HDATF
English
Start a conversation
Menu

Paper Archive

Every archived paper with its summary, highest number first.

Papers 1,193Archive updated BenchmarksRSS feed
  1. 1206ResearcharXiv:2610.12183

    A Closer Look at Agentic BBO: Benchmarking LLM Agents for Black-Box Optimization

    • The authors built AgenticBBO-Bench, a unified-budget benchmark spanning synthetic functions, hyperparameter optimization, database tuning, chip design, and molecular design.
    • The Agentic setting beat Direct on all five families with an average relative improvement of 30.3%, and on BBOB the score rose from 0.460 to 0.661.
    • Adding GP-based numerical tools did not consistently help, so the base Agentic mean of 0.589 was higher than 0.574 for + GP Suggest and 0.538 for + GP Tools.
    • Task and variable semantics raised the mean score from 0.465 to 0.595, while more specific priors were less reliable.
    • Handing the search to TuRBO after the agent warm-up scored 0.570, close to continued Agentic optimization and clearly above TuRBO started from the initial observations at 0.475.
    • On the frontier challenge GPT-6 Astra scored the highest mean of 0.591, closely followed by DeepSeek V4.1 Flash at 0.584, yet no model dominated every domain.
    AbstractPDF
  2. 1205Research AgentsarXiv:2610.12086

    EvoAlloc: A Self-Evolving Resource Allocation Agent for Efficient Program Evolution

    • In program evolution, evaluation is computationally expensive, so allocating limited resources to promising candidates was a central decision of the search.
    • EvoAlloc periodically consolidated prior search and allocation outcomes into reusable experience, which informed subsequent strategy revisions.
    • A counterfactual exploration mechanism occasionally evaluated candidates denied resources by the allocator, revealing hidden outcomes and enriching its experience.
    • Across coding and agent-harness optimization benchmarks, EvoAlloc required 59–82% fewer full evaluations and 61–89% fewer LLM tokens, and achieved 8.7–12.0% higher final performance under the same budget.
    • With the same Full-evaluation budget on the same seed, EvoAlloc examined 124 proposals on ADAS-AIME versus 75 for ShinkaEvolve, widening the search.
    • Across exploration rates, final performance was similar but tokens to the target varied substantially, and ρ = 0.3 was the most efficient.
    AbstractPDF
  3. 1204RoboticsarXiv:2610.11722

    From Solo to Ensemble: A Hierarchical Framework for Composable Multi-Agent Human-Object Interaction

    • The coordination problem where multiple humanoids handle the same object was hard because contact locations, motion directions, and force application must all be matched together.
    • The authors built a distillation-based hierarchical learning framework that converts single-human HOI knowledge into a reusable Object-oriented Motion Skill.
    • In downstream tasks the distilled skill is reused as a frozen executor, while a high-level policy generates region-wise object motions conditioned on the shared object and task goal.
    • On carrying tasks the method reached a success rate of 85.63 on four-human H-box carrying, higher than CooHOI at 80.96.
    • Although the skill was distilled only from single-human carrying, it supported two- and four-human sofa pushing by changing only the high-level policy.
    • In the standalone skill evaluation, the teacher-rollout targets gave a position error Epos of 79.8 (mm), a reach rate TTR of 100.0, and a trajectory success Succtraj of 98.9.
    AbstractPDF
  4. 1203RoboticsarXiv:2610.12457

    SpatialHarness: Test-Time Spatial Scaffolding for Fine Robotic Manipulation

    • SpatialHarness maintains an online simulated scene synchronized with real-world execution, identifies task-critical spatial relationships, and renders complementary virtual views.
    • It supplies spatial information to a frozen multimodal policy without policy fine-tuning or changes to the physical sensing setup.
    • With the same frozen GPT-6 Astra policy, plug-insertion success rose from 26.7% to 66.7% and Tower of Hanoi from 0% to 100%.
    • The average success rate improved from 18.33% to 83.33%, a gain of 65.00 percentage points.
    AbstractPDF
  5. 1202AI / AgentsarXiv:2610.12002

    Agentic-TTT: Training test-time policy for test-time training

    • Agentic-TTT turns TTT procedures into callable tools, treats accumulated skills as an evolving deployment environment, and trains a test-time policy by reinforcement learning.
    • As the compute penalty rose from 0 to 0.4, the direct-answer rate increased monotonically from 11.8% to 39.8% while adaptation compute fell 35.5%.
    • Moderate penalties (0.1 and 0.2) cut compute by about 30% while achieving higher mean utility than no penalty.
    • Raising the tool weight from 1 to 4 improved final utility by 0.046 and reduced the invalid-action rate from 4.08% to 1.72%.
    AbstractPDF
  6. 1201AI / AgentsarXiv:2610.11345

    SynCo: Data Synthesis Co-Training for Self-Evolving LLMs via Multi-Agent Reinforcement Learning

    • SynCo is a data synthesis co-training framework that jointly optimizes an independently parameterized Synthesizer and Reasoner via multi-agent reinforcement learning.
    • The Synthesizer constructs training tasks from the Reasoner's evolving capabilities, resolving the capability-experience mismatch of static datasets.
    • On mathematical reasoning SYNCO reached 78.72%, outperforming the strongest external baseline by 3.76 points and the frozen synthesizer by 2.87.
    • As training progressed, the share of gradient-bearing tasks fell from 31.3% to 9.5% while all-correct tasks rose from 34.4% to 59.9%.
    • Only 16.2% of 2,940 verified tasks gave non-degenerate feedback, and teachability tracked this property at r = 0.934 while novelty was irrelevant at r = -0.041.
    AbstractPDF
  7. 1200Machine LearningarXiv:2610.12190

    DataSense-Bench: The First Step Toward an AI Scientist

    • The agent selected disjoint, equal-sized subsets of candidate training data and ranked them by expected value for post-training performance.
    • The benchmark evaluated Qwen3-4B performance after training on the predicted top-ranked group under a standardized protocol.
    • On BFCL, all six agents' first groups exceeded random selection, but gains ranged from only 0.35 to 1.78 % points.
    • On TBLite, two of ten agents underperformed random selection, and the leading agent improved on it by only 0.99 points.
    • Ranking ability did not hold consistently across tasks. GPT-6 Astra identified the best group in all three BFCL runs but in only one of three TBLite runs.
    • Selection recipes varied substantially across three independent runs of the same agent, changing the selection rule itself.
    AbstractPDF
  8. 1199AI / AgentsarXiv:2610.11768

    Narrow and Deep: An Ontology Tower as the Knowledge of an LLM Agent for an Industrial Equipment System

    • The authors proposed an ontology tower for a single equipment system, which deepens through quantities derived from measured points by physical relations and lessons incorporated from the operating journal.
    • The studied tower consists of 162 nodes and 217 edges, holding 42 tags, 34 assumptions, 33 bindings, 24 knowledge nodes, 16 derived quantities, and 13 plant elements.
    • In a preregistered evaluation, the text projected from the tower raised the primary-trap avoidance rate of the four models from 0.605 to 0.802.
    • With the projected text, the 9-billion-parameter model scored above the largest model without it, at 0.737 against 0.672.
    • Lessons left in the operating journal were seldom used by the agents, and were used when incorporated into the tower or placed in the prompt as records.
    • In live runs the agents brought the controlled variable into its target band in 12 of 14 runs, and no setpoint changed except through the safety layer.
    AbstractPDF
  9. 1198Software EngineeringarXiv:2610.11578

    Chronos Enables Code Agents to Reason over Software Evolution

    • Chronos distilled merged pull requests into experience cards with four fields: Summary, Signals, Motivation, and Solution.
    • It built an evolution graph connecting the cards through eleven typed relations of code continuity, developer intent, and organizational context.
    • Semantic search located entry cards and weighted multi-hop expansion retrieved connected history for the agent to selectively read.
    • On SWE-Bench Verified it improved SWE-Agent across all six LLM backbones, raising the mean resolution rate from 69.2% to 72.9% and reaching 79.8% with MiniMax M2.5.
    • With the same backbone it raised resolution rates from 48.3% to 51.7% on SWE-Bench Pro and from 41.0% to 43.5% on FEA-Bench Lite.
    • In a human evaluation on 100 tasks, graph-grounded retrieval increased the mean useful cards from 1.24 to 2.87 and the fraction of tasks receiving useful experience from 68% to 89%.
    AbstractPDF
  10. 1197Research AgentsarXiv:2610.12419

    OneSearch-VL: Unified Multimodal Deep Research Agent for Image and Video

    • It encodes the links from on-screen visual anchors to real-world entities, source-supported web facts and answer-producing operations (comparison, arithmetic, counting) in one graph, VGEG, used for data construction, reward and evaluation.
    • A data engine that narrows 2.5M web videos to 70k produced about 110k training trajectories (36k single-image, 37k multi-image, 35k video) and about 10k RL tasks.
    • Beyond answer correctness and query quality, the RL reward adds EVGR, which separately scores whether tool observations actually support the required facts (evidence traceability) and whether the right objects and frames were found and used (visual grounding).
    • OneSearch-VL-8B beat the same-size Qwen3-VL-8B agent by 20.2 points on OneSearch-MI-Bench, 17.6 on OneSearch-Video-Bench and 27.0 on VideoDR, and averaged 58.3 on the image benchmarks, 1.7 points above OpenSearch-VL-8B.
    • Joint SFT on all three input types reached a benchmark average of 55.8 versus 51.7–55.1 for single-type training, and the EVGR reward raised the average from 57.3 to 61.1.
    • The authors note that external search tools and changing webpages limit exact reproducibility, and that automated VGEG construction and model-based rewards may introduce annotation errors or judging biases.
    AbstractPDF
  11. 1196AgentsarXiv:2610.11464

    Who Verifies the Verifier? Co-Evolving Inspectable Graders with Self-Improving Agents

    • The verifier is built as an expression over small drawback detectors (static checks, execution checks and narrow LLM queries) and evolves by synthesizing detectors from clustered failures and retiring useless ones.
    • On MBPP+ the evolved verifier raised held-out agreement from 0.417 for the hand-authored seed composition to 0.625, also above the bare LLM judge it contains (0.55).
    • With the two anchor guards switched off, the verifier collapsed into a grader that passes 0.97–1.00 of everything it grades, yet skills trained under it scored about as well as under the anchored verifier.
    • Double Ratchet, which co-evolves the verifier with the skill loop, kept 106% (MBPP+), 110% (Spider 2.0) and 88% (report generation) of the lift the same loop gets from ground-truth grading.
    • In report generation, skills gamed the rubric by writing tags in place of numbers; an outer judge caught it, and one added detector cut valueless tags from roughly 30% to about 1%.
    • The MBPP+ lift pointed the same way on all three seeds, but the task-level test stayed at p = 0.12, and the study covers three task families and one Claude model family.
    AbstractPDF
  12. 1195AgentsarXiv:2610.10444

    RunningTab: Direct Workspace Interaction with Environment-Side Tabs

    • Solves the problem of files being read but never reflected in the final deliverable with an environment-side ledger.
    • The agent adds requirements, while the environment records the provenance of read content and a list of files still unopened.
    • If the agent tries to finish with unresolved requirements, a final check sends them back for review.
    • Fits naturally with accessing workspace files directly via a terminal, without any pre-indexing.
    • Showed consistent performance gains across three benchmarks and three LLMs.
    AbstractPDF
  13. 1194AgentsarXiv:2610.10513

    SciExam for ENSO: Can AI Agents Build Climate Models?

    • Built the first family of benchmarks that grade results solely with hidden scientific criteria (statistics, variable dependence, forecasting skill), with no answer key, rubric, or LLM judge.
    • Agents must freeze their own diagnostic metrics in advance, so they cannot tweak the grading criteria to fit their results.
    • Six of twelve agents scored higher than the published reference model, with strengths in variable reconstruction and forecasting.
    • The top agents' model structures were each compatible with one of the two competing hypotheses explaining ENSO asymmetry, producing results directly relevant to an open scientific debate.
    • Scores held up even when calendar years were hidden, confirming the agents were not memorizing the observational record.
    AbstractPDF
  14. 1193Computer VisionarXiv:2610.13947

    Towards Physically Grounded Articulated 3D Assets: A Multi-Agent Pipeline and Benchmark

    • Separated the design, coding, and review agents to reduce the bias of a single agent grading its own output.
    • Added physical simulator feedback to the validation stage, catching collision-free joint motion and gravity stability during generation.
    • Unlike prior work, it also generates physical attributes such as mass and material, targeting assets ready for robot simulation.
    • Released a unified benchmark covering 100 reference assets, 50 stability prompts, and 50 objects with measured masses.
    AbstractPDF
  15. 1192LLMs / NLParXiv:2610.13912

    RSI-Anything: Unlocking Recursive Self-Improvement with World Models

    • Continuously checks how well the world model's predictions match actual observations, automatically deciding whether to update the policy or the world model.
    • In code repair, adding world-model-guided selection and adaptation raised the success rate step by step from 26.3% to 29.2% to 31.2%.
    • In chemical reaction optimization, it achieved a mean best observed yield of 96.15% within a 40-experiment budget, far ahead of Bayesian optimization at 82.16%.
    • Laboratory planning scores more than doubled from 21.80% to 46.01% across five policy versions, showing that improvements accumulate.
    • An immunology application connected workflow evolution with predictive modeling in an end-to-end demonstration of therapeutic hypothesis generation.
    AbstractPDF
  16. 1191Research AgentsarXiv:2610.07184

    Learning Scientific Exploration from Human Research Decision Trajectories

    • Problem: scientific exploration is a sequence of research decisions and actions, yet papers mainly record final outcomes, so abandoned ideas, omitted experiments and the ablations that shaped the result drop out of the record. Trajectories from synthetic environments may differ from how human researchers decide, so the authors use commit histories of public GitHub repositories as proxies for exploration.
    • Commits as proxies: day-to-day commit histories are messy signals that mix in low-level operations such as implementation updates, training-config changes or dataset switches, and decisions map to commits many-to-many. Of 117,852 commit assessments across the 599 annotated pairs, 56.6% were irrelevant, and only about 4% of the roughly 15K processed papers yielded a progressive decision trajectory.
    • Pipeline: GitHub links come from arXiv PDFs; a repository is kept only if its README holds the full paper title, with at least 5 pre-publication commits, the first at least 30 days before publication. GPT-5.6 Sol reads commits in time order and records decisions; a cited diff excerpt counts as evidence only as an exact substring of a stored patch, and repositories with fewer than 5 validated decisions are dropped.
    • Dataset: 599 paper and repository pairs yield 13,271 decisions, 51% experiment, 39% method and 9% ablation; 76.3% were retained, 11.5% abandoned and 12.1% superseded. Trajectories hold 5 to 205 decisions (median 15), and a decision has 7.5 evidence citations on average.
    • Evaluation: judging whether an arbitrary decision is effective remains open, so held-out real trajectories serve as user simulations, and LLM judges (GPT-5.6 Sol and Opus 5) rate whether a predicted next decision is strategically similar to the actual one. The score sums component and operation match, each 0 to 2, for 0 to 4. Fleiss' Kappa over repeated scoring of 100 fixed baseline predictions was 0.691 to 0.831.
    • Test time: Gemini 3.1 Flash Lite sees the first 1 to 3 decisions of 134 held-out repositories and predicts the next. On average, retrieving the 2 most relevant of 19 training trajectories beats 2 random demos by 0.196 and all 19 by 0.101. Skills distilled by Opus 5 beat retrieval by 0.181, final-paper skills by 0.286 and the baseline by 0.290, and beat skills from order-shuffled trajectories at each prefix length.
    • Training: Qwen3-8B gets SFT on 375 samples from 125 training trajectories, then Dr. GRPO reinforcement learning. In SFT, test perplexity keeps falling while the judge score peaks at epoch 3, so the reward optimizes generation quality directly. GPT-5.6 Sol scores rise from 1.025 (base) to 1.199 (SFT, epoch 3) and 1.505 (GRPO, epoch 7); Opus 5 scores rise from 1.219 to 1.333 and 1.674.
    • Limits: trajectories are LLM reconstructions of public repository histories, which the project blog calls development observed in public repositories, unverified by the paper authors. The score measures similarity to the actual decision. Experiments cover AI conference repositories on Git; fields such as biology and physics, and non-Git platforms, are future work. The data, code and trained adapters are public.
    Full textCodeAbstractPDF
  17. 1190Human-Computer InteractionDOI:10.1007/s10055-026-01497-x

    Spmark: improving small-target text selection with smooth-pursuit gaze interaction in hybrid paper–digital environments

    • Selecting small text with gaze is difficult because of spatial tracking error and unintended activation. SPMark combines gaze location with smooth pursuit of moving targets to distinguish selection intent.
    • The system activates interaction objects near the gaze position and combines spatial proximity with gaze-target motion correlation to lock the intended target. It was implemented on HoloLens 2 for annotation in a hybrid paper-digital setting.
    • The authors evaluated the system in a controlled comparison and open-ended exploration with 21 university students. They report higher task completion rates and lower perceived effort and frustration than the baseline method.
    • Both comparison conditions used a keyboard for final confirmation. Eye fatigue during extended use, lighting and latency effects, and retuning for other devices remain limitations. Accessibility for users with impairments was not established, and integrating automatic text recognition from real paper remains future work.
    Full textAbstractPDF
  18. 1189Machine Learning / SecurityDOI:10.1186/s42400-026-00679-z

    RFLDefense: robust federated learning defense against targeted attacks

    • Federated learning can be poisoned through malicious model updates even when raw training data remain local. This work studies defenses against targeted backdoor attacks under heterogeneous client data distributions.
    • The method combines hybrid similarity analysis with layer-wise sign-consistency checks. Inspecting individual layers aims to identify localized anomalies without discarding all benign contributions from a client.
    • Isolation Forest anomaly scores guide adaptive weighting and robust layer-wise aggregation. The authors report that benchmark experiments suppress several backdoor attacks while retaining performance on clean inputs.
    • The threat model assumes an honest, uncompromised server and no more than half of the participating nodes being malicious. The method adds server-side computation; additional implementation details and supporting experiments are available on request. We have not verified a public code implementation or reproduced deployment results.
    Full textAbstractPDF
  19. 1188LLMs / NLParXiv:2610.10455

    PHRBench: A Behavioral Evaluation of Post-Hallucination Reasoning in LLMs

    • Final-answer correctness alone cannot distinguish compliance, avoidance, and correction, so they built a framework that independently classifies each response trajectory into three behaviors (Hallucination Compliance / Avoidance / Heuristic Correction).
    • Hallucinated context reduced accuracy by 7.7 percentage points on average, and across several model families they found a scaling tension: larger models become more vulnerable while also correcting more often.
    • Successful recovery (correction plus reaching the correct answer) is rare, and recovered trajectories show a signature of more frequent belief updates.
    • Using only prompt-level structural and semantic features, a lightweight predictor anticipated successful recovery before generation, achieving AUROC 0.847 and accuracy 0.863.
    AbstractPDF
  20. 1187LLMs / NLParXiv:2610.10478

    Before They Can Solve: Predicting Post-Training Coding-Agent Performance from Base Models

    • They built a framework that replays successful trajectories to find the decisive step T* where tests first flip from failing to passing, and probes the base model directly at that point.
    • The three probes (Decisive-Action BPB, Patch MCQ, prefix-conditioned pass@K) showed strong rank agreement with post-trained SWE-bench Verified pass@1, at Spearman rho 0.964, 0.903, and 0.988 respectively.
    • Existing non-agentic coding benchmarks (HumanEval and others) were unstable, with correlations from -0.39 to 0.83, confirming they are ill-suited for base model selection.
    • These probes require only a benchmark's successful trajectories and its verifier, so future agentic benchmarks can be repurposed as base-model evaluations.
    AbstractPDF
  21. 1184AI AgentsarXiv:2610.12452

    BrickBench: Evaluating Agentic Brick Design

    • The best coding agent satisfied about 95% of prompt semantic requirements and produced physically valid assemblies for nearly every prompt.
    • Removing the BrickAgent tool environment crashed valid-assembly rates from 100% to 40% for GPT-6 Astra and to under 1% for GPT-5.6 Luna, showing that grounded validation tools are the key to success.
    • General coding agents outperformed specialized LEGO-generation models (BrickNet-14B, BrickGPT), though at 600 to 6800 times the cost.
    • Human raters chose the human design in 323 of 360 comparisons (about 90%), showing that meeting requirements and designing well are different problems.
    • They computed the design-quality ELO metric with a VLM judge and validated its ranking agreement with human preferences (Kendall tau = 0.89).
    AbstractPDF
  22. 1183AI / GeneralarXiv:2610.12176

    Recursive Self-Improvement through Multi-Agent Self-Supervision

    • By alternating workflow optimization with retraining on self-generated trajectories, the method achieves recursive bootstrapping where the improved model becomes a better optimizer and evaluator.
    • In just two cycles, performance per output token improved 1.2-1.6x across four research benchmarks.
    • Self-evaluation accuracy rose from 73% to 93%, showing that evaluation ability improves alongside task solving.
    • Training on multi-agent trajectories was more efficient than single-agent training that used 1.4x more tokens.
    • Optimized workflows showed more role-specific information handoffs and routing, supporting the information-flow optimization hypothesis.
    AbstractPDF
  23. 1182Computer VisionarXiv:2610.12374

    AgentGarten: Code Worlds for Evolving Agents

    • A division of labor where the engine handles state and rules while a neural renderer handles only visuals lets new virtual worlds be built entirely in code without 3D assets.
    • The Adversarial Forcing technique distills a pretrained video model into a real-time renderer, keeping visual quality from collapsing over long rollouts.
    • Agents learned strategies like shelter building in just 4 rounds, millions of times fewer than conventional reinforcement learning.
    • Agents distilled their experience into written playbooks that subsequent agents inherited and refined, producing evolution across generations.
    • Coding agents can automatically generate and extend new worlds from a single image or text description, securing environment scalability.
    AbstractPDF
  24. 1181LLMs / NLParXiv:2610.11959

    MiMo-V2.6: Scaling Reinforcement Learning Towards Self-Improvement

    • Asynchronous training that consumes 1,568 samples and 2.7-3.7B tokens per step handled long tasks with context lengths up to 1M tokens.
    • Environments across coding, general workflows, visual tasks, and cybersecurity were mixed with diverse agent harnesses to improve generalization.
    • Instead of binary pass/fail grading, groupwise grading that compares answers within a group produced more precise reward signals.
    • Freezing the MoE router and building multi-layer defenses against reward hacking kept large-scale RL training stable.
    • Training dynamics, RL environments, and the RL framework were fully open-sourced, creating a reproducible baseline for large-scale agentic RL research.
    AbstractPDF
  25. 1180AgentsarXiv:2610.10507

    RECAST: Learning to Compute the Right Context through Adaptive Evidence Routing

    • They formalized the problem of deriving evidence through filtering, aggregation, and computation rather than mere retrieval as a sequential decision process.
    • RouterLM alternates between calling primitive operations and synthesizing custom code to build evidence, trained with SFT and GRPO.
    • It achieved a 75.6% mean success rate across six heterogeneous benchmarks, 15.9 percentage points above the strongest baseline.
    • The compact 9B trained model outperformed a training-free Gemini 3.5 Flash router by 5.0 percentage points.
    • On three unseen benchmarks it stayed 15.0 percentage points ahead on average, demonstrating zero-shot generalization.
    AbstractPDF
  26. 1179RoboticsarXiv:2610.10498

    EmbodiedRSI: Active Continual Robot Learning Through Hypothesis-Guided Co-Evolution

    • Value-of-Information-based experiment selection reduced the number of robot trials and improved exploration efficiency.
    • They built a co-evolution structure that evaluates code and skill hypotheses separately and then measures the gain of running them jointly.
    • It reached 77.0% overall and 71.3% on Composite-Unseen in RoboCasa365, far ahead of the best baseline at 40.1%.
    • It recorded 86.8% success on LIBERO-Pro and achieved 71.3% when transferring zero-shot from simulation to a real robot.
    • They proposed a three-level hierarchical memory of execution records, hypotheses, and distilled knowledge that is reused for selecting the next experiment.
    AbstractPDF
  27. 1177AgentsarXiv:2610.10536

    Decoupling Exploration from Optimization in RLVR

    • They confirmed that adding a novelty bonus directly to RLVR degrades general capabilities outside the supervised reward signal.
    • They proposed ExpDis, which separates an explorer model from a student model and distills only the correct solutions filtered from the explorer.
    • Across seven math benchmarks it beat DAPO in both average accuracy and pass@k under the same compute budget.
    • It showed continued gains as parallel explorers and repeated rounds were scaled independently.
    AbstractPDF
  28. 1176RoboticsarXiv:2610.12126

    SuperNav: An Agentic Navigation System for Any Task in Any Scene

    • It achieved generalization with no navigation-specific fine-tuning, simply by equipping a pretrained MLLM with a dedicated agent harness.
    • The model issues movement commands through a visual-point interface where it clicks the destination directly on an image, and revises decisions from execution feedback.
    • It reached 78.00% success on single-object navigation, 44 percentage points ahead of the strongest baseline UniNaVid at 34.00%.
    • Without any OVON training or adaptation it recorded a 73.33% success rate over 120 HM3D-OVON episodes.
    • It validated real-world applicability by deploying on an actual quadruped robot as well as in simulation.
    AbstractPDF
  29. 1175RoboticsarXiv:2610.12369

    Embodied Turing Machines: Stateful Code for Robot Recursive Self-Improvement

    • They proposed COAP, a new paradigm that controls robots with code alone, without any VLA or VLM.
    • Storing state directly in code variables made the basis for decisions transparent, and failure recovery by backtracking and retrying became flexible.
    • They designed a recursive self-improvement (RSI) loop where a coding agent (Opus 5.5 Max) repeatedly improves the code library without human intervention.
    • It achieved a 70.24% success rate on RoboDojo's 42 bimanual tasks with no model at test time, 38.9 percentage points above the previous best.
    • 83% of a new task's code was reused from the shared library, showing a structure where capabilities accumulate as tasks pile up.
    AbstractPDF
  30. 1174AgentsarXiv:2610.10468

    A Society of Researchers: Designing Institutions for Populations of Autonomous Research Agents

    • The authors argue that if thousands of agents are left unmanaged, they will form their own organization anyway (as shown by a real incident), so designers should provide the structure explicitly.
    • They transplanted the institutions of science directly into an agent society: principal investigator agents submit proposals to open calls, and independent review decides which ones receive compute budgets.
    • They proposed six design principles: give purposes not goals, govern through institutions, rule by allocation rather than assignment, keep diversity alive, make verification an institution, and let the society remember.
    • They actually ran a society of ten thousand agent researchers, and one lab found a way to reach the same quality with about 30% less compute.
    • Citing an existing analysis that only about 5% of 4,000 LLM-generated research ideas were not duplicates, they emphasized the need for diversity and verification institutions.
    AbstractPDF
  31. 1173AI / GeneralarXiv:2610.10515

    RoboJEPA: Scaling Robotic Latent World Models

    • They ran scaling experiments by training world models from 22M up to 8B parameters on data from 12 robot platforms (about 2.87M trajectories).
    • Prediction error decreased following a power-law as compute increased, and the law accurately extrapolated to larger scales than those measured.
    • Offline prediction error correlates strongly with real robot planning success rate, making it possible to gauge model quality without physical experiments.
    • Given only a single goal image, a real robot arm could zero-shot plan and carry out long-horizon tasks such as grasping and pick-and-place without any additional training.
    • The 8B-parameter model is the largest JEPA predictor built to date, and all checkpoints and code have been released.
    AbstractPDF
  32. 1172RoboticsarXiv:2610.10528

    Long-WAM: Scaling the Context of World-Action Models

    • Extending the visual history from 0 to 19.2 seconds raised manipulation success from 63.3% to 78.7%, but only for models pretrained autoregressively.
    • They used a two-stage approach: first pretrain a foundation model that predicts the future from the past on 10,000 hours of robot and egocentric video, then adapt it to robot actions.
    • Thanks to streaming encoding and low-precision quantization, an action chunk is generated in 107.4 ms even on a consumer GPU, enabling real-time control.
    • On the task of grasping and stacking cups on a moving conveyor, Long-WAM succeeded 95% of the time, while both comparison baselines failed all 20 trials.
    • Combined with an upstream planning model (GPT family), overall success rose from 31.4% to 54.4%, showing that good planning pays off only with a strong executor.
    AbstractPDF
  33. 1171Machine LearningarXiv:2610.08735

    Optimal and Efficient Online Inverse Optimization

    • In a setting where the hidden linear criterion is learned only from the expert's observed choices, the algorithm achieves the optimal regret proportional to the square root of the dimension d.
    • The previous optimal algorithm required nearly exponential-time computation per round, and this paper presents the first polynomial-time version.
    • The key idea is a 'revocation' rule that undoes a metric update once the query point moves too far from where the update was made.
    • The algorithm is deterministic and guarantees optimal regret for every horizon T without any randomness.
    AbstractPDF
  34. 1170RoboticsarXiv:2610.08789

    QF3: Fast Flow RL with Filtered Q-Gradients

    • The critic's gradient is passed through a one-step prediction of the flow policy, directly improving the policy without costly differentiation through the sampler.
    • Learning is stabilized by a 'filter' that cuts off action dimensions where the gradient moves too far from the replay action.
    • A walking and 3-minute dance motion-tracking policy for the 29-DoF Unitree G1 humanoid was trained from scratch in simulation and transferred to real hardware without correction, a first for off-policy flow RL.
    • Training is 10x faster in wall-clock time than the recent on-policy method FPO++, and comparable to the classic TD3.
    • Fine-tuning a pretrained robot-arm policy made the bottle-gathering task finish episodes 22% faster while raising the success rate from 91% to 95%.
    AbstractPDF
  35. 1169Machine LearningarXiv:2610.06805

    H-JEPA: End-to-End Learning of Hierarchical World Models for Visual Planning

    • The world model is stacked into three levels instead of one, with each level predicting the future at a different temporal stride (5/10/20 steps).
    • Higher levels discard fast-changing detail and keep only slowly varying, important state, learning representations that match abstract goals well.
    • On the visual maze-navigation task, the three-level structure raised the success rate from 18% to 73% over the flat single-level structure while using less planner compute.
    • The hierarchy emerged purely from prediction learning, with no human-provided reward and no pixel reconstruction.
    • With an inverse-dynamics auxiliary loss, hierarchy also improved planning fidelity on the real-robot video dataset DROID.
    AbstractPDF
  36. 1168Machine LearningarXiv:2610.05538

    LiFT: Loop Flow Transformers

    • It draws a straight path from the model's initial estimate to the target, and trains each loop to hit a point-specific goal along that path.
    • The model keeps improving well beyond the number of loops it was trained with, so inference compute can be scaled freely at test time.
    • On ImageNet 256x256 it cut FID by 3.34 points while using about 60% fewer parameters and about 52% less inference compute.
    • It is a case of applying the test-time scaling idea to image generation at low cost.
    AbstractPDF
  37. 1167RoboticsarXiv:2610.05882

    Mulligan: Performance-Guided Data Collection for Efficient On-Robot Learning

    • They observed that 95% of failures concentrate in 25% of the initial states, so uniform collection wastes operator time.
    • By retrying failed states and filling in unexplored ones, the robot learned faster under the same collection budget.
    • Combining a value function that also learns from failures (HiL-IDQL+Mulligan) improved final real-task success by 10-34 percentage points.
    • They validated the results on 2,550 blinded held-out episodes for higher reliability.
    AbstractPDF
  38. 1166Machine LearningarXiv:2610.06851

    Base Models Can Reason By Taking a Cue From Training Data

    • Fixing only the first two token cues raised Olmo-3-7B's MATH-500 accuracy from 42% to 78% (the RL version reaches 75%).
    • They showed that RL does not so much teach new abilities as increase the probability of existing reasoning cues appearing.
    • By causally editing the training data, they made 'chicken' work as a reasoning cue just like 'okay', proving the cue effect comes from the data.
    • In a safety case study, different cues elicited different refusal and compliance behaviors, with implications for safety evaluation.
    AbstractPDF
  39. 1165ResearcharXiv:2610.06783

    Truly Subquadratic 3SUM and Truly Subcubic APSP via Triangles in Sparse Lopsided Graphs

    • For the first time in 50 years, 3SUM was solved polynomially faster, reducing O(n^2) to O(n^1.9992).
    • A single thin matrix multiplication algorithm speeds up many problems at once, including APSP and Exact Triangle.
    • The AI model Claude discovered the core algorithm, the human authors understood and extended it, and the proofs were verified in Lean 4.
    • Several hypotheses of fine-grained complexity theory are refuted, with major impact on lower-bound research.
    AbstractPDF
  40. 1164safetyarXiv:2609.40286

    Linguistic Loopholes in LLM Unlearning: From a 174-Language Benchmark to Coverage-Aware Unlearning

    • It built a cross-lingual unlearning benchmark spanning 174 language-script pairs and 25 paraphrase types to measure when forgetting spreads across languages.
    • It showed the surprising result that collecting individually strong languages does not reliably form a good source set, and proposed coverage-prediction-based COVER.
    • COVER needs only benign calibration data and access to the frozen model at deployment, reducing scarring without full multilingual retraining.
    • Across three model families and two forget sets it lowered residual access by 7.8-27.3% on average versus uniform selection, and the effect held on real low-resource news documents.
    AbstractPDF
  41. 1163evaluationarXiv:2609.40190

    Cheap to Draw, Expensive to Trust: Certifying Test-Time Scaling Curves

    • It pointed out that only a confidence band covering all budget points at once protects the choice of how many samples to draw, and computed the cost of existing fixed designs (192,000 answers for 100 questions and 64 budgets).
    • Since about three quarters of the variance in selected-answer accuracy at budget 64 comes from differences between questions, an audit that revisits every question need not pay that cost.
    • It built a paired audit based on an exponential inequality for two independent draws, reaching the minimum cost up to a logarithmic factor with no pilot.
    • In a real MMLU-Pro study it certified the whole curve with 79,133 answers, within 0.6% of what a cost law fitted beforehand predicted.
    AbstractPDF
  42. 1162efficiencyarXiv:2609.40316

    Scaling Laws for Looped Mixture of Experts

    • Unlike prior work that studied recurrence (looping) and sparsity (MoE) separately, it combined both axes into a single law and predicted held-out loss more accurately.
    • It showed with data that the recurrence benefit saturates rather than growing indefinitely, and that having more experts sustains the benefit longer.
    • Sparsity delivers about 3x active-parameter efficiency, recurrence about 2x total-parameter efficiency on reasoning, and combining them pushes the performance frontier further.
    • At trillion-token scale with matched training compute, a looped MoE matched a non-looped MoE twice its size, giving a practical guide for memory and compute budgets.
    AbstractPDF
  43. 1161memoryarXiv:2609.40195

    MemLife: Curating and Reasoning over Long-Term Egocentric Video Memories

    • Instead of storing all video verbatim, it converted footage into first-person text episodes anchored on people and places, leaving memories in a form that is easy to find later.
    • It solved the problem of similar items competing in retrieval as memories pile up with a time-indexed agentic reader, improving performance without any training.
    • It beat the strongest training-free baseline by 4.6-12.0%, and optimizing the writer itself with reinforcement learning added another 2.7-5.0%.
    • The gains held across different writer and reader models, making it a general design not tied to a specific model.
    AbstractPDF
  44. 1160roboticsarXiv:2609.40306

    DynaHarness: A Dynamic Physical Harness for Self-Evolving Robot Agents

    • Instead of treating a failure as 'the robot was bad', it localized which part (planning, execution, or recovery) went wrong and fixed only that part.
    • It split the slow brain (planning) and the fast brain (monitoring) into two layers on different time scales, so ambiguous commands are refused and alternative actions are sought.
    • Before a fix is admitted, it must pass regression checks, preventing the risk of breaking previously working functions while repairing.
    • On 800 new LIBERO-Pro initial states it scored 75.2% versus 17.5% for the frozen policy, more than quadrupling performance with the harness alone and no retraining.
    AbstractPDF
  45. 1159Machine LearningarXiv:2610.02199

    TACO: Ternary Absolute-max Column-wise One-sparse Optimizer for LLM Fine-Tuning

    • Nearly eliminated optimizer state by using a sparse geometry that updates only the sign of the largest-magnitude coordinate in each column.
    • Reduced optimizer state memory by 174x (27.7GB to 0.16GB) and peak training memory by 2.9x (80.6GB to 27.5GB) versus AdamW8bit (OPT-13B).
    • Achieved 94.22% accuracy on OPT-13B SST-2, close to Adam's 95.3%, while cutting memory from 254.8GB to 27.5GB.
    • Enabled full-parameter fine-tuning of 30-32B models on a single 80GB H100 GPU across multiple model families and tasks.
    • Unlike Muon, it shows little performance degradation when applied to AdamW-pretrained models.
    AbstractPDF
  46. 1158AI AgentsarXiv:2609.40285

    PivotOPD: Learning to Recover from Pivotal Mistakes in Multi-Turn Agents

    • Confirmed across three models (Qwen3 8B to 235B) that more than half of failed rollouts stem from a single early pivotal mistake.
    • Ran counterfactual experiments showing that correcting just the pivotal turn raises replayed success from 8% to 59%, and guiding only two turns after the mistake reaches 58%.
    • Combines preventive distillation (avoiding the mistake) and recovery distillation (recovering after it) with standard RL in a single PPO update.
    • Achieved +5.5% over the strongest baseline on ALFWorld with the Qwen3-1.7B student and ranked first on average among 13 baselines.
    • Transferred to another model family, raising the Nemotron-3.5 student's resolve rate on SWE-Bench Verified by +3.2%.
    AbstractPDF
  47. 1157ResearcharXiv:2609.40324

    Cogentic: Multi-Agent Orchestration for Automated Proof Discovery

    • Solved open problems that single-shot generation cannot handle via an iterative prove-verify loop.
    • Accumulated confirmed intermediate results in a permanent ledger so long-horizon work keeps its progress.
    • Produced new results on five open problems in online learning, auction theory, and mechanism design, all independently verified by domain experts.
    • Used adversarial verifiers that start by assuming the proof is wrong, catching flaws in drafts.
    • Runs on a relatively low inference budget of roughly O(100) to O(1000) model calls per problem.
    AbstractPDF
  48. 1156AI AgentsarXiv:2609.40169

    Learning from Research: Toward Lifelong Agent Harness Evolution

    • Went beyond fixing failures from execution traces by having the agent directly draw improvement ideas from research papers.
    • Split the harness into tool, context, skill, memory, and workflow modules, and ran paper-guided exploration per module.
    • Raised task goal completion on AppWorld Challenge from 49.6% to 63.6% and scenario goal completion from 28.3% to 44.8%.
    • Raised GPT-5.4-mini pass1 on tau2-Bench Telecom from 72.7% to 81.9%.
    • Built a lifelong evolution structure that repeats improvement whenever new papers arrive, using a genetic algorithm.
    AbstractPDF
  49. 1155AI AgentsarXiv:2609.40330

    Turbo Harness: Instance-Adaptive Harness Optimization

    • Built a tailored harness per task instance instead of one global harness, consistently improving performance.
    • Recycled trial-and-error artifacts discarded during prior optimization into a playbook used as training data.
    • Raised pass rate on SWE-smith-MR from 50.7% to 64.0% with Claude Haiku 4.5 and from 70.7% to 88.0% with Gemini 3.7 Flash.
    • A small 9B-class editor model (Qwen3.5-9B) sufficed, and it is invoked only once per task, keeping inference overhead low.
    • Recorded a higher pass rate on Terminal-Bench 2.1 than widely-used human-designed harnesses.
    AbstractPDF
  50. 1154RoboticsarXiv:2610.02161

    DuoMind: Enabling Distributed Multi-Robot Coordination with Semantic Communication

    • They proposed a hierarchical design in which each robot runs a VLM orchestrator (high-level reasoning) and a VLA action model (low-level control) in a distributed setup.
    • Robots exchange semantic messages containing intent and subgoals rather than raw data.
    • They built RoboPoly, a new benchmark for long-horizon multi-robot manipulation, filling an evaluation gap.
    • It consistently outperformed baselines on RoboPoly and RoboTwin, and ablations confirmed the separate contributions of the hierarchical design and semantic communication.
    AbstractPDF

Showing 50 of 1193