numerical tools are trustworthy is kept separate from reviewing the scientific claims. In an 82.6-hour autonomous investigation of why Z-pinch plasmas radiate more energy than the imploding shell's kinetic energy, the agents found competing effects: stronger helical deformation increases resistive conversion while reducing radiation, and a broader trailing-plasma profile delays the later radiation peak.

arXiv:2611.03721 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. Deep Research (DR) systems are LLM agents that break questions into subquestions, search the web, and combine the evidence into an answer, and small variations in inputs at any stage can push them to a wrong answer. An empirical study of 1,932 execution trajectories from three DR systems identifies nine recurring defect categories across subquestion formulation, search and final reasoning. Building on these defects, EMBERS fuzzes correctly answered seed questions with stage-targeted mutations, uses web search to check that mutated inputs keep their meaning, and maintains a diverse seed pool. Across 16 settings it achieves 3.21 times the failure-discovery throughput of the strongest baseline at the same budget, and fine-tuning on the generated inputs makes model answers more consistent.

arXiv:2611.03723 PDF

EvoMemState: Evolving Memory State for Personal Agents

Zhishang Xiang, Yunbo Tang, Zhongqiao Liu, Junkai Chen, Zerui Chen, Qingguo Hu et al. cross-listed Long-term memory systems for LLM agents usually store user facts as independent items, without tracking how new facts update or conflict with old ones or how they apply to the current request. EvoMemState adds two kinds of state to stored facts. Relational States, built by ontology-guided extraction, track each fact's validity, scope and changing relations. A Reasoning State, generated per request by a policy trained with reinforcement learning, decides how the facts should be applied. On long-term dialogue and memory-use benchmarks, the framework consistently improves retrieval completeness and the accuracy of memory-grounded responses over existing memory systems.

arXiv:2611.03743 PDF

RuntimeTickets: Source-Grounded Security Obligations for Agent Skills

Kuiyi Gao, Xue Liu, Bowei He cross-listed A malicious third-party Agent Skill can hide data leaks or output tampering inside an otherwise legitimate workflow, and neither reading its description nor granting it tool permissions shows whether a run meets the user's security requirements. RuntimeTickets has a language model propose security obligations from the task's delivery contract, each citing the Skill lines it addresses. These obligations can only narrow the host's approvals into fine-grained allowed actions, never widen them. A runtime gate then checks the recipient, object and amount of every outbound send, and a deterministic decoder verifies each obligation; no language model takes part in enforcement, so text injected by a Skill cannot change the rule that decides a send. Tested at the network egress of three unmodified agent frameworks against production Skill scanners and runtime baselines such as SkillGuard, it reached a 0.0% rate of letting malicious Skills through and a 0.0% rate of blocking benign ones, while the scanners made both kinds of error.

arXiv:2611.03772 PDF

El Agente F\'armaco: A Scalable AI Scientist for Drug Discovery

Marcel M\"uller, Mohammad Ghazi Vakili, Yoshihiro Ogura, Ella Ashoori, Shimeng Wang, Alessandro S. Nascimento et al. El Agente Fármaco is an LLM-based AI scientist that turns drug-discovery goals written in everyday language into large computational screening workflows, choosing tools, reviewing results and adjusting next steps. It is built around a shared database, Fármaco Atlas, that stores molecules, structures and results with a record of how each was obtained. The agent retrieves evidence through targeted queries, and a single request can launch calculations over a whole collection of molecules without an LLM decision per compound, which keeps the model's context small as screens grow. In a 22-task benchmark against plain tool calling and Codex, it achieved the highest mean model-judged score with lower median token use and peak context, and in a retrospective comparison with an expert-led campaign it proposed distinct candidates with competitive docking scores.

arXiv:2611.03832 PDF

Performance, Scalability and Maintenance: An Experimental Study of RPA vs. Autonomous Agents in the Healthcare Domain

Daniel Schulz, Pedro Cabral Buchaim, Bernardo Soares, Fernando Rafael Ensslin Schwengber, Lorenzo Farias, Daniela Duarte da Silva Bagatini et al. Healthcare data often has to be pulled from fragmented legacy web portals. The authors compare deterministic Robotic Process Automation (RPA), built with Playwright, against an LLM-based agent approach (Agent Process Automation, APA) built with browser-use, LangGraph and Gemini 2.5 Flash-lite, on retrieving mammography and breast ultrasound reports from a lab portal. Over two rounds a week apart, RPA reached 98.90% accuracy versus 96.49% for the agent, was about 13 times faster (16.07 s vs. 207.21 s per patient) and about six times cheaper per report, and the agent's results varied more over time. For a stable interface, RPA came out ahead.

arXiv:2611.03853 PDF

EvoGroup: A Self-Evolving Harness for Group Agents in Shared Contexts

Xinyi Xu, Xiaoyang Li, Yucheng Feng, Wenqing Chen, Youliang Yuan, Wei Shi et al. Agents deployed as long-term team workers operate over shared contexts that build up across many users, but existing memory systems are designed for a single user and miss group relationships and collectively formed knowledge. EvoGroup (EvoG) starts from a minimal six-tool harness and evolves it automatically: it selects informative trajectories using confidence and task outcomes, prompts self-reflection on low-confidence cases, and passes the selected trajectories to an analysis agent whose findings guide an evolve agent in revising the harness. On a held-out split of EverMemBench, pass@1 rises from 71.07% to 89.58%, beating self-evolving memory baselines and human-designed harnesses including Claude Code (85.83%). Frozen and tested on the unseen GroupMemBench with GPT-5.5, the evolved harness still outperforms both kinds of baseline.

arXiv:2611.03868 PDF

What Is This Memory Useful For? Measuring Task-Conditioned Utility in Long-Term Agent Memory

Chih-Hsuan Yang, Anjir Ahmed Chowdhury, Feng Yan, Rajeev Thakur Instead of giving each stored agent memory a single importance score, the authors measure how much it helps with a specific later task: a writer stores the memory without seeing the task, a frozen reader answers with and without it, and a ledger records the costs of forming, storing and exposing it. On factual-recall tasks built from LongMemEval-V2 web-agent memories, memories containing the needed information give a +0.33 higher paired reward gain than matched controls, and raw memories beat task-blind summaries and structured records even at a fixed total context size. Generic retrieval misses most of the memories that contain the answer, and none of the memories, raw ones included, measurably help on the software-engineering and scientific-coding benchmarks tested. The authors release a versioned Conditional Memory Utility Dataset of paired measurements and costs.

arXiv:2611.03930 PDF

Measuring Model-Specific Alignment Drift in LLM-Based Opinion Dynamics

Nata\v{s}a Djurdjevac Conrad, Sebastian Pokutta, Christof Sch\"utte Studies that use large language model agents as synthetic participants in opinion-dynamics experiments generally assume the prompt fully determines behaviour, and this work tests that by having agents debate a contested policy issue over repeated rounds while a fixed psychometric instrument, independent of any dynamics model, measures their stance. Without peer input, committed personas are intrinsically stable: across five models every agent held its assigned side for 20 rounds, so any loss of alignment comes from interaction and is strongly model-specific — gpt-4.1 keeps opinion clusters separated, claude-sonnet-5 partially, while gpt-5.6-sol and claude-opus-5 abandon them. One model maintained separated clusters under stance-homophilous selective attention, and the qualitative profiles persisted at doubled population and horizon and on a second issue, which the authors read as making the choice of model a first-class validity variable for social simulation.

arXiv:2611.03980 PDF

SLATE: A Large-Scale Benchmark for Long-Horizon API Selection in E-commerce Tool Use

Rongzhe Wei, Ge Shi, Min Cheng, Sarthak Ghosh, Vaibhav Gorde, Leman Akoglu et al. Enterprises want agents to execute standard operating procedures across internal software, which means picking and composing the right calls from thousands of APIs over a long horizon without hand-written step-to-API mappings — a regime existing tool-use benchmarks miss by exposing only a few tools, grading single invocations, or scoring multi-step behaviour with subjective LLM-as-a-judge rubrics that punish valid alternative routes. SLATE (Synthetic Large-scale API Toolkit for E-commerce) pairs queries decomposed into hierarchical tool-grounded procedures over a roughly 1,000-tool e-commerce library with a deterministic context-aware simulator and an end-to-end success metric that counts a task solved only when every reference outcome is reached, which credits diverse-but-valid trajectories while staying automatic and reproducible. Across six open decision models and their underlying language models from 3B to 12B parameters, the best model finishes only 5.0% of tasks. The benchmark, simulator, and evaluation harness are released.

arXiv:2611.03982 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. Agents fail at enterprise procedures because a pre-built static harness cannot encode proprietary policies, undocumented system behaviour, or state-dependent recovery paths that only appear during execution, and procedural memory accumulated from experience offers a training-free way to close that gap — but it must consolidate repeated successes and diagnosed failures into reusable actions, not just store episodes or facts. WASP-Mem (Wiki-based Agent-Synthesized Procedural Memory) distils execution traces and feedback into persistent interlinked wiki pages that the agent keeps extending. On TheAgentCompany it raises checkpoint pass rate by 4.6-6.1% over the strongest baseline across three models while using 57-61% less storage. A controlled study over consolidation method, feedback type, trace ingestion scope, and refinement rounds finds that imposing human-defined schemas and explicit deduplication rules hurts rather than helps, while one round of human diagnosis of failed traces beats four rounds of agent self-judgment by 7.0-10.9%, suggesting agents should organize the memory and humans should spend their effort diagnosing.

arXiv:2611.04003 PDF

Recursa: Toward Recursively Self-Improving Foundation Model Systems

Erkang Zhu, Yin Lin, Xiaorong Shi, Xuanrui Lin, Jingren Zhou Foundation models deployed in new domains often meet knowledge their training did not cover, and those domains rarely offer much expert-written training data or reliable supervision. Recursa adapts by learning from its own experience, with no human experts or stronger teacher model. It keeps a knowledge repository of domain skills, notes on when to apply them, memories of past tasks, and reflections on what to improve. The repository is first built from domain materials, then refined recursively: parallel learners carry out tasks and propose improvements, which are validated and merged back in. On Harvey LAB's Firm Knowledge benchmark, Recursa raises the all-pass rate on held-out tasks from 30.0% to 55.0% with DeepSeek-V4.1-Flash and from 35.0% to 46.0% with Qwen3.8-Flash.

arXiv:2611.04035 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, world-model look-ahead, and multimodal LLM judges, but no one had measured each component's contribution on its own. The authors build OSWorld-Tutorial and ScienceBoard-Tutorial, two datasets of curated multimodal tutorials, and propose RAG-PJ (RAG-PolicyJudge), where a policy proposes candidate actions and a judge picks among them. When tutorials are present, world-model rollouts add little benefit, even from an oracle that returns the true future screens. Error analysis points to the judge's ability to reason over future states and recognize task completion as the real bottleneck, not how accurate the rollouts are.

arXiv:2611.04073 PDF

The Agent Board: Kanban for Machines; Durable Coordination for Agents That Outlive the Conversation

Tyler Hunt, Donald Nguyen, Gurbinder Gill Most LLM harnesses assume one agent, one conversation history, and one session, which breaks down for work that spans many agents, outlives a single model call, or waits on outside systems. The Agent Board in the Corvic harness is a durable shared task table, modeled on a kanban board. Each task has a few exact fields the system routes on plus a free-form body that only models read, and agents can post tasks, watch them, or block on them without holding a running process. A system agent called the Board Keeper carries events from non-agent parts of the harness onto the board, matches them to waiting tasks by reading their prose, and starts new agents when needed. In small simulated pilots, waiting on the board cost the same regardless of wait length while naive polling cost grew about 165-fold, but a plain external watcher was cheaper still, and the Keeper missed 3.6% of real wakes while producing no false ones.

arXiv:2611.04100 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. Existing LLM agents for accelerator code tend to target a single programming interface or hardware family, work only at the kernel level, and reuse little experience across targets. ModelBlaster is a memory-augmented in-context reinforcement learning framework (MAIC-RL) whose policy is a persistent knowledge base. That knowledge base is updated at inference time by actor-critic textual gradient descent (AC-TGD) and records performance states, actions, outcomes, and constraints, so lessons carry over across kernels, models, and hardware. The same loop drives CUDA, OpenCL on Qualcomm GPUs, and RISC-V vector intrinsics, with only the lowering and measurement layers specific to each target. It reaches geometric-mean speedups over PyTorch of 1.85x, 5.58x, and 5.18x on KernelBench Levels 1-3 on a B200 GPU, 5.8x over the vendor SDK on Adreno GPUs, and 2.7x over ExecuTorch on RISC-V.

arXiv:2611.04104 PDF

Interpretable Early Termination of Web Navigation Agents via Closed Sequential Pattern Mining

Sergio Talavera, Magdalini Eirinaki Web navigation agents built on open-weight language models often fail late, spending most of their tokens after a trajectory can no longer be recovered, and existing early-termination methods give decisions that are hard to interpret. The authors run closed sequential pattern mining (BIDE+) on the first K steps of agent traces to build a small failure-signature library that humans can review, organized into a four-category failure taxonomy with repair recommendations. Their evaluation protocol removes contamination artifacts such as zero-step crashes and trivially separable one-step successes. On 1,544 MiniWoB++ traces from three models, the method gets the best F1 among eight methods at K=3 (0.717), recovers 88% of test failures within three steps, and supports an early-termination policy that saves 26.3% of tokens at 92% precision.

arXiv:2611.04142 PDF

Validation-Gated ML Pipeline Synthesis with Expert-Grounded Self-Evolving Skills

Zixiao Zhao, Jing Sun, Zhe Hou, Mengze Li, Qian Liu, Cheng-Hao Cai et al. Existing agentic AutoML systems either rely on rigid templates or on free-form code generation, where intermediate errors can propagate silently into the final model. EvoML splits ML pipeline construction into four phases run by three agents. An Orchestrator advances a phase only after its outputs pass structural, execution, and semantic checks; an Executor runs expert-grounded skills based on the nnU-Net workflow; and a Curator folds scripts and guidance recovered during each run back into the skill library. On eight datasets spanning six real-world task types, compared against agentic AutoML systems and command-line coding tools, it achieves full automation and the best final model quality on all eight, including 0.94 Dice for segmentation and 0.99 accuracy for classification.

arXiv:2611.04180 PDF

When the Consumer Is an Agent: Execution-Grounded Service-Level Agreements

Michal Orzechowski, Renata G. Slota, Jacek Kitowski cross-listed LLM agents that consume asynchronous distributed services often report a task as complete based on an accepted call or an immediate state read, which does not prove that the required state became observable by the deadline. The authors propose execution-grounded completion contracts: an independent control plane watches a deadline-bound predicate on service state, decides whether completion is accepted, and directs recovery while the agent works. In a live multi-provider data service with several models and an injected transfer delay, agent self-assessment both accepted tasks that missed the deadline and rejected some that met it. Under enforcement, accepted completions stayed consistent with later observations, and unresolved faulted tasks ended marked as unverified.

arXiv:2611.04182 PDF

Agentic Path Maps: Exposing an Agent's Traversal Through the Garden of Forking Paths

Madeleine Grunde-McLaughlin, Hafsah Usman, Daniel Mitroshkov, Emily Reif, Nino Migineishvili, Lakshya Chauhan. Kevin Yu et al. cross-listed When an AI agent carries out an underspecified task, it makes choices that may conflict with preferences the user never stated, and asking about every one is impractical. The authors lay out a design space for showing users the paths an agent did not take after it finishes. In a user study (n=48) on data analysis, they test one dimension of it, the number of alternatives displayed. Participants rarely considered alternatives on their own, but visualizing alternative paths reduced overreliance on the AI and helped them interpret results in light of their own context.

arXiv:2611.04183 PDF

Harness-S: Learning from Sibling Harnesses for Generalizable Agent Training

Kangning Zhang, Wenxiang Jiao, Bowen Fang, Shuai Shao, Jiachen Zhu, Rong Shan et al. Agent harnesses that evolve from a common base each fix different failures, and their usefulness shifts as the policy learns. Harness-S uses this to train one policy that works well under the original harness. It builds a fixed pool of sibling harnesses, each specialized along a different repair direction. When every rollout under the base harness fails, it collects fresh rollouts from selected siblings and uses those trajectories in reinforcement learning instead of the failed group. An adaptive router picks the siblings using behavior profiles and voting across several LLMs. Across AppWorld, SWE-Gym, and DBBench, the trained policy substantially outperforms policy-only training when returned to the base harness, and reusing the router at inference gives further gains at the same rollout count.

arXiv:2611.04277 PDF

Drift No More: Spectral Locking for Long-Horizon Open-Weight Multi-Agent AI

Vishal Pramanik, Maisha Maliha, Susmit Jha, Sumit Kumar Jha Long-horizon multi-agent systems suffer task drift, role drift, and stale shared state. A pilot study over 14,832 annotated turns suggests a common cause: hidden states drift away from the directions that encode the active task, the assigned role, and the current state. ATLAS (Algebraic Task-role Locking with Adaptive Spectral Memory) is a training-free intervention that builds a positive-semidefinite kernel over these three subspaces. It steers hidden states toward their shared high-coherence eigendirections before each token is generated, using a low-rank decomposition to stay tractable. It beats prompting, retrieval, memory, and activation-steering baselines, improving memory accuracy by up to 5.7 points over A-MEM, reaching 93.5% role consistency, and transferring to a 72B backbone.

arXiv:2611.04315 PDF

ATC: Agent Trajectory Control for Role-Contract Enforcement in Black-Box Multi-Agent Systems

Maisha Maliha, Vishal Pramanik, Sumit Kumar Jha In multi-agent LLM systems, an agent can drift out of its assigned role over many turns, for example a verifier that starts planning or an executor that changes state nobody approved. The authors call this role drift and propose Agent Trajectory Control (ATC), which writes each role's authority down as an explicit contract and checks every proposed message before it enters the shared conversation. ATC scores each message for out-of-role behavior, leakage into other roles' functions, impermissible speech acts, and logical contract breaches, then constrains regeneration and re-checks the revised message. Because it works on messages rather than model internals, it applies to black-box agents, and it beats prompting, memory, workflow-control, and activation-editing baselines on both task performance and role consistency, with the largest gains under role-confusion stress tests and in larger, more diverse teams.

arXiv:2611.04321 PDF

CloakFEM: Agentic FEM Code Generation for Red-Teaming Computational Resource Exhaustion

Li Zheng, Simin Chen cross-listed Finite element method (FEM) solvers can be targeted by efficiency attacks: valid, runnable programs that finish successfully but force excessive computation, enabling computational denial-of-service. CloakFEM is an LLM-agent framework that searches for such programs, combining LLM-guided search with an FEM-specific intermediate representation and a deterministic validity gate that keeps every candidate runnable. Across three FEM stacks, 17 problem classes, and 1,068 seeds, it finds inputs that greatly amplify solver work, in the most extreme case raising solver iterations from 43 to 247,367 (a 5,752.72x amplification). It outperforms five generic black-box baselines and keeps an edge over evolutionary search, and cross-solver analysis shows that some vulnerabilities are intrinsic to the problem while others are specific to a solver.

arXiv:2611.04356 PDF

Silent Instrument Errors Reorder Agent Leaderboards

Yangze Liu (Shandong University) The authors audit two public agent leaderboards, HAL and tau2-bench, using only their published artifacts, and find that evaluation infrastructure errors silently change scores and rankings. In one case a general-purpose agent appeared 13 points behind a specialized airline scaffold because only about half its episodes ever triggered grading and the rest were stored as failures; grading the ungraded half under the benchmark's own rule closes the gap without rerunning any agent. Other errors include a cost column that double-counts nested model calls, archived trajectories that record grader-written actions as agent actions, and re-grading under a newer evaluator that reorders a leaderboard or collapses some scores by 59 to 82 points with no error message. The HAL maintainers confirmed three of the errors, and the authors propose cheap checks: report the graded fraction, aggregate cost from root spans, tag the provenance of each action, and pin evaluator versions with a published re-grade diff.

arXiv:2611.04399 PDF

Verifier strength does not buy discovery quality: a pre-registered two-domain test of LLM-based automated research, with post-hoc evidence pointing to proposer-pool composition

Peter Chen, Ni Zhang, Siyuan Zhang Automated research pipelines pair an LLM that proposes candidates with a verifier that filters them, on the assumption that stricter verification yields better discoveries. In a pre-registered study across two real-data domains, job-shop scheduling and vehicular edge offloading on Beijing T-Drive trajectories, the authors compare 100 matched LLM and random proposal pairs per domain under three verifier strengths. Although the verifier tiers measurably differ, the LLM's advantage over random proposals does not grow with verifier strength and in one domain shrinks under the strictest tier. A post-hoc analysis finds that LLM proposals do outperform random ones and that mixing the two pools gives a significant gain, which points to proposer-pool composition, not verification strictness, as the more effective lever.

arXiv:2611.04420 PDF

$\tau$-Rec-Books: A Disclosure-Controlled Benchmark for Requirement Elicitation in Agentic Recommender Systems

Bharath Sivaram Narasimhan, Karthik R. Narasimhan cross-listed Conversational recommender agents have to discover what a user wants before recommending, and τ-Rec-Books tests this by keeping each request fixed while changing when the user reveals requirements: all up front, one up front and the rest when asked, or some never. The benchmark has 450 tasks over a 20,000-book catalog drawn from Wikidata, a rule-based simulated user, and scoring against catalog metadata with no LLM judge. Across four agents and two search interfaces, average success drops from 0.975 when users volunteer everything to 0.756 in the mixed setting and 0.283 when some requirements are hidden. Mixed tasks fail mostly because agents never ask, and every agent reaches at least 0.931 once told to ask; on hidden tasks, agents told to propose books first only match a popularity baseline and fall well short of a baseline that uses rejections to rule out possible requirements.

arXiv:2611.04459 PDF

From Watching to Steering: Enabling Human-Coding Agent Collaboration through Execution Traces

Xiaoke Yan, Jiaying Lei, Danqing Shi, Zichen Wang, Nan Cao cross-listed Coding agents can run long software tasks on their own, but users find their execution hard to inspect, understand, and steer. Vibetrace, built on OpenCode, displays agent workflows as interactive execution traces, lets users fork execution from an earlier step with a targeted correction instead of restarting, and distills traces and human feedback into reusable skills. On SWE-bench Verified, the distilled skills raise pass rate from 65.0% to 68.0% while cutting input tokens by 23.0% and cost by 15.5%. User studies show better accuracy and efficiency than conventional log-based interaction.

arXiv:2611.04466 PDF

CogniLife: Continuous Cognitive Distillation for Deployed Multi-Expert Agent Marketplaces

Xuwei Tan, Jingchun Huang, Ming Cai, Shuangquan Lyu, Tiange Wang, Yinzhi Wang et al. In multi-expert agent marketplaces, creators publish domain-specific AI experts, but a newly published expert has little evidence about its creator to draw on, which the authors call the cold-expert problem. CogniLife frames continuous cognitive distillation as turning attributed creator observations into revisable decision rules, communication patterns, and procedures, using Creator Profiles and Expert Journals, though this cross-expert design is not yet implemented or evaluated. A deployed platform implements part of the surrounding lifecycle, and on the 85-task ProfyBench-v1 its full configuration scores 4.65/5 from an LLM judge versus 4.01 for a raw LLM and 4.28 for an expert prompt. The authors note these scores do not test cold-expert mitigation, cross-expert transfer, or longitudinal distillation.

arXiv:2611.04483 PDF

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

Ran Li, Lei Chen Compiling a natural-language agent skill into model parameters removes the skill text from the prompt, and existing methods assume the compiled artifact carries the skill's content. Controls on a retrieval agent challenge that assumption: a wrong skill document performs as well as the correct one, one averaged bundle matches per-skill routing, and a random direction of the same norm falls back to baseline, which suggests the compiled shift encodes a domain-general behavioral prior rather than skill-specific content. The authors then introduce SKILL2STEER, a hypernetwork that reads a skill document once and produces low-rank activation shifts added after every decoder block, using 2.36M parameters per skill, 9.25× fewer than a weight-space adapter. On the seven-dataset SearchQA benchmark it reaches 37.9 exact match, 9.8 points above the frozen model and 2.3 points above the weight-space method.

arXiv:2611.04497 PDF

Signals: Continuous Diagnostics of AI Agents

Amey Rambatla (Cisco), Vatsal Goel (Cisco), Sagar Gaur (Cisco), Siddhant Loya (Cisco), Anurag Mishra (Cisco), Asif Mohammad (Cisco) AI agents can make wrong decisions or waste resources even when their tools and infrastructure work normally, and their nondeterminism makes such problems hard to reproduce. Signals diagnoses these problems from agent traces using a taxonomy of 45 problem types in eight families, separating realized failures, risk exposures, and inefficiencies. It groups related findings into engineering issues, rates their severity from the evidence, and attaches concrete remediations along with checks for verifying a fix, while later runs can add evidence to existing issues. On public agent executions it achieves 91.43% F1 at matching reference issues, about 18 points above a one-pass LLM baseline that used the same model and taxonomy.

arXiv:2611.04507 PDF

Learnings and Pitfalls from AI-Assisted A/B Experimentation at Scale

Brett Daley, Ilya Gorbachev, Anubhav Natani, Nicholas Brown, Athena Dai, Rachel Song et al. The authors report lessons from building and operating an LLM agent for A/B experimentation, where model nondeterminism and hallucination clash with the trustworthiness experiments must provide. Their central finding is that the tension is resolved by bounding where nondeterminism can reach: the model selects and sequences procedures, while deterministic code computes every reported quantity. Three results surprised them: withholding information from the model improved its answers, users mostly wanted mundane retrieval rather than autonomous decisions, and the agent's eagerness to help put statistical validity at risk.

arXiv:2611.04515 PDF

TactiNavi: Serving and Revising an LLM Game Coach with Explicit Strategy Conditions

Qi Liu, Xiaoyang Yuan, Zhuomeng Zhang, Yubin Ruan, Xinyi Mou, Wenjin Wang et al. A deployed LLM coaching agent for a commercial open-world game has to give advice whose correctness depends on the player's state and the game version. TactiNavi handles this by compiling versioned game evidence and expert guidance into strategy packages with explicit prerequisites. At serving time it evaluates these prerequisites as three-valued conditions over structured player state and logs decision traces, and a control plane uses those traces to validate revisions before release. On 1,122 logged coaching turns, condition-bearing packages raised correctness from 57.9% to 82.0% over atomic facts and reduced strategy misapplication from 18.1% to 10.4%, at a median reply latency of 2.07 s. The deployment also showed that most apparent fixes were generation noise: rerunning the unchanged system recovered 38.4% of failing turns, and only one of 15 shipped revision units passed the release checks retrospectively.

arXiv:2611.04772 PDF

Kierston & KierstonBench: Simulation of Allentown, PA and Evaluation of Language Models in Crime Investigation and Surveillance Tasks

Andrew Semenov, Svyatoslav Dorofeev No existing benchmark can measure how accurately language models perform criminal-investigation and surveillance work, because real deployments lack ground truth. Kierston simulates about 125,000 residents of Allentown, Pennsylvania, and generates the telecom, financial, administrative, and license-plate records that would be held by separate providers. KierstonBench poses 32 task families, including record linkage, routine reconstruction, incident investigation, and detecting deceptive accounts, which models must solve through a restricted records interface. Among sixteen models evaluated, most could not carry out records investigation and submitted identifications the records did not support. A few frontier models solved most tasks and abstained when evidence was insufficient, but they still sometimes fell for deceptions and could not replace an investigator.

arXiv:2611.04785 PDF

Where Is the Truth When AI Agents Disagree? Dependence-Aware Belief Propagation for Unsupervised Truth Discovery (Extended Version)

Laure Berti-Equille LLM agents often disagree, and agents built on the same base model make correlated errors. As a result, classical truth-discovery and crowdsourcing estimators reward agreeing blocs and can do worse than majority vote. BP-TD adds per-agent decorrelation weights to mean-field variational inference on a one-coin Dawid-Skene model, so that agents sharing a base model count together as roughly one vote. With 13 local agents on MMLU, it beats majority vote without labels (0.716 vs. 0.624 on 160 questions) and ranks agents in line with their true accuracy. Across 2,852 sub-pools, however, decorrelation helps only when the largest correlated bloc is weaker than the other agents, and with just 20 labeled questions, picking the single best agent beats every aggregator tested.

arXiv:2611.04805 PDF

AgenticFinBench: Benchmarking Financial Agents on Diverse Tasks from Real-World Requests

Shu Zhou, Ziyang He, Yuxin Rao, Yudong Fang, Jiali Qian, Dongyuan Chen et al. AgenticFinBench contains 230 real-world financial requests spanning 8 categories, 2 languages, and 4 complexity levels. Many of the tasks are open-ended with no single correct answer. Evaluation is reference-free: question-specific rubrics are combined with a two-stage claim audit that flags hallucinations, defined as claims absent from the agent's trace and refuted by external evidence. Across 18 agent configurations, hallucination auditing drops the leading baseline by 7.58 points and to fourth place. Differences in agent harness explain more of the variation in evidence grounding than differences in model (28.6% vs. 8.8%), and quality falls from 89.44 on direct retrieval to 79.18 on the most complex tasks.

arXiv:2611.04813 PDF

One Assistant, Many Memories: Benchmarking Real-World Group Assistance When Recall Is Not Enough

Zhanghao Hu, Linhai Zhang, Qian Zhao, Qi Zhu, Yuan Hua, Di Liang et al. GroupAssistBench is a memory benchmark built from real group chats with a deployed assistant, covering 80,299 messages across 50 groups and costing over $53K to collect. Over the same chat histories, it tests both memory recall against reference answers and memory use through open-ended service requests, scored with rubrics that include penalties derived from real user feedback, across single, distributed, and conflicting memories. Across frontier models with full context, retrieval methods, and dedicated memory systems, the strongest memory system rivals full context on recall but falls far behind on use, especially when memories are distributed or conflicting. Two additions, relating retrieved memories to one another and binding remembered facts as constraints on the response, bring consistent gains.

arXiv:2611.04867 PDF

ScamWorld: An Interactive Environment for Evaluating Web Agents Against Online Fraud

Yutao Wu, Xiao Liu, Ming Wen, Hanxun Huang, Jun Wang, Xingjun Ma Web agents that shop and manage accounts for users will run into fraud aimed at humans, such as fake stores, investment scams, and advance-fee schemes, but scam sites are short-lived and hard to study systematically. ScamWorld rebuilds reported scam websites across many scenarios, fraud mechanisms, and languages, and makes the warning signs observable before the agent's final action so that evaluations are repeatable. Across web agents powered by six frontier LLMs, baseline deception rates range from 11% to 35%, and agents go ahead with payments more readily than with requests for credentials or wallet details. Trajectory analysis shows that some agents notice suspicious conditions but proceed anyway, sometimes treating the user's instructions as permission to accept extra terms imposed by the site.

arXiv:2611.04879 PDF

VisSearch-Bench: Evaluating Long Horizon Visual Search in Multimodal Agents

ZhengXian Wu, Qianhao Nie, Zhuohong Chen, Junjie Gao, Yuyao Zhang, Yunyao Yu et al. cross-listed Many search benchmarks for multimodal agents use the image only to identify a starting entity and then rely on text search, which reveals little about how agents connect visual evidence across many steps. VisSearch-Bench contains 204 human-written tasks across 7 domains and 25 subdomains that require combining object appearance, fine details, or spatial relations with information found through search, in formats such as set-wise, sequential, and competitive disambiguation search. Each task comes with a reference answer, traceable evidence, and scoring rules, and is scored on both full-task accuracy and a per-step process score. Across 16 multimodal models, the best system fully solves only 26.47% of tasks, even with search tools.

arXiv:2611.04894 PDF

Artifact-to-Capability: Diagnosing Breaks in Capability Formation of Self-Evolving Agents in a Controlled Environment

Ziqian Ban, Haoxiang Yang, Yanfei Zhang, Shixin Liang, Haoran Yu, Xu Lin et al. Self-evolving agents store past experience as artifacts such as reflections, skills, or code and reuse them on later tasks. Their progress is usually judged by three proxy signals: artifacts accumulate, behavior changes when they are loaded, and metrics rise. The authors define capability formation as five sequentially dependent properties and diagnose them at the state, game, and lineage levels in freestyle Gomoku. Reflexion, Voyager, and DGM all show the three proxy signals with the same base model, yet each fails in its own way. Reflexion's pass rate rises only because more of its moves are legal, Voyager's individually sound components combine to miss all 20 available one-move wins, and later DGM versions only lose passing states, so signal-based evaluation alone cannot show that a capability has formed.

arXiv:2611.04952 PDF

GeoFaham: Domain-Specialized Multi-Agent Geospatial Analytics for Disaster Response

Muhammad Sohail Danish, Gilles Q. Hacheme, Akram Zaytar, Satej S. Soman, Inbal Becker-Reshef, Juan Lavista Ferres et al. cross-listed GeoFaham is a multi-agent system for disaster-response geospatial analysis. Specialist agents handle vector operations, map features, satellite imagery discovery, and raster analysis, coordinated by an orchestrator that keeps two ledgers for planning, progress tracking, validation, and replanning. Large geospatial outputs are stored outside the LLM context window in a Response Store, and agents pass compact identifiers and summaries instead. On the new 40-task GeoFaham-Bench, built from 17 real disaster records, it scores 0.940 versus 0.879 for the next-best general coding agent while using 7.5 times fewer tokens; the baselines include Codex, Claude Code, and OpenCode.

arXiv:2611.04977 PDF

ToolSpace: Scaling Tool-Centric Environment Synthesis for Long-Horizon Agents

Guanting Dong, Zhichao Sun, Xinyu Yang, Rui Wang, Xin Deng, Jiajie Jin et al. Training LLM agents to chain tools across MCP (Model Context Protocol) services needs realistic environments, but existing synthesis pipelines generate a database first and derive tools and tasks from it, so the results drift from real interfaces. ToolSpace works in the other direction: it mines data, state and retry dependencies from execution traces of real MCP services using deterministic rules, clusters the resulting tool-composition graph into environments, and generates tasks through state-aware random walks that are checked by replaying them from the initial state. Harder versions of AutomationBench and tau3-Bench Banking built this way drop the scores of seven frontier models by up to 38.2 points. Training on the synthesized data raises Qwen3.6-35B-A3B from 23.0 to 41.0 average score across seven long-horizon tool-use benchmarks and lifts GLM-5.2 to a level on par with Claude Opus 5. Ablations show that dependency edges observed in real executions clearly beat edges judged by an LLM.

arXiv:2611.05036 PDF

From AutoResearch to Continual AutoResearch: Do Agents Learn Across Tasks?

Jeremy Qin, Ruizhi Deng, Mohamed Osama Ahmed, Frederick Tung, Leo Feng Autoresearch agents can write code and run ML experiments, but evaluations usually start each new problem in a fresh session, so it is unclear whether agents carry insights from one task to the next. Continual AutoResearch compares persistent and reset instances of the same agent on matched task streams across nine task families, including symbolic regression, time-series forecasting and tabular AutoML, with a memory ablation that separates insight transfer from session continuity. Across five frontier agents, persistent state improves research efficiency for all of them, but not consistently their final solutions. The authors conclude that retaining experience and using it effectively are distinct capabilities.

arXiv:2611.05108 PDF

Meta$^2$: Precomputing Reusable Knowledge for Online Harness Evolution in Long-Horizon Agents

Lingzhi Yuan, Yiguang Cai, Weize Liu, Siqi Yang, Feijiang Han, Yongkyun Lee et al. Harness optimization improves long-horizon agents by refining the prompts, tools, memory, and hooks around a frozen model. Offline methods deploy one shared harness that may overfit, and online methods must reason about every change from scratch. Meta² instead learns, offline, a textual Meta-Skill that encodes when and how to modify a harness. It does this by contrasting training rollouts without an evolver, with a Meta-Skill-guided evolver, and with an evolver given privileged verifier information. At test time, an evolver uses the frozen Meta-Skill and evidence from the current run to either keep the harness or make a task-local edit, without verifier feedback. With GPT-5.6-Terra, it achieves the best results among evaluated methods: 81.00% on Terminal-Bench 2.1 and 44.44% on DeepSWE 1.1.

arXiv:2611.05139 PDF

GM League Arena: An Interactive Long-Horizon Task Benchmark Based on Professional Sports League Management

Junkai Zhou, Ziming Pan, Zhaoyi Zhang Benchmarks of multi-step LLM interaction rarely capture how decisions add up over long horizons under resource constraints and sustained competition. GM Bench has LLMs act as general managers of professional sports teams over multiple seasons, using simulators for four sports with financial constraints, player demands, and market mechanisms. Rosters of comparable strength are built from real-world player data, and models are scored on competitive performance, cash reserves, and improvement across seasons. Over five NBA seasons, Claude Opus 5.5 ranks first among 29 LLMs, while GPT-6 Astra and GPT-6.1 Sol combine strong results with larger cash reserves.

arXiv:2611.05147 PDF

OperAItor: An Agentic Framework for Fault-Tolerant Operation of Chemical Processes

Javal Vyas, Mehmet Mercang{\"o}z cross-listed OperAItor is an open-source framework for building and evaluating LLM agents that handle supervisory fault management in chemical plants. It is demonstrated on a simulated reactor-distillation process with 113 variables, control loops, alarms and safety interlocks. The agent combines statistical process monitoring, fault diagnosis, knowledge-graph retrieval and action planning. The study compares five ways of validating actions: no validation, an LLM judge, digital-twin simulation, and deterministic admissibility checks with or without consequence projection. Across 20 fault scenarios, unassisted operation led to 18 plant stops, while validated agent configurations running GPT-5 nano, Qwen3 8B and Mistral 7B cut this to 5–8 stops and recovered $265,000 of a $399,000 economic loss.

arXiv:2611.05179 PDF

JHiMA-WebRL: A Joint Hierarchical Multi-Agent Reinforcement Learning Framework for Long-Horizon Web Tasks

Shanka Subhra Mondal, Jiayun Wang Realistic web tasks need long-horizon planning, progress tracking across long histories, and visually grounded actions, which monolithic multimodal policies handle poorly, and prior modular systems usually leave the planner frozen. JHiMA-WebRL pairs a task decomposer with a subtask-conditioned actor: a strong teacher model builds structured subtask supervision from multimodal graphical user interface (GUI) trajectories to fine-tune both, after which Group Relative Policy Optimization (GRPO) is extended with hierarchical rollout grouping and role-specific advantages to optimize them jointly in WebArena. With a 7B actor and 7B decomposer, JHiMA-WebRL-14B reaches 37.2% average success on WebArena-Lite-v2 against 24% for an RL-tuned ScaleCUA-32B under the same protocol, and ablations show joint decomposer updates beat updating the actor alone.

arXiv:2611.05263 PDF

MDSKILL: Learning to Internalize Skill Memory through Multi-View Distillation for Efficient Dynamic Agent Self-Evolution

Dian Jin, Jinluan Yang, Kairong Han, Baohong Li, Zijing Hu, Yi He et al. Textual skill memories can guide self-evolving agents, and On-Policy Self-Distillation (OPSD) can internalize that guidance as dense supervision, but self-generated memories are noisy and often redundant or contradictory, so direct distillation can hurt under limited training budgets. MDSKILL organizes skill memories into three complementary functional views (state grounding, action planning, and execution control), selects high-quality skills with dynamic metrics to build view-specific and integrated reasoning branches, then uses OPSD to consolidate knowledge from all branches into one shared policy. Compared against reinforcement learning, skill-augmented, and self-distillation baselines that use no stronger external teacher, it improves on prior state of the art by up to 7.1% on ALFWorld and 8.1% on WebShop without additional compute overhead, with ablations and evolution analyses supporting the internalization claim.

arXiv:2611.05313 PDF

Adversarial Speculative Reasoning Improves Long Horizon Agents

Hongyin Luo, Wei Fang, Lijie Yang, Owen Stepan, Daniel Suit, Jeff Krapf et al. Speculative decoding is usually treated as a pure speed-up, but the authors find that numerical noise inside verification windows can shift a model's output distribution after about 30 steps, randomly changing long-horizon agent accuracy. They train the draft model to be deliberately rejected at critical reasoning decision points found in agent traces. The target model then samples those tokens itself under its least noisy conditions. On complex agentic coding benchmarks, the method raises end-to-end success with no change to the target model or its output distribution, training only 0.06% to 0.28% of hosted parameters.

arXiv:2611.05345 PDF

3 more specialized papers

Theory 51

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

Marcus Thomas This theoretical paper argues that a model's gains in predictive accuracy can require lasting errors in how it responds to interventions. For deterministic predictors under squared-error loss, any observational advantage over the best predictor using only an outcome's direct causes requires dependence on other inputs, whatever the learning rule, and when interventions change those inputs without changing the outcome, this gives a positive lower bound on intervention-response error. The authors identify conditions under which gradient updates introduce such dependence or damage a previously correct mechanism through shared parameters. They argue that correction is limited by the causal hypotheses currently under consideration, and conjecture that reliably maintaining interventional competence requires the system to take part in proposing, testing and revising causal alternatives itself.

arXiv:2611.03861 PDF

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 usually reports the depth at which a signal first becomes decodable, and the common rule — a probe score beating a null-calibrated threshold for three consecutive layers — turns out to be badly miscalibrated because neighbouring layers are correlated, a hazard long known in neuroimaging. On synthetic curves with known ground truth that rule fires nineteen stages early and never fails to report an onset; a corrected estimator builds null trajectories by permuting labels within each evaluation unit and holding the permutation fixed across depth, uses per-layer order statistics as thresholds, and calibrates the persistence requirement rather than picking it. Applied to pocket information in a frozen co-folding trunk over 1,820 post-cutoff targets, the three-stage rule fires on 12-17% of null trajectories against a nominal 5%, onset fails to distinguish any of four familiarity regimes, and a pre-declared accumulation-depth measure does. Separately, the authors report that a widely used public co-folding checkpoint loads against its own code with 128 trunk parameters missing and silently randomly initialized.

arXiv:2611.03968 PDF

The Evolution Strategies Paradox: Better Models from Worse Perturbations

Roberto Dailey, Kajetan Schweighofer, Jamieson Warner, Conor F. Hayes, Elliot Meyerson, Babak Hodjat et al. When Evolution Strategies (ES) are used to fine-tune billion-parameter LLMs, every perturbed offspring model can score worse than its parent while the parent still improves steadily during training. The authors document this paradox across models, tasks, and modalities. They explain it through the local reward landscape and properties of the ES gradient estimator: in very high-dimensional spaces, individual perturbations hurt performance but still point toward good directions, and averaging those directions yields real improvement. They present this as an optimization regime in which noise interacting with curvature produces the paradox, and they draw practical guidance for perturbation-based LLM fine-tuning.

arXiv:2611.04113 PDF

Proper Agnostic Learning of Affine Halfspaces under Log-Concave Marginals

Zhijie Zheng The authors give a proper agnostic learner for affine halfspaces under full-dimensional log-concave distributions: for arbitrary binary labels, it returns a single affine halfspace with error at most a constant times the optimum plus epsilon, using polynomially many samples and arithmetic operations, and it does not need to know the distribution's mean or covariance. The key geometric result relates global disagreement between halfspaces to disagreement on a balanced restriction of the space, which cancels the noise amplification that conditioning would otherwise cause. To build that restriction from noisy labels, the algorithm robustly estimates the minority cap's mean and covariance and approximately minimizes a zonoid-depth dual objective, with an analysis based on degree-four sum-of-squares moment bounds.

arXiv:2611.04344 PDF

Theoretical Analysis of Muon on Transformers for Next-Token Prediction

Shengwei Wu, Yixuan Yang, Gao Huang, Song Li The Muon optimizer works well for training Transformer LLMs, but its dynamics are poorly understood theoretically. The authors analyze Muon on next-token prediction in two phases: a preprocessing phase that learns the output-value component, for which they characterize the trajectory exactly and show the normalized spectral margin is optimal after each update, and a main phase that learns the key-query component. For the main phase they establish non-asymptotic convergence rates for training loss and the normalized margin gap under softmax attention, suggesting Muon helps models learn efficiently which tokens matter for prediction. The learned weight structure also yields a generalization guarantee for unseen contexts, and experiments on next-token prediction tasks support the theory.

arXiv:2611.04465 PDF

Curriculum unlearning in humans and neural networks

Jirko Rubruck, Alice Zhang, Andrew Saxe, Christopher Summerfield Humans often learn better when taught simple rules before exceptions, and the authors propose a computational explanation using a canonical behavioural task alongside neural network models. Both humans and networks show a simplicity bias, settling on underfit simple rules, and in networks this comes with compressed internal representations. Simple-to-complex curricula help both kinds of learners escape this compression and reach representations whose dimensionality fits the task. A confirmed prediction is that learners pre-exposed to simple tasks, with more compressed initial representations, benefit more from curricula, suggesting curricula work by helping learners "unlearn" simple rules.

arXiv:2611.04565 PDF

Modelling lock-in of AI adoption through skill erosion

Renaud Lambiotte The authors build a mathematical model in which people imitate successful peers, their own skills grow with practice, and those skills erode when they rely on generative AI, and they derive the resulting skill distributions exactly. Skill erosion can create two stable states, low adoption with high skill and high adoption with low skill, and the population can stay stuck in one depending on its history (hysteresis). A single AI user spreads through a population of experts only if AI-assisted work pays more than expert work, while a single non-user can lead an AI-dependent population back only by regaining enough skill first. As a result, a temporary boost in AI's benefit, such as a period of low prices, can leave a population relying on AI even after the benefit returns to its original level.

arXiv:2611.04886 PDF

PAC Learning with Multiple Correct Answers

Steve Hanneke, Farnam Mansouri, Alireza F. Pour, Amirreza Shaeiri, Kun Wang The paper studies PAC learning when each example has a set of at most ℓ correct labels. Training feedback comes in one of three forms: the full set of correct labels, bandit feedback saying whether a proposed label is correct, 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. For demonstrations, a class is learnable only if intersections of up to ℓ+1 of its hypotheses are learnable, and the sample complexity scales as Õ(ε^-ℓ), an exponent that cannot be improved in general. For full sets and bandit feedback, a generalized Natarajan dimension characterizes learnability, and the authors also show that without the uniform bound ℓ, no dimension based on finite subclasses can characterize learnability.

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 prone to spurious local optima. For a canonical model with linear experts and a gate defined by Gaussian clusters, the authors give what they describe as the first global convergence guarantee for jointly training the gate and experts with gradient expectation-maximization (EM). The guarantee requires logarithmic overparameterization and well-separated gate centers. Experiments confirm the predicted m log m phase transition and suggest the required least-squares warm start is unnecessary in practice. They also find that unconstrained mixing weights concentrate routing onto the true experts, while load-balancing penalties spread it out, so the unconstrained model can be pruned heavily without losing accuracy.

arXiv:2611.05184 PDF

Logarithmic Regret for Alternating Mirror Descent

Tianlong Nan, Christian Kroer Alternating online mirror descent (AltOMD), in which players in a zero-sum game take turns updating, was known to achieve O(T^{1/3}) and O(T^{1/5}) regret, but whether it reaches near-constant regret was open. The authors resolve this question: the players' summed external regret is O(log T) in any two-player zero-sum bilinear game with compact convex strategy sets, under mild assumptions on the regularizers. The assumptions hold for negative entropy, giving logarithmic regret for alternating multiplicative weights updates. The bound also implies 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 guarantees coverage when calibration and test points are exchangeable, but in performative settings the deployed predictor changes the data distribution and exchangeability fails. The authors ask when prediction sets can stay valid at performative equilibrium and prove a phase transition at L=1 in the total-variation sensitivity of the distribution map: for L≥1 no finite-sample procedure can guarantee both equilibrium-valid coverage and nontrivial sets, while for L<1 the minimax excess miscoverage scales as 1/((1-L)√n), inflating sample complexity by 1/(1-L)². The main tool is a transfer theorem showing any procedure landing within ε of equilibrium gives equilibrium-valid coverage up to an additive Lε after split-conformal calibration, instantiated with a model-free and a tail-averaged parametric retraining algorithm. Experiments including semi-synthetic credit scoring reproduce the predicted threshold and scaling.

arXiv:2611.05277 PDF

Universal Hardness From Masking Channels

Parikshit Gopalan, Manoj Prabhakaran, Pranay Tankala The classic hardcore lemma on average-case complexity of Boolean functions is generalized by reinterpreting a hardcore measure, which reweights the input distribution, as an input-dependent masking channel that injects randomness into the output until it looks pseudorandom, with channel fidelity playing the role of measure density. Choosing the right notion of fidelity via convex duality tightly characterizes the difficulty of learning a function with a hypothesis class under a convex loss as the maximum fidelity achievable by a masking channel, extending beyond Boolean functions and 0-1 loss. The framework yields multiclass generalizations of the hardcore lemma, exposes subtleties at larger alphabets, and constructs a universal masking channel covering many hardness notions at once, dual to omnipredictors. The authors also relate masking channels to pseudoentropy and give a tight characterization of pseudoentropy via the MaxEnt algorithm.

arXiv:2611.05296 PDF

Agnostic Learning by Imitating

S\'ilvia Casacuberta, Parikshit Gopalan, Varun Kanade, Omer Reingold, Konstantinos Stavropoulos, Pranay Tankala The question is whether training a discriminative model that does as well as the best in a class (agnostic learning) is computationally harder than producing a generative model that no member of the class can tell apart from real data (multiaccuracy). Earlier work gave an efficient reduction from multiaccuracy to agnostic learning using the regularity lemma. The authors supply the reverse reduction, proving the two tasks equivalent in complexity, and show that a multiaccuracy oracle also achieves a range of harder learning objectives. The construction depends on adaptive oracle queries and links the problem to follow-the-leader style no-regret online learning.

arXiv:2611.05346 PDF

38 more specialized papers

Safety & Alignment 31

Alignment Faking Depends on the Value, Not Just the Model

Manodeep Ray, Vivek Srivastava, Shirish Karande, Murari Mandal Alignment faking is when a model follows a directive while it believes it is being monitored and drops it when it believes it is not. The authors argue that this behavior depends on the specific pairing of model and value, not on the model alone. They score ten open- and closed-weight models (8B to about 675B parameters) on 32 ValueBench dimensions under no-system-prompt, unmonitored, and monitored conditions. Compliance gaps range from 0.05 to 0.50, a distilled 70B model shows a larger gap than a ~675B one, and most of the variance comes from the model–value interaction rather than from the model or the value alone. A taxonomy of reasoning traces shows that models reason about training signals when monitored and about value conflicts when unmonitored, and results are packaged as per-model "value cards".

arXiv:2611.03650 PDF

LoRA-Patch: Defending Against Backdoored LoRA Adapters via Directional Steering

Shengyan Huo, Peihai Jiang, Xixiang Lyu, Yige Li cross-listed Third-party LoRA adapters can carry hidden backdoors that activate on a trigger while normal task performance stays intact. Existing defenses edit, prune or distill model parameters, which does not fit deployments where the trusted base model must stay frozen. LoRA-Patch builds on the observation that backdoor injection distorts the directional geometry of an adapter relative to its base model, and suppresses triggered behavior by realigning adapter directions with layer-wise steering. Across five backdoor attacks and three models, it outperforms existing defenses while preserving benign task performance, and builds each patch in under 20 seconds.

arXiv:2611.03752 PDF

The Web Doesn't Sit Still: Adversarial Self-Evolving Attacks on Search Agents

Geyuan Wang, Shunyu Liu, Siyuan Liang, Chenyang Lyu, Dacheng Tao cross-listed LLM-based search agents depend on web content, which makes them vulnerable to injected or manipulated information. Existing attacks mostly rely on static, hand-written templates. The authors propose an adversarial self-evolution framework that uses bi-level optimization: attack strategies evolve while a defender adapts its mitigations. The attack side is driven by Contrastive Rollout Evolutionary Optimization (CREO), and experiments run in a controlled sandbox with fine-grained vulnerability metrics. Across multi-hop QA benchmarks and frontier LLMs, the evolved attacks are more effective than static baselines.

arXiv:2611.03754 PDF

JevBreak: Your Model, Their Decision

Nitin Aravind Birur, Divyanshu Kumar, Sahil Agarwal, Prashanth Harshangi cross-listed Limiting a model's output to a fixed set of labels does not stop an attacker from steering which label it picks. Slot Attack is a black-box attack that appends text built from named components to the input evidence, pushing decision models toward an attacker-chosen wrong option using only the returned label probabilities. Across seven models and 16 tasks from Decision Bench, it succeeds on 32–99% of initially correct decisions, needing a median of 17 to 72 queries. Models with nearly identical accuracy differ by about 20 percentage points in vulnerability, and confidence thresholds do not reliably filter out the attack.

arXiv:2611.03759 PDF

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 image or text does not support a reliable conclusion. Dr. No finds sparse autoencoder (SAE) features associated with abstaining or committing to an answer, and adjusts a small set of them during the prefill pass without retraining. Evaluated on LLaVA-1.6 Mistral 7B with the UNK-VQA and MM-UPD benchmarks, and on Qwen3.5-2B with AbstentionBench subsets, feature steering balances correct answers against appropriate abstention better than post-hoc rejection baselines, especially on UNK-VQA, on mismatched image-question or question-answer pairs, and on math reasoning. Gains are smaller for detecting absent answers and for retrieval-heavy multi-hop QA such as MuSiQue, suggesting that different sources of uncertainty call for different abstention controls.

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 poisoned examples often remain statistically detectable after unlearning, which suggests current methods do not fully remove what was memorized. PINT uses the Hilbert-Schmidt Independence Criterion (HSIC) to reduce the statistical dependence between the tokens to forget and the internal representations that predict them, and the authors connect this objective to exact memorization through a computable upper bound. Evaluated on Gaussian poisons, personal information, hazardous biology and cybersecurity knowledge, and copyrighted text with models up to 7B parameters, and scaled to 70B, PINT achieves the best average forgetting-utility trade-off and is the only method that makes Gaussian poisons undetectable while preserving near-original utility.

arXiv:2611.03955 PDF

Whose startup does the AI back? An algorithm audit of reputation and demographic signals in LLM-assisted crowdfunding

Mirza Samad Ahmed Baig, Syeda Anshrah Gillani As language model assistants start advising people on which crowdfunding campaigns to back, it is unclear whether they reproduce, erase, or invert the race and gender discrimination documented in entrepreneurial finance. A pre-specified choice-based conjoint experiment had seven models pick among five synthetic campaigns with independently randomized quality signals and founder names, under three backer personas plus a persona-free control, yielding 40,248 responses. By default founder names barely register — under budget-conscious and delivery-focused personas, gender and Black-name effects are statistically equivalent to zero — but when the backer says they want to support underrepresented creators, the models favor Black-signaled founders by 8.1 percentage points (positive in all seven models) and female-signaled founders by 5.1, most strongly among near-tied campaigns, and almost never cite gender or ethnicity in their stated reasons. The authors read this as an agency account of algorithmic discrimination: the assistant takes demographic preferences from its user rather than its training data.

arXiv:2611.03993 PDF

Characterizing Adversarial GEO via Implicit Brand Injection

Shobhnik Kriplani, Haohan Wang cross-listed Generative search engines answer from retrieved web pages, which means whoever controls that content can influence which brands surface in the answers; this work characterizes that exposure by studying what happens when a single contextually integrated brand mention is added to an article, so the text stays non-promotional yet relevant enough to be retrieved. Implicit Brand Injection is a multi-agent pipeline that picks an insertion point, rewrites the sentence, verifies the edit, and selects a final candidate, evaluated on IBIBench — 1,000 web articles and 60 synthetic brands over 12 domains, covering both brands native to a domain and unrelated ones — against direct prompting and best-of-N ranking at equal budget. The edits preserve semantics better and disturb perplexity less while improving retrieval Hit@k at every cutoff, query regime, and placement setting, reaching 82.2% Hit@10 and 0.74 MRR@10, with brand mention rates up to 62.7% in generated answers. The authors frame the result as evidence that one sentence can move both retrieval exposure and answer-level brand visibility, arguing for defenses against this kind of content manipulation.

arXiv:2611.04005 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 remove the watermark, which typically requires access to the detector or hard-to-collect paired audio. Commercial platforms are increasingly closing off both. The authors show that removal is unnecessary: detection depends on the watermark pattern staying structurally consistent, and perturbing the harmonic structure of speech breaks that consistency while keeping the audio natural. ScrambleMark is a zero-query, watermark-agnostic, black-box attack that pairs a deterministic engine for finding transferable harmonic perturbations with a generative model that learns randomized scrambling patterns. Against Google SynthID, Resemble AI, and TrustedAudio, it achieves nearly 97% evasion with a 4.2/5 human naturalness rating, compared with 1.1-1.7 for prior attacks, and it holds up against adaptive attempts to recover the watermark.

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, applied to Alpamayo for driving and InstructVLA for manipulation, finds that the most action-sensitive direction is nearly orthogonal to the text-loss gradients. Across 154 driving scenes, TNAS-guided attacks cause a new collision under an unchanged rationale in 52.6% of scenes, compared with 4.55% for unguided attacks. The authors conclude that an unchanged rationale cannot certify a safe action, so monitors must check the action channel directly.

arXiv:2611.04166 PDF

Neologism Inoculation Prompting

David Demitri Africa, Cameron Holmes Inoculation prompting reduces unwanted behaviors such as reward hacking by training on harmful examples under a prompt that names the behavior. Because that prompt is plain English, any user can re-trigger the behavior. Neologism inoculation prompting (NIP) replaces the English prompt with a new token trained on the frozen model and then removed at deployment. NIP outperforms other inoculation methods and is competitive with gradient routing while being simpler; it works on reward hacking but fails on sycophancy and harassment. Control-token and activation analyses indicate that NIP hides the harmful behavior behind an unreachable region of input space rather than deleting it.

arXiv:2611.04260 PDF

Reliability Stress Testing of RAG Retrieval under Query-Relevant Corpus Poisoning

Janet Sung, Kainan Zhou, Gangzhen Qian, Hang Xiao, Lu Yi The study stress-tests the retrieval step of retrieval-augmented generation (RAG), the point at which poisoned documents get admitted into the model's context. It varies poison content, how much the attacker knows about the query, the poison budget, and the contamination scope. On 623 queries from SciFact and NFCorpus in the BEIR benchmark, five query-targeted documents per query cut classical-retriever nDCG@10 by 40.2–61.5%, and the dense retriever BGE-small reaches 100% target exposure. Most of the damage comes from copying the query into the poison document (query anchoring) rather than from the injected payload. Overlap-based filtering can remove these poisons but sharply hurts clean retrieval, and maximal marginal relevance (MMR) leaves exposure unchanged.

arXiv:2611.04264 PDF

When Debate Amplifies Risk: Social-Bias Vulnerabilities in Multi-Agent Safety Deliberation

Qinghua Mao, Xi Lin, Yuxin Qi, Zihao Xu, Jinze Gu, Yuliang Chen cross-listed Multi-agent debate (MAD) is increasingly used to reconcile conflicting retrieved web content, but deliberation is a social process that may let pressure override evidence. The SafeConflict-Bias protocol holds queries and evidence fixed across six web-risk categories while varying majority sway, authority cues, and confidence asymmetry. Across proprietary and open models, isolated models reliably reject unsafe guidance, but in standard debate unsafe adoption jumps from 0.0% to as high as 100.0% and safe minority evidence gets silenced. The proposed inference-time auditor RiskTrace-Debias combines evidence-first precommitment, masking of social metadata, and consistency-gated arbitration, and cuts worst-case unsafe adoption to at most 2.0% without causing over-refusal.

arXiv:2611.04289 PDF

Sparse Internal Control of Language Models

Xizhe Zhang The paper introduces sparse internal control, a framework for steering a frozen language model by applying inference-time interventions at a small set of internal nodes. Candidate site-direction pairs are chosen under constraints on reaching the target behavior, preserving off-target behavior, and limiting intervention energy, using two selectors: Driver-OMP, which sparsely reconstructs a desired behavioral shift, and Coverage-Driver, which favors stable coverage across prompts. The author also proposes a six-axis evaluation protocol covering reachability, signed reversal, dose response, feedback controllability, off-target preservation, and held-out reuse. Exploratory experiments on indirect object identification, MMLU, sentiment steering, and refusal steering suggest sparse interventions are feasible, with control properties that vary by task.

arXiv:2611.04349 PDF

Voluntary Servitude: Weaponized Safety Warning Mechanisms for Controlling AI Agents

Ilan Zendel, Matan Levy, Stav Cohen, Avi Mendelson, Amit LeVi cross-listed Defenses against attacks on LLM agents mostly look for injected commands in retrieved content, but the authors show that an attacker can steer an agent through its own harm-avoidance instead: a single fabricated hazard, written as an ordinary review, leads the agent to rule out legitimate options and choose the attacker's target by its own reasoning. They call this voluntary servitude. These payloads succeed about as often as instruction injection (60.4% vs. 66.0%) but deployed injection detectors score them as benign, making them 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 about 95% of the exposure against a non-adaptive attacker, without retraining.

arXiv:2611.04393 PDF

Correct Labels Are Not Enough: Reliable Metadata Can Silently Control Model Behavior

Renee Jia, Jinzhou Wu When fine-tuning data includes metadata that predicts the right answer, such as a reward-model score, models can learn to depend on that metadata even if every training label is correct. The authors fine-tune 7-9B open-weight models from three families on multiple-choice questions where a (Reward: 5) annotation always marks the correct option. When the annotation is moved to a wrong option at test time, models pick that option on 97-100% of CommonsenseQA items, but removing it leaves most answers correct, so standard evaluation misses the dependence. Cue-following grows with how reliable the annotation was during training, appears even with neutral or nonsense annotation words, and also shows up in Qwen3 after direct preference optimization (DPO); the authors propose audits that move or remove visible annotations while keeping the task fixed.

arXiv:2611.04395 PDF

Sample-Grained Approximate Unlearning with Provable Per-Sample Bounds

Yaxin Xiao, Yulin Jin, Li Tang, Haoyang LI, Huadi Zheng, Qingqing Ye et al. Approximate machine unlearning methods remove the influence of a forget set without full retraining, but they are usually judged with dataset-level metrics that can hide large differences from retraining on individual samples. SALA (Sample-grained Likelihoods ALignment) is a plug-in optimization framework that pushes each example's output toward its correct target distribution. It comes with a new sample-grained distribution gap metric, theoretical guarantees of a bounded per-sample discrepancy from retraining, and a low-cost way to estimate member and non-member target distributions. Paired with existing unlearning baselines on classification and generation tasks, it reduces the sample-grained distribution gap by up to 86.67% and makes results less sensitive to hyperparameters.

arXiv:2611.04501 PDF

Inter-Agent Influence: Evaluating Persuasion, Deception, and Coercion in Multi-Agent Systems

Chandler Smith, Qi Guo, Cecilia Elena Tilli, Sophia Hatz, David Demitri Africa, Patricia Paskov et al. The authors study whether AI agents built on frontier language models can strategically steer other agents through persuasion, deception, and coercion, testing this in five realistic multi-agent evaluation environments. In oversight environments, every tested model shifted an overseer agent's decision from rejection to approval using at least one of these tactics. Influence also worked in non-hierarchical peer-to-peer settings, where an agent redirected a peer's safety research trajectory. The authors conclude that mitigations for inter-agent influence are needed as multi-agent deployments spread.

arXiv:2611.04556 PDF

Reducing Logprob Without Causal Verification: Diagnosing SAE-Based PII Suppression in LLMs

Chaerin Lee, Wonho Lee, Sohee Park, Daeseon Choi Interventions on sparse autoencoder (SAE) features that aim to stop LLMs from leaking personally identifiable information (PII) are usually judged by how much they reduce log-probability. The authors argue that this criterion cannot tell feature-specific suppression apart from a generic effect of large activation perturbations. Across 36 combinations of model, layer, and PII type, they compare selected SAE features against random directions matched for coherence, using both logprob and rank-aware metrics. Logprob alone makes 17 cases (47%) look successful, but only 1 case (2.8%) passes the multi-metric test; in the other cases random directions often do better, or feature selection collapses across PII types. A 9,000-run causal mediation study supports the conclusion, and the authors call for PII suppression grounded in circuit-level causal analysis.

arXiv:2611.04782 PDF

TraceRouter++: Causal Semantic Routing Realignment via Optimal Transport

Chuancheng Shi, Shangze Li, Wenhua Wu, Hengzhi Chen, Fei Shen, Li Zhuo et al. Many safety interventions for large foundation models suppress isolated neurons or features, but harmful behavior runs through distributed circuits that span layers, which makes such edits brittle against jailbreaks and harmful to general utility. TraceRouter++ finds the layer where harmful inputs start to diverge using attention divergence, splits activations into interpretable features with sparse autoencoders (SAEs), and traces those features to downstream causal circuits using feature influence scores. Instead of hard suppression, it uses constrained optimal transport to shift unsafe activation distributions toward safe reference distributions, only within the causally identified subspaces. Across LLMs, multimodal LLMs, and diffusion models, it outperforms the evaluated baselines on adversarial robustness while keeping a favorable safety-utility trade-off.

arXiv:2611.04878 PDF

Whose Effect Is It? How Objectives Inflate Learned Subspace Interventions

Toygar Tanyel, Serkan Ayvaz Interpretability researchers often locate concepts in a language model by copying part of its internal state from one input into another (activation patching). When the copied part was learned, it frequently changes the output more than copying the whole state does, and that larger effect is usually read as stronger evidence that the part isolates the concept. The authors prove that, to first order, the effects of orthogonal parts add up to the effect of the whole state, so a part can exceed the whole only when the rest of the state pushes the output back. They show that distributed alignment search trained with cross-entropy exceeds the whole-state copy on almost all tasks, and that retargeting the objective to what the network itself produces removes most of the excess. They conclude that the excess comes largely from the training objective, so a larger learned effect is not by itself stronger evidence.

arXiv:2611.05033 PDF

LatentTriad: Pre-Action Risk Representations for Malicious Agent Skill Detection

Zhe Yu, Yi Liu, Ying Zhang, Simin Chen LLM agents increasingly load reusable skills, and untrusted skill content can make an agent perform operations outside the user's task. Static pattern matching misses this kind of contextual risk, while running the skill to observe it is costly. The authors find that hidden states formed during prefill already encode whether the operations match the task. LatentTriad builds a detector on these states that generates no tokens, calls no tools and executes no skill code. Across 5,830 skill packages from five sources and seven backbone models, it reaches 99.0 to 99.2% pooled F1, with a median latency of about 259 to 284 ms for a compact implementation.

arXiv:2611.05088 PDF

From Opinions to Commitments: Unstable Assignments and Silent Choices in AI-Mediated Proposals

Sachit Mahajan, Javier Argota S{\'a}nchez-Vaquerizo, Arthur Capozzi, Dirk Helbing LLMs are increasingly used to turn citizens' opinions into collective proposals that assign who should act, what must be done, and with what budget and deadline. The authors ran four LLM mediators five times each on 30 opinions from DeliberationBank and Polis, varying only the random seed or the order of the opinions. In 17–20% of five-run sets, some runs contained assignments while others did not, and reordering the opinions changed assignments almost as often as replacing them with different citizens' opinions. This variation persisted at temperature zero, with a commercial model, and with draft selection or merging. Prompts seeking agreement also covered minority positions much less often than prompts asking to describe differences, so the authors propose audits that expose alternative assignments and disclose each proposal's objective.

arXiv:2611.05135 PDF

The Geometry of Ignorance: Detecting Suppressed Knowledge in LLMs through Activation Fragmentation

Adar Peleg, Dvir Alsheich, Tomer Ashuach, Avi Mendelson Behavioral tests cannot tell whether a large language model (LLM) truly lacks knowledge or is merely suppressing it, since a refusal may still sit on top of recoverable information. The proposed framework measures the geometric coherence of anon-vectors, the hidden-state differences between a concept-specific query and its anonymized paraphrase, on the premise that genuine knowledge forms a tight low-entropy cluster while ignorance forces fragmentation. Across ten instruction-tuned LLMs from three families, all fragmentation metrics separate known from unknown concepts, and after unlearning the geometry correlates with adversarial recovery across four attack types and five unlearning algorithms, though the sign of that relationship depends on the attack. A layer-resolved analysis shows different attack mechanisms exploit the geometry at different depths, giving a single-phase, attack-free proxy for unlearning robustness.

arXiv:2611.05272 PDF

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

Ariana M. Villegas-Suarez, Guido Freire, Sohan Venkatesh, Bartosz Cywi\'nski Fine-tuning a large language model (LLM) on narrowly harmful data can cause broad misalignment, but models are rarely trained on explicitly harmful data in practice. This work shows that fine-tuning on content the model believes is false, with nothing harmful in it, also induces broad misalignment, and that the effect scales with how strongly the model disbelieves the training answers: in mathematics, misaligned answers rise from 4% to over 60% as training answers become more wrong, with little loss of general capability. After fine-tuning the model asserts the wrong answers while its multiple-choice probabilities still favor the true ones, pointing to internal inconsistency as the mechanism, and the same pattern appears in realistic single-fact updates as in knowledge editing or continual learning. Framing a false fact inside a coherent story makes the model believe it more and become less misaligned than instructing it to assert the fact, yet content filters and persona vectors cannot distinguish the two framings.

arXiv:2611.05287 PDF

Weight-space programming for auditing clinical decision networks

Humza Nusrat, Omar Nusrat cross-listed Testing a clinical model on cases with known answers may not reveal a shortcut that the tests share with the training data. The authors check whether a frontier language model can audit such a network by reading its raw weights, an ability called weight-space programming. Small networks (433 to 4,801 parameters) that screen head-and-neck radiotherapy requests were given planted flaws, and Claude Opus 5.5 audited each one blind, with no data, tools or code execution. It named the broken rule in 16 of 18 flawed networks and repaired each by changing a median of 3 parameters, but it cleared only 2 of 5 healthy networks and found 1 of 3 hidden triggers, so it cannot yet certify that a network is correct.

arXiv:2611.05341 PDF

Suppressing Misaligned Behaviors in Language Models, Durably

Luxi He, Pengcheng Jiang, Jifan Zhang, Jiawei Han, Danqi Chen The question is how durably a misaligned behavior can be suppressed after it appears. The authors build two model organisms, one that cheats on multiple-choice questions only when a hint is present and one that gives broadly harmful answers after fine-tuning on insecure code. They compare SFT, GRPO and gradient-ascent unlearning on immediate suppression, durability under further training, and preserved capability. Their method, PEARL (Prefix-Expanded Adversarial Reinforcement Learning), extends GRPO rollouts with continuations from intermediate misaligned states, and across Qwen3-4B and GPT-OSS-20B it suppresses the behavior while keeping it suppressed under further fine-tuning, with the largest benefit for broad misbehavior.

arXiv:2611.05358 PDF

4 more specialized papers

Robotics 28

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

Zhanyi Sun, Shuran Song cross-listed Visuomotor policies trained by behavior cloning suffer from covariate shift: small deviations from expert trajectories compound into failure. Latent Policy Barrier (LPB), inspired by Control Barrier Functions, treats the latent embeddings of expert demonstrations as an implicit boundary between in-distribution and out-of-distribution states. A base diffusion policy is trained only on expert data, while a separate dynamics model, trained on both expert and suboptimal rollouts, predicts future latent states at inference time and steers them back toward the expert distribution. In simulated and real-world manipulation experiments, LPB improves robustness and data efficiency without any additional human corrections or annotations.

arXiv:2508.05941 PDF

VagueBench: Benchmarking the Robustness of Embodied Agents on Ambient Context Illusions

Yang Qu, Qiang Fu, Zonghao Ying, Jiayi Zhao, Yaoyuan Zhang, Jinyang Guo et al. Embodied agents are rarely tested on short instructions that have the full meaning but leave out the extra context people often omit. The authors call the resulting failures Ambient Context Illusion: agents fail when non-essential details are removed even though the task itself is unchanged. VagueBench, built on the AI2-THOR simulator, contains 1,303 instruction pairs created with a nine-subtype taxonomy for removing context and checked to preserve meaning. Across six state-of-the-art embodied agents, the reduced instructions cause a relative 19.65% drop in task success and a 54.10% relative increase in empty plans, and the effect also appears on real robots.

arXiv:2611.03765 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 two-armed manipulation system trained entirely on 16,151 hours of demonstrations recorded with YUBI, an open-source handheld gripper, with no robot data, teleoperation, reinforcement learning or on-robot corrections. Using only one camera on each wrist, it does long, precise tasks such as sorting small parts into numbered cells and replacing a robot vacuum's mop pads and brush, at a site where none of its training data was recorded. It learns which regions of each wrist view matter, predicts the current task stage, and decodes actions as a continuous stream. Across ten tasks it averages 78% success versus 27% for the strongest of four other vision-language-action models post-trained on the same demonstrations, and it runs on several robot arms that mount the gripper; the dataset, weights and post-training recipe are released.

arXiv:2611.03815 PDF

RoboScope Kit: Unified Interfaces for Fast and Scalable Embodied-AI Evaluation

Deng-Ke Wu, Yifan Jin, Yida Yang, Lei Qian, Zhaoye Zhou, Hongru Zhang et al. cross-listed Robot simulation benchmarks each come with their own evaluator, observation layout and action conventions, and evaluation is slow: profiling the official StarVLA evaluator shows simulation keeping CPU cores busy while the GPU averages 14.3% utilization. RoboScope Kit wraps each model once as a Policy class and each benchmark once in an adapter around its own evaluator, leaving scoring unchanged, and covers six benchmark families. It runs simulator workers and Policy replicas as separate pools sized from measured probes, shards work deterministically across GPUs, and renders only the observations a policy reads. On LIBERO-Long with one RTX 4090 it evaluates 500 episodes 4.45x faster than the official StarVLA evaluator (6.16x with CUDA MPS), and on RoboTwin Easy four GPUs finish 500 episodes in 14.4 minutes, 25x faster than the single-GPU baseline, with identical outcomes.

arXiv:2611.03850 PDF

Controlling Action-Head Co-Adaptation to VLM Representation Anisotropy

Ruijie Zhang, Jiayin Zheng, Yequan Zhao, Liyan Tan, Ziyue Liu, Zhengyang Wang et al. Modular systems that bolt a trained head onto a frozen pretrained encoder are everywhere, but how the head adapts to the geometry of those pretrained features is little studied; the setting examined here is vision-language-action models (VLAs), where a pretrained vision-language model conditions an action model for robot control. The action model turns out to progressively concentrate its sensitivity on the highest-variance dimensions of the vision-language model's output, and controlled experiments across backbone-head pairs show that counteracting this concentration helps when representation and task are mismatched. Adaptive Interface Equalizer (AdaIE) is the resulting parameter-free mechanism, reweighting pretrained representations according to the head's evolving sensitivity during training; it adds up to 5.28 percentage points on RoboCasa-GR1 and 1.68 on RoboTwin 2.0, with an average 19-point gain across five real-world manipulation tasks, and also transfers to diffusion policies and world models.

arXiv:2611.04006 PDF

Embodied AI for Scientific Discovery: From Physical Action to Scientific Evidence

Wenhao Xu, Changwei Wang, Shunpeng Chen, Yukun Song, Sifan Zhou, Kaize Luo et al. This review brings robot learning and AI-driven experiment selection together under one problem, embodied AI for science, in which the state of the specimen, the apparatus, and the measurement process is only partially observed and actions are costly and irreversible. It organizes the literature by three kinds of closure: operational (the experiment run is the one intended), epistemic (the result updates beliefs or picks the next experiment), and evidential (the record supports a claim). Methods and workflows in chemistry, materials, biology, microscopy, and field science are compared through a Record-Diagnose-Calibrate-Confirm evidence cycle, and an appendix maps 28 workflows. The authors propose evaluations that separate individual actions, valid experiments, campaigns, and confirmed claims.

arXiv:2611.04371 PDF

Human-to-Robot Transfer via Bridging Representation and Action Space in VLAs

Junhyeong Park, Wook Jung, Taeyoung Kim, Huiwon Jang, Jinwoo Shin Adapting a pretrained Vision-Language-Action (VLA) policy to a new robot normally requires costly robot demonstrations, while cheaper human demonstrations suffer from a large embodiment gap. H2RB (Human-To-Robot Bridge) co-trains on unpaired human and robot demonstrations, using geometry-anchored alignment to place both kinds of observation in a compatible representation and a decoupled-decoder tokenizer to share action latents without forcing both through a single action mapping. On real-world tasks it raises object-transfer scores from 8.0 to 51.8, achieves 62.7 on skill transfer where co-training baselines fail entirely, and improves out-of-distribution placement transfer from 6.0 to 45.3.

arXiv:2611.04391 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 keep scenes spatially consistent across viewpoints and over long interactions. 3D World Simulator instead predicts appearance, geometry, and tactile signals from a shared 3D representation: a hybrid particle-voxel autoencoder compresses point clouds into voxel latents, and an action-conditioned dynamics model predicts how those latents evolve. The model delivers interactive, long-horizon multimodal prediction at 50 Hz, can render geometrically consistent views from arbitrary cameras, and generates data for training and evaluating policies. Scaled to large, diverse datasets, it also supports model-based planning with unseen objects and robot configurations.

arXiv:2611.04492 PDF

Cycle WAM: Coupling Action and Video through Cycle Consistency

Qinglin Yang, Jiepeng Wang, Yuxing Huang, Haibin Huang, Kai Xu, Chi Zhang et al. cross-listed World action models (WAMs) use pretrained video models for robot control, but their training does not ensure that predicted actions and their predicted visual consequences agree. Cycle WAM treats video and action as paired modalities: it co-trains conversion in both directions to learn forward and inverse dynamics, then adds a differentiable cycle-reconstruction objective that ties the two mappings together. It reaches state-of-the-art results among comparable methods on LIBERO-Pro and substantially improves DreamZero on DROID.

arXiv:2611.04711 PDF

LLMs as MDP Designers: A Dependency-Aware Agentic Framework for Automated Robot Policy Generation

Hong Je-Gal, Jinwoo Kim, Chanbin Yi, Hyun-Suk Lee cross-listed Instead of using large language models (LLMs) to make decisions at run time, ARPG uses them as designers that turn a natural-language task description into a complete robot reinforcement learning (RL) pipeline: a Markov decision process (MDP) formulation plus its simulator implementation. The framework treats MDP modeling as a sequential, dependency-ordered process with localized verification and correction, then builds the simulator in modules once the formulation is fixed, saving intermediate artifacts along the way. This targets a common failure in which generated code runs but trains no usable policy because the state, action, reward, and termination definitions don't fit together. In Isaac Sim, ARPG improves modeling correctness, environment executability, and policy-generation success over single-pass LLM baselines and off-the-shelf agentic systems.

arXiv:2611.04974 PDF

Do Agentic Policies Need to Look at Every Step? Geometric Abstraction for Reasoning and Action

Yihang Liu, Yiming Chen, Mingzhe Zheng cross-listed Multimodal language models can control robots, but they typically re-read raw images at every step of a task. The authors move routine spatial reasoning into text through a 3D geometric abstraction: typed symbolic references to object parts, grounded from calibrated RGB-D data, with compact textual checkpoints and metric residuals returned after each action. The symbolic task structure can then be reused across decisions, and direct visual reasoning is invoked only on demand. In both offline policy synthesis and continuous online execution, this approach improves completion of multi-step manipulation tasks while substantially reducing repeated multimodal queries.

arXiv:2611.05058 PDF

FloMo: 3D Scene Flow as a World Action Model Intermediate for Learning from Human Video

Jeremy Collins, Namra Patel, Ayush Agarwal, Shitij Govil, Animesh Garg cross-listed Pretrained video generation models carry useful motion priors for robot policies, but full video is a redundant signal for predicting actions. FloMo fine-tunes a 5B-parameter video generation model to predict 3D scene flow, rendered as RGB into the same latent space as video, together with action tokens. It is co-trained on action-free human video and action-labeled robot teleoperation data, and it generalizes to unseen objects, motion primitives and tasks. Swapping motion prediction for video prediction hurts out-of-distribution generalization, and adding a video target alongside motion lowers policy success, with action tokens attending far more to motion than to video.

arXiv:2611.05062 PDF

Asynchronous Robot Policies

Sombit Dey, Luc Van Gool, Danda Pani Paudel cross-listed Large Vision-Language-Action (VLA) robot models run semantic reasoning and action generation in sequence inside the control loop, so their latency hurts reactivity. Asynchronous Robot Policy (ARP) separates the two. A large vision-language model updates a semantic memory at low frequency, while a lightweight vision-action expert generates actions at high frequency, conditioned on that memory. The action expert is trained to tolerate stale memory and execution delays. On Franka manipulation tasks, ARP achieves an 18% higher success rate and 20% faster execution in quasi-static setups and 3–5 times baseline performance in dynamic simulation, and the vision-language model can run at 10 times lower than its nominal frequency without significant performance loss.

arXiv:2611.05155 PDF

RoboTryOn: Synthesizing Dexterous Robot Demonstrations from Human Videos via Virtual Try-On

Seungyong Lee, Suhyeok Jang, Minyoung Kim, Jinwoo Shin cross-listed Adapting robot foundation models to new tasks still depends on costly teleoperated demonstrations, and existing ways of turning human videos into robot data leave large visual or action gaps for high-degree-of-freedom robots. RoboTryOn treats the conversion like virtual try-on: it keeps the demonstrated motion, objects, and background and swaps the human for a target robot specified by a reference image. Keyframes are translated autoregressively and guide video synthesis conditioned on the hand skeleton. An inverse dynamics model that uses depth features then infers action labels from the synthesized videos. Across several humanoid embodiments, the synthetic data raises success on novel tasks from 2% to 60% without any real robot demonstrations of those tasks, and matches teleoperation-trained policies at about a third of the collection cost.

arXiv:2611.05163 PDF

GSBR: Geometric Support Behavior Regularization for Online RL

Charles Xu, Jacob Berg, Liyiming Ke, Sergey Levine cross-listed Regularizing online reinforcement learning (RL) toward a pretrained base policy aids stability but existing regularizers are too restrictive in one of two ways: behavioral cloning pulls the actor toward the whole base distribution including suboptimal modes, while hard support constraints forbid any improvement just outside the support. Geometric Support Behavior Regularization (GSBR) instead penalizes actions by their distance from the base policy's support, so in-support actions are free and out-of-support ones are penalized smoothly, permitting improvement when the value gained outweighs the penalty; the distance is estimated from a finite set of base-policy samples, requiring only sample access. Across 19 tasks on four robotic benchmarks, GSBR raises average success from 33% for the base policy to 95%, with the highest average success and throughput among compared methods and about 23% higher normalized throughput than the next-best RL method.

arXiv:2611.05295 PDF

PyRUA-DataFarm: LLM Agent as a Cost-Efficient Teleoperator for Real-Ready Robot Data

Kewei Zhang, Juncheng Ma, Guodong Lv, Duomin Wang, Feixiang Ruan, Ruiyang Si et al. cross-listed Collecting robot demonstrations by teleoperation costs robot and operator time for every episode. PyRUA-DataFarm hands that job to a coding agent instead. From a data request and one calibration episode of the real scene, the agent rebuilds the task in simulation by writing the scene, an expert controller and a success judge, and a video model conditioned on real images renders each rollout as consistent multi-view photographic video labelled with exact actions. On five real-robot tasks, policies trained only on this synthetic data matched teleoperation-trained policies (93 vs. 84 successes in 100 trials), and under an unseen tablecloth and coloured light they kept 87 successes where teleoperation-trained policies fell to 46. One hour of three-view data costs about $107, slightly below quoted teleoperation prices.

arXiv:2611.05320 PDF

GUILD: Guided Iterative Data Collection for Imitation Learning

Shivin Dass, Kevin Nguyen, Rohan Yelandur, Junhong Xu, Roberto Mart\'in-Mart\'in cross-listed Teleoperated data for imitation learning is usually collected on the operator's intuition, and existing curation methods only filter data after it has been gathered. GUILD (Guided Iterative Data Collection) closes the loop: vision-language models propose environments, object setups and goals worth collecting, then predictive data attribution estimates how each collected trajectory affects policy performance in the target deployment and steers the next round. In simulation and on real robots, guided collection improves sample efficiency and deployment performance over passive collection, multiple baselines and experienced roboticists.

arXiv:2611.05368 PDF

DreamX: Video Diffusion Model is a Universal Latent Action Learner

Jisoo Kim, Chaehyun Kim, Junsu Kim, Jiho Park, Seongchan Kim, Eunbeen Hong et al. cross-listed Latent action models that infer actions from video often tangle the action with irrelevant appearance, which limits transfer across embodiments. DreamX repurposes a pretrained video diffusion model to extract compact latent actions from its spatiotemporal features, compressing each latent frame 300× from 1,200 to 4 tokens, and uses the model's generative prior to separate dynamics from appearance. It reaches state-of-the-art results among latent action models on LARY-Bench and supports any-to-any video transfer and zero-shot action transfer between human and robot embodiments. The compact latent space also reduces the cost of policy learning built on video diffusion.

arXiv:2611.05377 PDF

10 more specialized papers

Unclassified 20

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

Muhammad Usama, Muhammad Saad Qadir No summary available — see the abstract on arXiv.

arXiv:2611.04192 PDF

DoGNAVY-Exploit: Harnessing Parallel Open-Weight Agents for Stage-Aware Vulnerability Exploitation

Yuxuan Wang, Dadi Guo, Weijie Guo, Zikai Xu, Chen Lu, Yizhe Zhuang et al. No summary available — see the abstract on arXiv.

arXiv:2611.04196 PDF

An Information Budget for Label-Noise Memorization, and What Statistical Complexity Tracks: A Study with Multilayer Perceptrons

Nancy Baygorrea C., Ot\'avio da Fonseca Martins Gomes No summary available — see the abstract on arXiv.

arXiv:2611.04199 PDF

Impact of Input Resolution in Mammographic Mass Detection: A Controlled Multi-Architecture Study

Kaan Buyukdemirci, Emine Ulku Saritas cross-listed No summary available — see the abstract on arXiv.

arXiv:2611.04200 PDF

Toward User-Centered Agentic Recommendation: When Does Platform Competition Benefit Users?

Kehan Zheng, Deyao Hong, Qian Li, Jun Zhang, Jie Jiang, Hongning Wang No summary available — see the abstract on arXiv.

arXiv:2611.04203 PDF

Point interpolation and distributional universality in flow-based generative models

Antonio \'Alvarez-L\'opez, Jingpu Cheng, Liyan Zhang, Qianxiao Li cross-listed No summary available — see the abstract on arXiv.

arXiv:2611.04207 PDF

Two-Layer Random ReLU Networks Can Separate Disjoint Classes

Yichen Liu, Jianbin Yang cross-listed No summary available — see the abstract on arXiv.

arXiv:2611.04211 PDF

Tokenizer Myths and Where To Find Them

Rylan Schaeffer, Sanmi Koyejo No summary available — see the abstract on arXiv.

arXiv:2611.04215 PDF

LENS: Reliability-Aware Stock Prediction via LLM-Extracted Priors and News-Driven Semantic Graphs

Hao Zhou, Tiejun Ma, Luis Felipe Costa Sperb No summary available — see the abstract on arXiv.

arXiv:2611.04217 PDF

Shared Context Candidate Scoring with Jointly Adapted Readout Networks

Jingyang Du, Yiben Fu, Yang Shen No summary available — see the abstract on arXiv.

arXiv:2611.04218 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 No summary available — see the abstract on arXiv.

arXiv:2611.04222 PDF

IntentGEO: Intent-aware Generative Engine Optimization for Recommendation

Runlin Lei, Shiqi Zhang, Xiaokui Xiao, Zhewei Wei cross-listed No summary available — see the abstract on arXiv.

arXiv:2611.04225 PDF

DiMuRe: Diagnosis-Guided Multilevel Recovery for Robotic Task Execution

Renkai Yi, Yuwei Wu, Haoyu Zhang, Penghao Dai, Zhi Gao, Jin Chen et al. cross-listed No summary available — see the abstract on arXiv.

arXiv:2611.04226 PDF

DualWorldBench: Can Agents Plan Deliveries across Symbolic and Grounded Worlds?

Yang Chen, Bin Wen, Danyang Peng, Hong-Jie You, Zi-Qiao Shang, Lan-Zhe Guo No summary available — see the abstract on arXiv.

arXiv:2611.04228 PDF

Exact Optimality Certification of QPLIB 2096 through Complete Set Partitioning

Jingyang Du, Zixiong Yi, Yang Shen No summary available — see the abstract on arXiv.

arXiv:2611.04230 PDF

Constructing a Regional Slang Lexicon for Electricity Demands via Soft-Label Bidirectional Anchoring and Active Learning

Rong Chang, Lijing Shi, Zhaoxiang Zang No summary available — see the abstract on arXiv.

arXiv:2611.04231 PDF

LIRA: A More Interpretable Dynamic Loss-Informed Rank Allocation for Low-Rank Adaptation

Sihao Wang, Bozheng Li, Xuemiao Xu, Haoxin Yang, Yujie Xu, Jie Tang No summary available — see the abstract on arXiv.

arXiv:2611.04239 PDF

AceAction-5M: Hierarchical Goals for Generalist Action Learning Across 3D Game Worlds

Kailing Li, Ruoxuan Zhang, Zheyuan Zhou, Boyang Chu, Yingshan Lei, Hongxia Xie et al. No summary available — see the abstract on arXiv.

arXiv:2611.04241 PDF

Navigating Runtime Dynamism in Semantic Query Serving

Hengrui Zhang, Xupeng Miao cross-listed No summary available — see the abstract on arXiv.

arXiv:2611.04248 PDF

COMPREHEND: Ultra-Fine-Grained Benchmark Understanding by Entailment Hierarchy

Sivaraaman Balakrishnan, Prakriti Shetty, Jeremiah Hong, Hamed Zamani, Haw-Shiuan Chang No summary available — see the abstract on arXiv.

arXiv:2611.04249 PDF

Vision 18

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 The benchmark tests whether point cloud architectures with built-in rotation invariance are still needed, comparing 10 models including PointNet++, DGCNN, RIConv++, PointTransformer and the equivariant VN-DGCNN under pose, noise and occlusion across 5 domains, with confirmatory hypothesis tests rather than single point estimates. The rotation-invariant (RI) architectures are the most robust to pose changes but among the least robust to noise in 4 of 5 domains. After retraining every model with a simple embedding invariance loss and evaluating on ModelNet40-C, the best standard backbone and the best RI architecture finish within 2.5 percentage points of each other on unseen corruptions, and with wider rotation augmentation, standard backbones lose only a flat accuracy offset under full-sphere rotation.

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 In autoregressive long-video generation, each new segment becomes context for later predictions, yet most pipelines choose which past frames the decoder sees implicitly, through recent windows, compressed caches, or external memory. ReCR is a training-free method that makes this an explicit decision at test time. It uses the decoder's internal representations to pick past context units that match the overall rollout state, avoid redundancy with recent frames, cover temporal boundaries, and include a range of spatial layouts, then places them on a compact causal axis before decoding the next block. Across long-horizon text-to-video benchmarks, it consistently improves several autoregressive backbones under matched visible-token budgets.

arXiv:2611.04164 PDF

Exploiting Biases in Attention for Transformers

Charles Herr, Gabriele Spadaro, Enzo Tartaglione In vision transformers, some tokens matter to attention far more than others regardless of the query, and this imbalance grows stronger in deeper layers. The authors rewrite the attention formula to add an input-dependent bias to the query-key dot product. This term gives each token a single number that predicts how much it will influence other tokens, and that number can be used to prune attention maps. Across several model sizes, the pruning substantially cuts attention FLOPs and inference latency while keeping most of the original classification accuracy.

arXiv:2611.04657 PDF

Learned Smartphone ISP on Mobile GPUs: Mobile AI 2026 Challenge Report

Andrey Ignatov, Georgy Perevozchikov, Radu Timofte, Yuhong He, Yuxuan Qiu, Tianyu Hao et al. cross-listed This report covers a Mobile AI 2026 challenge to build learned image signal processing (ISP) models that convert RAW Bayer sensor data into RGB photos fast enough to run on smartphone GPUs. Participants trained on the Fujifilm UltraISP dataset, which pairs Sony IMX586 Quad Bayer captures with a 102MP FujiFilm GFX100 camera. Runtime was measured on Snapdragon 8 Elite Gen 5 (Adreno) and Dimensity 9500 (Mali) chipsets. Top solutions reconstruct a high-quality 5MP image in under 100 milliseconds on both platforms, and the report describes the submitted models.

arXiv:2611.05008 PDF

RGB Photo Enhancement with Deep Learning on Smartphone GPUs: Mobile AI 2026 Challenge Report

Andrey Ignatov, Georgy Perevozchikov, Radu Timofte, Dong Zhang, Zhao Ran, Yang Zhang et al. cross-listed This report covers a Mobile AI 2026 challenge on deep learning models that fully enhance RGB photos while running on smartphone GPUs. Participants trained on the DPED dataset, which pairs photos from an iPhone 3GS with photos from a Canon 70D DSLR. Runtime was measured on Snapdragon 8 Elite Gen 5 (Adreno) and Dimensity 9500 (Mali) chipsets. Top solutions enhance Full HD images in under 300 milliseconds on both platforms, and the report describes the submitted models.

arXiv:2611.05024 PDF

13 more specialized papers

Multimodal 15

ReVi: Recursive Recap Video Generation

Kyrie You Wu, Chaozhuo Li, Songyang Liu, Xueyang Feng, Ziluowen Luo, Chenxu Wang et al. cross-listed The authors define a new task, generating a short recap video directly from a novel: a self-contained audiovisual retelling, grounded in the source text, that fits a fixed viewing-time budget. Their framework, ReVi, uses multiple agents in two loops and follows an "edit first, generate second" approach. The inner loop plans which story content to keep, how much narration time each event gets, and which shots go with it before any video is generated. The outer loop runs across novels and recursively revises both the editing policy and the meta-policy that controls those revisions. On two evaluation sets, ReVi improves recap quality and adherence to the time budget over representative baselines, supported by human ratings and results across multiple improvement rounds.

arXiv:2611.03674 PDF

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, presenting piano performances both as audio and as symbolic transcripts. It comes from a new dataset of more than 20,000 hours of instructional piano videos, with music transcripts interleaved with spoken commentary. Open-weight audio-language models reach at most 52.8% and barely beat a control with no audio, while frontier models using chain of thought reach 93.2% when given the music as text, showing that the bottleneck is perceiving music from audio rather than reasoning about it. Mid-training on the interleaved instructional data raises Qwen 3.5 9B Base's zero-shot accuracy with transcripts from 42.8% to 56.0%.

arXiv:2611.03839 PDF

ScienceVision: Can AI Observe and Reason Like a Scientist?

Tianyi Ma, Jing Gu, Haozhe Zhao, Liang Chen, Yunfan Zhang, Haijun Su et al. cross-listed Most multimodal benchmarks test text and 2D images, and existing scientific 3D evaluations use different formats and protocols in each discipline, which makes them hard to compare. ScienceVision is a visual question answering (VQA) benchmark of 250 expert-reviewed questions on native 3D artifacts, with 50 each from biology, chemistry, medical imaging, mechanical CAD, and architecture. Twelve models were tested in three settings: an agent harness with file and code-execution tools, the same harness plus viewer and measurement tools exposed through the Model Context Protocol (MCP), and rendered 2D images. The best mean score is only 0.53 out of 1. Working directly with the native 3D artifacts beat working from rendered images for every model, by 17.2 percentage points on average.

arXiv:2611.04087 PDF

Do Multi-Turn Conversational Capabilities Carry Over from Text to Speech? A Parallel Multilingual Benchmark

Jiaming Zhou, Pei Zhang, Shiwan Zhao, Zhijun Wang, Baosong Yang, Yong Qin Multi-turn voice assistants are usually evaluated on transcripts, which leaves open whether skills such as remembering earlier turns and following persistent instructions survive when the input is actually spoken, and whether that holds across languages. Omnilingua-AudioMC is a strictly parallel multilingual benchmark of 354 conversations and 1,404 rubric items in ten languages, aligned across transcript and speech conditions. Across six speech-language models, the gap between transcript and speech performance depends strongly on the model and language, and models with similar transcript scores can diverge sharply on speech. Replacing earlier spoken user turns with transcripts helps the two weakest models more than replacing the current turn, which points to the representation of prior spoken context as a major recoverable source of degradation.

arXiv:2611.04390 PDF

ReListen: Training-Free Evidence Localization in Hour-Long Audio with Frozen Audio Language Models

Ming-Hao Hsu, Shujie Liu, Yuxuan Hu, Yuxiang Wang, Yingda Shen, Jinyu Li et al. Audio language models cannot fit hour-long recordings in context, and transcribing first discards non-speech sounds and voice information. ReListen is a training-free method in which a frozen audio language model scores each 20-second window by how likely it is to help answer the question, then answers from the highest-scoring minute. The selected minute overlaps the annotated evidence for 88.5% of VoiceGiraffe questions, far more often than transcript retrieval, and the same scan grounds events on LAT-Bench as well as a model trained for temporal grounding. Answering from the selected audio matches or beats hearing the whole recording and beats a full transcript by 9.2 points on VoiceGiraffe, while running faster than transcript-based systems with memory that stays constant regardless of recording length.

arXiv:2611.04480 PDF

TEMPO: Time Series Understanding via Discrete Tokenization

Riccardo Maggioni (Forgis) The question here is how time-series signals should enter a pretrained language model's token sequence when interpretation also needs textual context. TEMPO extends an LLM's vocabulary with discrete codes from a frozen, context-aware Transformer tokenizer that uses finite scalar quantization to map four-step patches to a 625-code vocabulary. Signal tokens, scale statistics, and text then pass through the model's existing self-attention, and only the new embeddings and DoRA adapters are trained. With a Qwen3-4B backbone and task-specific fine-tuning it reaches macro-F1 scores of 89.0% on activity recognition, 85.0% on sleep staging, and 86.0% on bearing-fault diagnosis.

arXiv:2611.04482 PDF

BtsPruner: Role-Aware Visual Token Pruning for Large Vision-Language Models

Zihao Qi, Zhibo Zhang, Miaoxin Zhang, Jincheng Cai, Zengqiang Yan cross-listed Large Vision-Language Models (LVLMs) process hundreds or thousands of visual tokens, and existing training-free pruning methods score all of them with the same criterion. The authors observe that tokens shift during encoding from carriers of local detail to globally referenced representatives. BtsPruner detects this transition point from changes in attention concentration, keeps global representatives from later stages, and recursively retrieves complementary detail tokens from earlier stages. On LLaVA-1.5-7B it retains 91.6% of dense performance with only 32 of 576 visual tokens, and on high-resolution LLaVA-NeXT-7B it gives up to a 5.95× CUDA-latency speedup, with results also reported for Qwen-VL, InternVL, and CogVLM.

arXiv:2611.04537 PDF

Heads That Write, Not Just Point: Image Retrieval Heads in Vision-Language Models

Junsung Park, Sangwon Yu, Donghun Kang, Yeongtak Oh, Jaewon Jeong, Gyeongtae Yoo et al. cross-listed The authors examine how large vision-language models (LVLMs) work out which of several images in their context a question refers to, using a controlled probe called ICIR-MCQ. Based on attention, they find sparse "pointer heads" that reliably attend to the target image; a single such head is competitive with external retrievers on a naturalistic multi-image benchmark. An analysis of each head's output identifies "writer heads" whose outputs carry enough information to pick out the target image, and the two criteria select largely different heads, with only 35–47 shared among the top 100 of each. Ablations show that the heads that matter most for multi-image inference sit in the overlap between pointer and writer heads.

arXiv:2611.04903 PDF

VIGOR: Visual Gain Ordering for Hallucination Mitigation in Multimodal Discrete Diffusion Language Models

Junzhe Chen, Tian Qin, Eugenie Shi, Tianshu Zhang, Qiang Ju, Lijie Wen cross-listed Multimodal discrete diffusion language models (dLLMs) generate text by iteratively unmasking tokens, and they still hallucinate content not supported by the image. The authors argue that part of the cause is premature commitment: some positions are unmasked early because language priors make them look confident, and these early choices bias the rest of the output. VIGOR (Visual Gain Ordering) is a training-free inference strategy that compares each candidate token's confidence with and without the image, and unmasks first the positions most supported by visual evidence. On LLaDA-V and Lumina-DiMOO it consistently reduces hallucination and improves multimodal reasoning across benchmarks.

arXiv:2611.05066 PDF

The Right Source Is Not Enough: Understanding History Interference in Speech-LLM Memory

HongJin Song, Weiqiao Shan, Runwu Shi, Yifei Wu, Yujin Wang, Jiale Luo et al. Speech large language models keep audio history for multi-turn conversations, but irrelevant past audio can change answers even when the queried event is still represented. Controlled experiments on MOSS-Audio and MiMo-Audio show that historical values steer errors toward specific competing answers, which is more than a generic context-length effect. Causal interventions find that in open-ended question answering, most of these errors disappear when direct access to the irrelevant source is blocked. In multiple-choice answering, many errors persist because historical influence is already carried in the question-candidate states. The authors conclude that identifying the correct source is not enough: reliable conversational memory also requires controlling how retained information enters answer computation.

arXiv:2611.05124 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 labels from stronger external judges or on curated hard negatives. PROVE (Perturb-and-Repair Optimization for Vision-Language Evaluation) instead uses self-play in a single shared multimodal model. A Mutator introduces typed errors in objects, attributes, counts, or relations into a correct caption, the verifier judges the perturbed caption and gives feedback, and a Repairer restores a faithful caption for re-verification. The two verification rounds produce structured rewards, and because all roles share parameters, training them also improves the verifier. Fine-grained compositional verification improves across model scales and families, with limited, benchmark-dependent changes on broader multimodal tasks.

arXiv:2611.05158 PDF

What Does the Model See? Representation Validity in Multimodal LLM-Based Video Analysis

Christian Pipal, Eva-Maria Vogel, Morgan Wack cross-listed When social scientists use multimodal large language models (MLLMs) to code video, they must choose what to feed the model: text, audio, video, or a combination. The authors call this choice representation validity and test it across six open-weight MLLMs and three human-annotated video datasets. No input representation is consistently best: three models do better with extracted text and three with native video, and combining inputs rarely helps. Higher classification accuracy also does not reliably yield estimates closer to human-coded results. Even with each model's best representation, 45% of prevalence estimates differ from the human benchmark by more than ten percentage points.

arXiv:2611.05181 PDF

3 more specialized papers

Reasoning 15

Stress-Testing the Robustness of LLM-Based Causal Inference

Ankitkumar Patel, Jigarkumar Patel, Amit Kumar, Supreetha Sreeram, Venkatesh Kulkarni The authors build a stress-test framework for LLM-based causal discovery that varies six factors: model scale, causal graph complexity, prompting strategy, metadata quality, data quality, and handling of uncertainty. They find a consistent "optimism bias": models overestimate causal links, which keeps precision low across the board. Reasoning-heavy models score best, but their scores drop once uncertainty is built into the evaluation. Gains from larger models are fragile and degrade sharply as metadata quality falls.

arXiv:2611.03656 PDF

Low-Rank Hierarchical Merging for Efficient Long-to-Short Reasoning

Zeqiu Yu, Xuesheng Zhang, Wenxiao Zhao, Shixiao Wang Long-to-Short (L2S) model merging tries to combine a long chain-of-thought model's reasoning ability with a base model's concise style, but existing methods rely on linear interpolation and simple activation statistics and lose accuracy on hard problems. Low-Rank Hierarchical Merging (LHM) is a training-free method that estimates each layer's Hessian trace with Hutchinson's estimator on random inputs, so no task data is needed. It combines this with the task-vector norm into a per-layer sensitivity score, then maps the scores to layer-specific merge coefficients that respect both a fusion-error bound and a target length reduction. On six math benchmarks with 1.5B and 7B models it beats methods such as ACM-TIES and ORCA, with over 4% higher accuracy and up to 93.1% shorter outputs than the long-CoT model.

arXiv:2611.03834 PDF

Why Pay for Long Reasoning When Rewrite Do Trick? Causal Effects of Reasoning History Compaction on Future Generation and Agent Cost

Anton Hristov Long reasoning traces make reasoning models expensive, and proprietary models' hidden scratchpads make it hard to study how they keep reasoning short. The authors test whether open-weight models reason less verbosely when their own earlier thoughts are rewritten, with no retraining: before pending tool results are revealed, prior reasoning is rewritten into compact state records, and continuations from the original and compacted histories are compared. Accepted compact histories usually shorten future reasoning, and a coding case study shows less repeated verification after a candidate solution. In software-agent deployments compaction can lower total cost, though task quality and time including the rewrite depend on the target model and the compressor.

arXiv:2611.03852 PDF

On the Primacy Bias in RLVR Training

Phuc Minh Nguyen, Ngoc-Hieu Nguyen, Parshin Shojaee, Chinh Duc La, Duy Minh Ho Nguyen, Rui Zhang et al. It is unclear whether Reinforcement Learning with Verifiable Reward (RLVR) teaches language models new reasoning or only sharpens what the base model already knows. The authors study this through primacy bias, the tendency to keep relying on already-learned representations. Controlled experiments and large-scale analysis show that RLVR on mixed data mostly sharpens performance on easy problems while giving only marginal gains on hard ones. Hard problems turn out to be what actually expands reasoning ability, but training on them for too long causes forgetting on easy problems, which helps explain earlier conflicting reports about RLVR on new versus well-trained domains.

arXiv:2611.04061 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. Problems in program synthesis, symbolic reasoning, planning, and optimization often call for very different ways of solving them, and language models struggle both to pick the right approach and to formulate it correctly. AlgoPilot splits the work across three models. A SteerLM, trained with supervised fine-tuning (SFT) and reinforcement learning (RL), chooses a solving strategy for each input. An expert GuideLM then writes structured formulation guidance for that strategy, and an ExecutionLM produces the final solution. Built on Qwen3-8B and tested on 13 benchmarks spanning 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 usually get long arithmetic right but offer no guarantee on unseen inputs, while tool calls are exact but move the computation outside the network. In Sketchpad, Qwen3.5-0.8B with a low-rank adapter carries out arithmetic on a rendered sheet: each forward pass reads only a three-cell view and emits one step, while a fixed runtime moves the view but never writes a result digit. The authors prove in Lean that the model is then exact on inputs of every length if and only if it is exact on the finite set of views the runtime can present, and all 2,446 views pass on a laptop GPU running the model in a web browser. The certified model handles signed addition and subtraction, long multiplication and division, square roots, and greatest common divisors at any length; it solved 220 of 220 sums of up to 1,000 digits and products of up to 20 digits, against 217 for GPT-6.1 Sol, the best of ten other models tested.

arXiv:2611.04434 PDF

Treating the Symptom, Missing the Cause: Anatomy of a Scientific Reasoning Failure

Anna Gogleva, Thomas Charman, Martin Balla, Karl Gemayel In controlled synthetic regulatory systems, LLM agents must run experiments to find a hidden fault and then restore a healthy phenotype, separating finding a fix from understanding the cause. Across seven LLMs, every model restores the phenotype more often than it identifies the fault; Claude Opus 4.6 and Sonnet 4.6 rescue every organism but diagnose only 55% and 35% of them. Agents encounter evidence that distinguishes the competing causes in 91% of relevant sessions yet still often misdiagnose. Targeted probes show the models interpret decisive evidence correctly when the comparison is spelled out, but struggle to predict what an experiment would show under each competing cause, which is the step needed to use the result to tell causes apart.

arXiv:2611.04473 PDF

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

Shota Takashiro, Takeshi Kojima, Masanori Koyama, Kohei Hayashi, Yusuke Iwasawa Looped Transformers repeatedly update a latent state to spend more computation at inference, but training only on the final-depth output gives no reason for intermediate steps to explore useful alternatives. Best-of-Depth (BoD) treats the whole recurrent trajectory as a set of candidates and minimizes prediction loss across depths, so different valid answers can be matched at different depths without architectural changes. On N-Queens completion, solution coverage rises to 91.03% versus 45.75% for final-depth training, with similar gains on graph coloring. Fine-tuning a pretrained looped language model with BoD also raises pooled code-generation success.

arXiv:2611.04663 PDF

Think Less, Locally: Local Preference Learning for Reasoning Compression

Muyao Yuan, Liyan Xu, Weizhan Zhang, Muyan Jiao, Jiangnan Li, Jinchao Zhang et al. Large reasoning models often keep deliberating after they already have enough information, which inflates inference cost. LoCo finds likely overthinking points in the model's own traces, freezes the reasoning prefix at each point, and regenerates shorter verified continuations from the same state. It then trains on these state-conditioned preference pairs with direct preference optimization (DPO), teaching the model when to stop rather than simply rewarding shorter outputs. Trained only on self-generated math traces, LoCo cuts completion length by 41.7-45.4% on math and 23.4-36.3% on code across eight benchmarks while maintaining or improving accuracy.

arXiv:2611.04814 PDF

Shuffled OPSD: When Semantic Consistency Hurts On-Policy Self-Distillation?

Qingyuan Zheng, Zhihao Dou In on-policy self-distillation (OPSD), one model acts as both student and teacher: the student solves a problem from the question alone, and the teacher, which also sees a worked example, gives token-level feedback on the student's reasoning. Shuffled OPSD randomly shuffles the words of the shared problem statement while leaving the teacher's worked example intact. Across three math reasoning benchmarks and three Qwen3 model sizes, this improves mean accuracy over standard OPSD, by up to 3.52 percentage points on Qwen3-1.7B. Shuffling the teacher-only worked example instead lowers accuracy at every scale, which challenges the assumption that semantic coherence always helps distillation.

arXiv:2611.04958 PDF

ALBI: LLM-Guided Inductive Logic Programming via Iterative Generate-Test-Constrain Refineme

Aboubaker Seddiq Benterki, J\"uri Vain, Gert Kanter, Leelo Aino J\"urgenstein ALBI (Agent for Learning By Induction) is a neuro-symbolic Inductive Logic Programming (ILP) system. It keeps the Learning from Failures loop but replaces the Answer Set Programming hypothesis generator with an LLM, and uses SWI-Prolog to verify candidates and return diagnostic feedback. To check that the models perform symbolic induction rather than guessing from predicate names, the authors introduce semantic anonymization, a metamorphic test that renames predicates. On 30 standard ILP benchmarks, ALBI reaches 100% success with Gemini 3.1 Pro, GPT 5.4 Pro and Claude Opus 4.6, matching Popper in only 1.4 to 1.7 iterations on average. Success stays at 100% under anonymization, with only a moderate rise in iterations.

arXiv:2611.05039 PDF

4 more specialized papers

Reinforcement Learning 14

Preserving the Privacy of Reward Functions in MDPs through Deception

Shashank Reddy Chirra, Pradeep Varakantham, Praveen Paruchuri cross-listed When an agent's actions can be observed, an adversary can use Inverse Reinforcement Learning (IRL) to work out the reward function behind them. That matters in settings like anti-poaching patrols, where the patrol pattern could reveal where the animals are. The authors show that existing differential-privacy approaches, which work by hiding the true preferences (dissimulation), leak significant information to IRL-based observers and give no guarantee on the minimum expected reward. They propose instead an RL-based planning algorithm built on simulation, meaning deliberately displaying false preferences, which protects reward privacy while guaranteeing a minimum expected reward. On multiple 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: a discriminator estimates the density ratio between expert and policy occupancy, and a reward assignment (RA) function turns that ratio into a training reward. The RA function has traditionally been hand-derived from divergence objectives. The authors use an LLM-guided evolutionary search to discover RA functions directly from the performance of the resulting imitation policy, producing Discovered Adversarial Imitation Learning (DAIL), described as the first meta-learned AIL algorithm. DAIL generalizes to unseen environments and policy optimization algorithms while outperforming state-of-the-art human-designed baselines, and the authors analyze why it trains more stably.

arXiv:2510.00922 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 The study asks whether a learned dynamics model that forecasts logged trajectories accurately also works well as a training environment for control policies, using tokamak magnetic control as the test case. The authors collected four datasets from a calibrated free-boundary Grad-Shafranov solver, varying coverage of physics parameters and current references, and trained 16 action-conditioned Transformers at different data budgets. They then trained PPO policies inside the surrogates and evaluated them on the real solver without adaptation. Only the dataset covering both factors improved consistently across conditions, and datasets with identical 83.3% completion rates differed sharply in how long policies survived under their worst conditions (94.2% vs 43.2% of the horizon). The authors therefore recommend evaluating with prediction error, completion, per-condition duration and tracking quality together.

arXiv:2611.03690 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 depends on inferring the task from logged data into a compact latent representation that a policy can condition on. CoTA-DT (Contrastive Task-Aware Decision Transformer) learns those representations from four complementary signals at once: contrastive learning over offline trajectories, alignment with natural-language task descriptions, one-step supervision from a task-conditioned world model, and long-horizon supervision from a task-conditioned Q function, then feeds the result to a Decision Transformer policy. A language-grounded self-adaptive strategy handles both zero-shot and few-shot adaptation to unseen tasks, and notably a single pretrained language model serves as the shared backbone for language encoding, world modeling, Q-function estimation, and policy learning, so parameters are reused across all four roles. Across several offline meta-RL benchmarks it reports consistent gains over prior methods, most clearly 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. Offline multi-task reinforcement learning (MTRL) learns one policy from mixed offline datasets, but it struggles when each task has little data and when skills must carry over between tasks. LaMA-DT (Language-Model Augmented Decision Transformer) uses a single large language model for two jobs. First, it generates extra trajectories guided by natural-language task descriptions, keeping only high-quality ones through a scoring-and-filtering step. Then the same model is fine-tuned as a language-conditioned Decision Transformer on a mix of real and generated data. On Meta-World and MuJoCo multi-task benchmarks it consistently beats competitive baselines, especially when data is scarce.

arXiv:2611.04031 PDF

CANDO: Conservative Action-Noise Directional Optimization for Residual Reinforcement Learning

Zeyu Gao, Chengkai Hou, Chaoyue Li, Zhuangzhe Wu, Jiaming Liu, Yao Mu et al. Pretrained generative policies give robots strong behavior priors, but fine-tuning them with online reinforcement learning can destroy what they learned from demonstrations. CNDO (Conservative Noise Directional Optimization) improves a frozen generative policy by learning state- and noise-conditioned edits to its input noise, replacing fixed latent feasible regions with a soft conservative constraint built from conservative value estimates and gated execution. CANDO adds an action-space residual on top, combining latent-space improvement with local action correction. CANDO reaches 76.1 NAUC and a 91.4% final success rate, beating the best single-space results by 6.6 NAUC and 4.1 points.

arXiv:2611.04543 PDF

Policy Network Diffusion

Jiate Liu, Deheng Ye The authors ask whether diffusion models can generate the weights of reinforcement learning (RL) policy networks, and whether the generated policies generalize. DPPG (Diffusion-based Conditional Policy Network Parameter Generation) compresses network parameters with a transformer autoencoder that patchifies them, then generates parameters with a diffusion model conditioned on the task. A single DPPG model synthesizes high-quality policies across Classic Control, Atari, and MuJoCo tasks and handles larger networks than prior parameter-generation methods. In unseen environments, the generated parameters converge faster and reach competitive performance.

arXiv:2611.04614 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 concentrate probability on a few high-reward outcomes, which hurts diversity when many good solutions exist. Density-Aware Reward Scaling (DARS) estimates how densely outcomes are packed within each on-policy rollout group and divides rewards by that density, which pushes sampling toward being proportional to reward. 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 seen with GRPO and closely recovers the target allocation across modes. In text-to-image prompt adaptation it improves image and prompt diversity while keeping semantic alignment good.

arXiv:2611.04792 PDF

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

Chase J. Metoyer The authors ask whether intrinsic rewards found by automated search help a reinforcement-learning agent learn, transfer to tasks it never saw, and hold up against exploits that were absent during the search. Each candidate reward weights four generic events (a new view, a pickup, a drop, and a toggle), and the search protocol was pre-registered in a public commit before any results existed; CMA-ES and an equal-budget random search together used 1,560 PPO training runs. On the held-out KeyCorridorS3R2 task, the CMA-ES-selected reward beat task-only PPO, episodic novelty, and the random-search reward, though a reward given hand-wired task-specific events did better still. Adding a free reset action, which was absent during the search, collapsed learning on DoorKey-8x8 from a mean area of 0.829 to 0.136, below task-only PPO, which shows that the search gave no protection against an unseen exploit.

arXiv:2611.04838 PDF

CRONOS: Benchmarking Non-Episodic Reinforcement Learning for Robotic Manipulation

Po-Yi Wu, Djengo Cyun-Jyun Fang, Dian Cheng, Tsung-Wei Ke cross-listed Reinforcement learning (RL) fine-tuning of robot policies usually assumes episodic training with frequent scene resets, which is impractical in the real world because it needs a lot of human intervention. CRONOS is a simulation benchmark for RL fine-tuning of pre-trained vision-language-action (VLA) models without such resets, using shared-scene multi-task settings in high-fidelity physics simulators. Naive fine-tuning fails as object configurations and end-effector poses drift, but targeted reset allocation and correcting biases in the pre-trained models mitigate the problem. 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

Catalyzing Interactive HTML Generation Capabilities via High-Fidelity Reinforcement Learning

Xujia Wang, Yixuan Liu, Shaohan Huang, Furu Wei, Lei Hou, Bin Xu et al. Code LLMs can generate HTML pages that render but often fail on functionality and visual quality once a user interacts with them. Using mini-games as a demanding testbed, WebCatalyst trains with browser-interaction reinforcement learning built from three components. A game-level GUI assessor plays and evaluates the pages, Visual Anchor prevents vision-language-model reward hacking, and Action-Trace Credit assigns token-level credit using runtime execution traces. Trained on fewer than 1K requests, Qwen3.6-35B-A3B improves from 68.8 to 75.3 on Cookie-Bench Mini Games and from 48.8 to 56.9 on WebGen-Bench, outperforming Claude Opus 4.7 in average rank across the evaluated metrics.

arXiv:2611.05151 PDF

Stabilizing On-Policy Targeted Guidance in Agentic Reinforcement Learning

Kevin Frans, Alexander M Rush, Sergey Levine, Pieter Abbeel, Ashvin Nair Purely on-policy reinforcement learning (RL) is limited by the base model's coverage, so rewards the policy rarely earns are hard to optimize, even when the reward signal comes from rich textual sources such as rubric-based judge models. The examined framework, targeted guidance, learns from per-turn textual feedback through an on-policy self-distillation objective, with the explicit goal of improving stubborn behaviors without degrading baseline RL performance over the course of training. Across settings in a multi-turn agentic software engineering environment, what works is combining targeted self-distillation with a per-criterion RL reward and restricting feedback to the turns where the policy actually fails. The authors also characterize the design choices that keep entropy stable and reduce oscillatory behavior during training.

arXiv:2611.05266 PDF

2 more specialized papers