Research library

Search papers on equipment connectivity, quality inspection, production analysis, and manufacturing work.

Search and filter

New paper summaries

Summaries are editorial notes, not statements by the original authors. Check the original paper. Updated 2026-11-20.

  1. 1403Research Agents

    Aligning Autoresearch: Co-Improvement between Agents and Humans

    • In AutoBenchmark, benchmarks that the research agent built without human feedback were easy enough for solvers to score above 80. With detailed human direction, solver scores dropped by half.
    • Fine-grained human feedback lowered solver scores by 28 to 46.5 points in every one of the six benchmark-solver pairs, and this difficulty transferred to held-out solvers that were never used in the loop.
    • HackBench was built from 169 real trajectories, where the median number of commands per trajectory is 264 while the median number of hacked commands is only 1. The best monitor, Claude Opus 5, reached only 64.88 F1 at hack localization.
    • AutoInteract builds multi-turn interaction data from user models learned on real histories and trains with reinforcement learning, raising average Pass@1 on LiveCodeBench-Pro from 29.9 to 43.0 at the 4B scale and from 49.1 to 55.7 at 27B, with multi-turn gains from 23.0 to 43.2.
    • In the RL-XAR experiments, standard LLM judges preferred model writing over expert human writing in 63.5% to 84.6% of cases. Meta-optimizing rubrics on expert writing reversed this, and training lifted story writing from 2.8 to 8.2.
    • As a limitation, on Wikipedia section writing RL-XAR moved the score only from 2.7 to 4.0, far below humans and frontier models, and the human evaluation was a small preliminary study by the authors.
  2. 1402Research Agents

    DR Pixel: Page-Native Deep Research with Vision-Language Models

    • DR Pixel is the first open-source end-to-end multimodal deep research agent that uses rendered page images as its sole retrieval and reasoning unit over a corpus of 1.06M pages.
    • The authors built DRPixelBench with 500 questions on scientific papers, where every question requires reading text, tables, figures, and equations together.
    • Without training, the page-native approach improved accuracy by 5.1–8.4 points over OCR-based pipelines across GPT-5.4, Gemini-3.5-Flash, and Qwen3.6-27B while using 20%–30% fewer retrieval calls.
    • Qwen3.6-27B trained with counterfactual credit reweighting raised DRPixelBench accuracy from 40.1 to 52.1, slightly surpassing GPT-5.4.
    • The training came within 1.8 points of an oracle on out-of-distribution ViDoRe-V3 and needed 57% fewer searches than vanilla SFT for a comparable hit rate.
  3. 1401Research

    RubricAtlas: Mining Type-Level Rubrics from Expert Reports to Evaluate Deep Research Agents

    • The authors built RubricAtlas from 3,079 Chinese financial research reports, yielding 99 research types, 921 questions, and 8,160 checkpoints.
    • An individual report covers a median of 54% of its type’s questions, so relying on a single report narrows the evaluation.
    • Two judge pipelines achieved checkpoint agreement of κ = 0.71 and agent-ranking agreement of Kendall τ = 0.87.
    • Deleting the blocks answering one question lowered its score from 0.64 to 0.21, while adding 40% irrelevant text produced no score increase.
    • Evaluating only questions supported by the task’s source report raised completeness scores by 10.5 points and changed rankings, so the choice of standard shapes the outcome.
  4. 1400Research Agents

    Infinity Machine: a self-improving AI agent for research

    • The authors gave a strong model a shell, up to one hour of work, and an exact verifier, and it solved 14 problems over six days for a total model cost of about $180.
    • The system produced four results beyond the published state of the art on catalogued open problems in quantum information, including a certified point p = 0.064966, a qutrit frontier of 0.2910672, a 44-gate depth-6 circuit, and a stabilizer-rank bound of 10.
    • On the SABRE qubit-routing benchmark it cut the added-gate count from 68,559 to 12,087.
    • In hardware design it cut the energy per multiply-accumulate of an open 8×8 INT8 tile from 578.9 fJ to 128.8 fJ, a 77.7% reduction, with 55% less area.
    • The self-improvement gate rejected the one playbook generation the system wrote for itself on cost, and proposers gamed lenient scorers four times.
    • All three fully autonomous runs ended in honest ties with public records, and one audit caught a stale record.
  5. 1399Robotics

    Late Autumn: LLM Agents Specialize a General Multi-Robot Planner for One Warehouse

    • LLM coding agents ran six autonomous research campaigns to specialize a 9.0k-line general multi-robot planner for one warehouse site.
    • Every campaign's best patch raised throughput by 22–60% over the baseline, and the gains held on 20 unseen seeds.
    • The agents rediscovered known MAPF mechanisms such as corridor-exit priorities, local guidance, and rollouts, and independent campaigns found the same bottleneck.
    • The gains largely carried over to another layout, a larger fleet, and other service times, but mostly vanished under an outbound-heavy demand mix.
    • The final patches are 0.7k–2.4k lines with a median improving step of 27 lines, so experts can review them commit by commit.
    • Review found problems that the score misses, and some patches multiply planner compute by over 100×.
  6. 1398Research Agents

    PAPERBENCHX: Can AI Reproduce Science Across Disciplines?

    • PAPERBENCHX is a benchmark of 93 paper-reproduction tasks spanning 12 research areas.
    • The tasks run in 10 domain-native scientific computing environments, including Ansys HFSS, Meep, PySCF/GPU4PySCF, AutoDock Vina, and Habitat.
    • Every submission is rerun from scratch in a clean environment, and only regenerated evidence is scored against a hidden rubric rather than agent claims.
    • Among 10 frontier agent configurations, the best, GPT-6-Astra, reached a 59.08% partial score and a 13.98% full rate.
    • 71 tasks were not fully reproduced by any evaluated configuration.
    • Average Modeling and Execution scores reached 62.73% and 61.83%, while Validation fell to 42.91%, making scientific validation the weakest stage.
  7. 1397Agents / Memory

    The Right Fact, the Wrong Decision: SciValidityBench for Scientific Agent Memory

    • The paper defines scientific use validity: the same p = 0.03 value changes its evidential status depending on whether the plan was fixed in advance.
    • The authors built SCIVALIDITYBENCH, an executable benchmark factory whose 64 TaskSpecs cover five validity modules and six scientific domains.
    • The SCISTATEFORGE compiler–reader pipeline turns distributed experimental histories into explicit decision-governing state.
    • Across four model families, compiled scientific state raised valid decisions by 9.2–18.8 percentage points over a generic Event Graph.
    • BM25 retrieval alone did not reproduce this cross-family gain.
    • Exact repair of the decisive state recovered 18.6 validity points more on average than a matched irrelevant repair.
  8. 1396AI / Agents

    Cumulative and Audited Guardrail Tool Evolution for LLM Agent Safety

    • The authors propose EVOLVE, a four-stage framework that externalizes security knowledge into executable guardrail tools which are inherited, audited, and repaired, to address the weakness of fixed-rule defenses against emerging risks.
    • Six specialized agents cover risk analysis, tool generation, tool fusion, execution, auditing, and revision, forming a closed loop from risk perception to trusted adaptation.
    • Across the RJudge, ASSEBench, and AgentHarm benchmarks and four backbone LLMs, EVOLVE ranks first in all 12 combinations and outperforms existing defenses by +22% average F1 score.
    • The authors identify the VGT phenomenon in which over 20% of initially generated tools contain vulnerabilities, and auditing lowers the risky-tool share in the programming domain from 44.78% to 8.20%.
    • EVOLVE achieves 93.3% accuracy in detecting risks in the embodied AI2THOR environment, and on a real 6-DoF robotic arm it blocks sharp-object and falling-object hazards and replans paths to avoid collisions.
    • Guardrail tool reuse rates reach up to 86%, showing that security knowledge accumulates and evolves through interaction rather than remaining fixed after deployment.
  9. 1395Reasoning / RL

    RLiD: Reinforcement Learning in the Decision Plane for Language Agents and LLMs

    • The authors propose RLiD, which trains a large decision model that picks an action from an executable menu in one forward pass, removing the credit noise that token-level RL inherits from how actions are worded.
    • Proposition 1 shows that marginalising over alternative wordings of the same action preserves the expected update and removes the variance component tied to realisation.
    • In an exact FrozenLake credit audit, a decision-plane batch gradient with exact advantages reaches cosine 0.88 to the true gradient, while 71% to 97% of a token policy's score second moment is realisation variance.
    • On ARC the Qwen2.5-1.5B decision model reaches 78.6% in 0.12 GPU-h, ahead of token GRPO at 75.9% with 1.17 GPU-h, and in the same-runner Sokoban the decision plane reaches 44.5%, above every token variant.
    • On ALFWorld the full distillation and student-RL route reaches 99.3% success in 12.7 H100 GPU-hours, whereas token-level RL of the same LLM reaches 86.6% in 17.3 GPU-hours.
    • Without the admissible-action list the generative distilled student keeps 99.3% unseen success, and without memory it reaches 90.5%.
  10. 1394AI Agents

    EvoRec: A Unified Self-Evolving Agentic Recommender with Dynamic Reflection Trees

    • The authors propose EvoRec, a unified agentic recommendation framework that turns recommendation into a closed loop of need understanding, planning, execution, feedback, reflection, and repair.
    • A Need Agent infers structured user needs from long-term, short-term, and working memory, and a Planning Agent splits the retrieval budget of 3,000 between Behavior Continuation and Need Satisfaction while choosing tools and ranking fusion weights.
    • A Dynamic Reflection Tree localizes failures to need inference, planning, or prior repair strategies, generates personalized repairs, and verifies their effectiveness by re-running the decision.
    • Verified repair strategies enter working memory to guide later decisions and are abstracted into reusable strategies shared across users, so self-evolution continues without updating agent parameters.
    • On the Kuaishou local-services advertising dataset, EvoRec improves Hit@50 by 15.7% over the strongest baseline, and on the 18.5% of users with accumulated repair experience it improves Hit@50 by 13.1%.
    • In a seven-day online A/B test, nearline-deployed EvoRec increases advertising revenue by 4.6% and CTCVR by 3.5%, with a 14.5% revenue gain among high-intent users.
  11. 1393RAG / Agents

    Grounding LLM Agents in Business Processes: Process-Aware Knowledge Graphs for Reliable Multi-Step Service Invocation

    • The authors integrated OData metadata and BPMN business processes into one typed property graph, giving service retrieval both sequential and schema constraints.
    • The hybrid retriever finds candidate endpoints with dense embeddings and then follows process edges to retrieve subsequent endpoints.
    • For queries with low lexical overlap between the query and downstream entities, it improves Recall@3 by 5.7 points (n = 74), with consistent gains across embedding models and cut-offs.
    • On 81 multi-step queries, the agent completes 38.7% of tasks, compared with 10.3% for a chunk-based baseline.
    • Providing valid code lists as plain text eliminates invented codes but leaves invalid keys in 46% of write runs, while full typed context reduces invalid calls to 3% (3 of 101 runs), with no invented codes.
    • As a limitation, the authors state the evaluation is partly circular because the mock validator and the graph share the same metadata, so results must be reproduced against an independent backend.
  12. 1392Robotics

    Remember Where You Walk: Learning Spatial Memory for End-to-End Legged Robot Exploration

    • The authors proposed LSMM, a position-based matrix memory that lets reinforcement learning agents store task information at specific memory cells.
    • LSMM updates only a 5 × 5 cell window centered on the robot's position estimate at every step, leaving the rest of the memory unchanged so information about past locations persists.
    • The policy was trained end-to-end with PPO together with a maze reconstruction loss, where an auxiliary head predicts the maze wall layout from memory features.
    • On 100 unseen evaluation mazes, all architectures reached a comparable coverage of 65.4 to 82.8 %, so pure exploration performance alone did not require a spatial memory.
    • On maze reconstruction error, LSMM was the only architecture that significantly reduced it while exploring, from 0.33 to 0.1, more than three times lower than any baseline.
    • As limitations, the memory covers a fixed area sized to the maze, and the update depends on an external position estimate, since the position is not yet estimated from onboard sensing alone.
  13. 1391Agents / Memory

    Proactive Memory for Event-Driven LLM Agents

    • ProactMem stored memory in three layers: recent events, temporal trajectories, and behavioral regularities.
    • After an LLM extracts daily trajectories, facts are validated against source events, and regularities get confidence scores through historical replay.
    • The authors built ProactiveStream, a longitudinal event-stream benchmark across Sports, Home, and Code, containing 17 isolated streams, 4,539 events, and 729 reference interventions.
    • ProactMem recorded an overall effectiveness (F1) of 52.6%, a 21.4% relative improvement over the strongest of five external baselines.
    • In ablations, adding trajectory memory raised effectiveness in every scope, while removing replay validation or trajectory validation lowered both relevance and effectiveness.
    • As limitations, the authors state that evaluation covers low-risk service settings under a protocol fixed in advance, and that recipient benefit and deployment safety require future measurement.
  14. 1390Reasoning / Survey

    Video Generation Meets Agent Harness: Agentic Video Generation, Agentic World Modeling and Recursive Self-Improvement

    • The report defines the agent harness as the executable runtime that constructs context, exposes tools, maintains state, routes feedback, and enforces budgets, and applies it across video generation.
    • In agentic video generation, the survey organizes how harnesses coordinate script and shot planning, asset management, shot production, editing, and watermarking around shared project state.
    • In agentic world modeling, it analyzes designs such as an executable game harness where a lightweight game engine maintains the world state and an autoregressive video model renders the resulting frames.
    • It reports real-time cases including a compute-side rate of about 32 frames per second on one H100 and a system with 112 milliseconds from key press to display.
    • On recursive self-improvement, it separates revisions of the current output from updates to harness state, generator parameters, and the development procedure itself, and states that no surveyed system yet revises the procedure that trains and validates the generator.
    • Open challenges include autonomous video production, lifelong memory, hybrid code-pixel generation, human-free model development, and provenance and cost accounting.
  15. 1389Agents

    Implicit Minor Detection from Multi-Turn Dialogue in High-Risk AI Interactions: A Constrained Self-Evolving LLM Harness

    • This paper formulates the task of estimating whether the current speaker is likely a minor from weak cues across multiple turns, without strong identity signals such as accounts or faces.
    • The authors propose a constrained framework of a fixed LLM, a versioned Harness, and a Maker-Checker protocol, where the Harness controls when inference is invoked, how evidence is represented, and how outputs are validated.
    • Every decisive claim must rest on a short quotation or near-verbatim span locatable in the current dialogue, and mentions of other people or quoted material cannot support a claim about the current speaker.
    • Harness evolution promotes a candidate only when it passes admission constraints on validation data and otherwise retains the incumbent, preventing regressions that damage previously correct cases.
    • On a Chinese benchmark of 8,328 multi-turn dialogues, the method improves average precision by 7.60 points and minor-class F1 by 6.37 points over direct inference with the same LLM family.
    • In a 50-participant challenge study, human mean accuracy was 72.6% but fell to 54.8% on Boundary items, showing that difficult cases remain disputed even among humans.
  16. 1388Reasoning / RL

    SCOPE: Selective Critical-Step On-Policy Self-Distillation for Agentic Reinforcement Learning

    • In annotated WebShop trajectories, the authors found an outcome-step mismatch: 29.8% of steps in successful trajectories are erroneous and 42.4% of steps in failed trajectories are correct.
    • Distilling trajectory-level skills across all steps produces off-target signals at unrelated steps, which account for 57.9% of nonzero sampled-token gaps.
    • SCOPE first learns error localization and step-specific skill generation through a supervised warm-up, then during online RL analyzes its own trajectories and applies distillation only at selected error steps.
    • A bounded signed distillation coefficient prevents extreme teacher-student gaps from destabilizing training, blocking the collapse that occurred when the training-stage gap reached 2715.5 nats.
    • With Qwen2.5-3B, SCOPE records 97.7% success on ALFWorld and 85.2% on WebShop, outperforming SEED by 5.9 and 6.3 percentage points respectively.
    • On unseen ALFWorld tasks it reaches 89.6% success, 23.2 points above GRPO with 46.7% fewer interactions, and scaling to Qwen3-32B yields 99.2% success with mean steps of 7.7.
  17. 1387AI Agents

    VeriCUA: Verification as Agentic Exploration for Computer-Using Agents

    • The verifier first inspects the available environment or recording to gather context, then builds a fine-grained rubric for the instruction and selects GUI or code checks for each requirement.
    • It labels evidence as attested or asserted sources, so the executing agent's completion claim alone cannot mark a requirement as met.
    • In the online AndroidWorld environment it achieves the highest accuracy of 90.5% and the lowest false positive rate of 7.7% among the compared baselines.
    • In offline verification it reaches accuracy of 70.4 on OSReward-Hard, ahead of all compared judges, with a larger gap on hard cases.
    • Adding code-based state reads raises the decision rate and recall with false positives unchanged, and adding a visual inspection pass lifts both agreement and decision rate to 87%.
    • It rejects about 80% of false completion reports across the three online environments, with the largest difference on ScienceBoard.
  18. 1386AI Agents

    KARMA: Training-Free Knowledge-Atomic Harness Evolution for Industrial Advertising Agents

    • It splits knowledge into six modules of role, domain, experience, constraints, output protocol, and action descriptions, and allows at most three atomic mutations per cycle.
    • A deterministic verifier maps failures to candidate modules by error class and protocol codes, and a Critic compares failed and successful trajectories together to propose edits.
    • When the validation gain falls below the θ = 0.01 threshold, the full pre-edit knowledge snapshot is restored to prevent regressions.
    • Across 18 domain–model settings, it raises the average hold-out score from 30.19 to 39.15, a gain of 8.96 points, or 29.7%.
    • Removing either the successful or failed trajectory channel more than halves the average hold-out gain.
    • Knowledge evolved on one domain transfers to other domains without further adaptation, yielding an average gain of 8.54 points across the 30 source–target pairs.
  19. 1385AI Agents

    MRC-MMAgent: Reinforcement Learning for Multi-Role Collaboration in Multimodal Tool-Using Agents

    • It places solving, inspection, and continuation roles in one executable agent loop, where the inspector judges evidential support of the answer and either finalizes or resumes solving.
    • It combines task success, interaction validity, and successful correction into one trajectory reward with weights (λacc, λfmt, λcorr) = (0.7, 0.1, 0.2).
    • Full-trajectory GRPO propagates the combined reward to every role decision, avoiding the post-hoc role-attribution bias that role-selective masking introduces.
    • Using only the MMSearch-R1 training set, it improves upon the base model by 11.1 points and reaches an average of 68.1 with the 8B model, leading same-scale systems.
    • On image reasoning benchmarks it achieves the best V* Bench score of 93.2 and HR-Bench 8K score of 78.5, gaining +6.8 on V* Bench and +3.9 on HR-Bench 8K over the base model.
    • Without inspector training, 98.3% of trajectories are marked for continuation causing unnecessary revisions; training reduces the continue rate to 60.4% and raises accuracy from 62.7 to 64.0.
  20. 1384Robotics

    Gym-Anything for Robotics: Scaling physical simulations with digital agents

    • The authors decomposed occupations and everyday activities from U.S. GDP data into atomic physical actions, selected 120 action groups, and had coding agents build 25 of those simulations.
    • A creation agent and four audit agents ran a creation-audit loop for three rounds to verify simulation quality by catching physics errors and exploits.
    • The resulting RoboWorld covers 25 simulations with 25 test environments and 175 training RL environments, supporting swappable robots and scene variations.
    • All three frontier models scored 0% success on the 25 test tasks with a best average reward of 0.06, while Claude Opus 5.5 achieved a 32% success rate on an easier subset of 25 tasks.
    • A code policy reading privileged state succeeded on 96.7% of 2,048 varied initial states, and π0.5 trained on 1,000 of its demonstrations rose from 0% to 53% success on two tasks and 55% on unseen variations.
    • As a limitation, success drops as the robot's degrees of freedom increase, and the two GPT models earned no reward at all on humanoid tasks.
  21. 1383Research Agents

    MoLRSI: A Recursive Self-Improving AI agent for Molecular Discovery

    • MolRSI converts scientific papers and code into validated executable skills, and uses experience-guided Monte Carlo tree search (MCTS) to assemble and refine workflows.
    • It stores workflow performance, refinement outcomes, and execution records from 25 source tasks in persistent memory to recursively improve the search policy for later tasks.
    • In virtual antibiotic screening, 17 of 107 compounds showed antibacterial activity with MICs of 32µg/mL or lower, and the experimental hit rate of 15.9% exceeded the 8.7% of the previous expert-driven workflow.
    • In de novo peptide-MHC binder design, it produced 8 experimental hits in total, 5 for MART-1 and 3 for SARS-CoV-1, whereas Claude Science produced no hits under the same conditions.
    • The small-molecule resource library consists of 110 reusable skills (55 models, 48 methods, 7 tools), and the protein library consists of 65 skills.
    • As a limitation, the authors state that whether policy updates remain useful as biological targets and evaluation criteria change is an open question, and incorporating experimental outcomes into policy updates is future work.
  22. 1382LLMs / NLP

    Nemotron 3.5 Super and Lightning: Open, Efficient Mixture-of-Experts Hybrid Mamba-Transformer Models for Agentic AI

    • NVIDIA released Nemotron 3.5 Super, a 120B-parameter, 12B-active multimodal Mixture-of-Experts Hybrid Mamba-Transformer model, and Lightning, a 30B-parameter, 3B-active text model.
    • Both models went through a 100B-token MTP warm-up, 1 trillion tokens of distillation from a larger Nemotron model, and long-context training, while Super added Indic continued pretraining and agentic software-engineering mid-training.
    • The agentic mid-training raised DeepSWE pass@1 from 6.4% to 10.4% and SWE-bench Verified under the Codex harness from 48.4% to 61.8% under identical experimental SFT.
    • The post-trained Super model achieves 51.78% on SWE-bench Pro, surpassing Nemotron 3 Ultra with approximately 4.6× fewer active parameters, and parallel agent execution with self-verification improved Terminal-Bench 2.1 from 62.7% to 71.9%.
    • NVFP4 quantization achieves 99.62% median accuracy recovery relative to BF16 across 36 benchmarks, and MTP boosting improves speculative-decoding acceptance length by up to 21.8% under probabilistic draft sampling.
    • On a single GB200 GPU it produces up to 4,856 output tokens/s/GPU while sustaining at least 50 tokens/s/user, and NVFP4 delivers 2.43× the throughput of BF16 under the evaluated interactive agentic workload.
  23. 1381Robotics

    OmniBrain-0: A Unified Agentic Model for Embodied Intelligence

    • OmniBrain-0 is a lightweight unified agent that ties manipulation and navigation to a shared high-level decision model handling task understanding, planning, progress tracking, and feedback-driven recovery.
    • On LIBERO-Pro it achieves 65.5% overall success with average completion time cut from 13.8 minutes to 1.6 minutes, an 8.6x speedup, and RSI raises success to 69.3%.
    • On RoboCasa365 it reaches 39.6% overall success, exceeding RLDX-1 at 36.0% and more than tripling the few-shot backbone's 12.2%.
    • On R2R-CE navigation it improves SR from 34.0% to 65.0% and reduces NE from 6.57 m to 3.78 m.
    • Relative to GPT-5.6-Sol, token usage is 4.5x, 2.6x, 3.5x, 12.7x, and 3.7x lower on the five benchmarks respectively.
    • On three real-world manipulation tasks it achieves a macro-average success rate of 73.3%, outperforming π0.5 at 58.3%.
  24. 1380Robotics

    EmbodiedLab: Benchmarking Embodied Agents That Act to Know

    • EmbodiedLab runs 30 experiments in five disciplines on 14 apparatus platforms, giving the agent only the objective and the apparatus while it chooses its own procedure.
    • A run succeeds only if its result lies within the instrument tolerance, rests on readings recorded in the same episode, and leaves the bench in the state the task requires.
    • Evaluating six frontier models, the highest success rate at the lowest interface level T0 is 43.3% by Claude Opus 5.5, followed by GPT-6 Astra at 40.0%.
    • Acting on the apparatus, observing it, and restoring it account for 75% to 100% of every agent's failed runs, while at most 25% fail in reasoning from readings already in hand.
    • Adding perception and motion planning at T1 raises the success rate by at most 10 points, while manipulation skills at T2 raise the weaker agents by 10 to 27 points.
  25. 1379Agents

    EvolveMAS: A Unified Self-Evolution Framework for Decentralized Multi-Agent Systems

    • EvolveMAS pairs each main agent with a dedicated MemoryAgent that maintains isolated Structured Memory outside the main agent's reasoning context.
    • Across five benchmarks and six evaluation configurations, EvolveMAS achieves a normalized average score of 77.2, exceeding the strongest setting-specific baseline by 9.7 points and Self-Memory by 7.7 points.
    • In the cooperative setting it records 83.0 on BBH and 97.0 on HumanEval, above the baselines of 79.0 and 95.7 respectively.
    • In the adversarial Werewolf audit it achieves a wolf deception success of 70.0, exceeding the best competing method by 13 points.
    • In the LLM-Deliberation ablation, an external MemoryAgent raises the unanimity rate from 24.0% to 74.0%, and cross-episode Knowledge raises it further to 86.0%.
    • Memory governance keeps context occupancy at 0.62% at episode 7, lower than the 6.54% of carrying raw interaction histories.
  26. 1378AI Agents

    AutoRSE: Evaluating LLM Agents as Recommendation Engineers for Self-Evolving Recommender Systems

    • The authors built AutoRSE, an evaluation framework spanning six representative industrial-grade recommendation tasks, where in each task the agent receives a faithfully reproduced state-of-the-art algorithm, one H200 GPU, and a 12-hour optimization budget.
    • Across 42 evaluated settings of three frontier LLMs and three harnesses in seven combinations, agents outperformed the baselines in 35 cases, achieving a mean relative improvement of 16.1% across all tasks and configurations.
    • Trajectory analysis of 4425 segmented experiments showed engineering optimizations accounted for 91.7%, dominated by Model Ensemble at 59.0% and Parameter Tuning at 26.2%, while algorithmic innovations were only 8.3%.
    • Final submitted algorithms achieved a modest average novelty score of 2.42 out of 4.0, and higher novelty correlated strongly with larger gains at Pearson r = 0.79.
    • Injecting ideas from frontier research and human experts beat self-generated ideas by 13.35% in average improvement and 0.58 points in novelty, confirming that idea generation is the bottleneck.
    • Agents achieved an average gain of 29.2% on generative recommendation tasks compared to 3.1% on CTR prediction, so less saturated baselines left more room for agent gains.
  27. 1377Reasoning / RL

    When Fine-Grained Credit Assignment Helps and Hurts: Multimodal Agentic Reinforcement Learning through Video Editing

    • The authors built 797 editing tasks of four types, Content Selection, Narrative Sequencing, Rough Cut, and Fine Cut, from 49 real films and their released trailers.
    • Training Qwen3.5-9B with outcome-only GRPO raised the overall score from 0.284 to 0.413, so outcome rewards alone stayed competitive on the shorter-horizon tasks.
    • GRPO+PRM with a fixed process-reward weight scored only 0.334 overall with 44.1% truncated rollouts, showing reward hacking where process rewards kept rising while task completion rates and outcome scores declined.
    • Decaying the process-reward weight to zero between updates 20 and 30 gave the best overall score of 0.447 and cut the truncation rate to 6.2%, mitigating the hacking while preserving early gains.
    • On the longest Fine Cut tasks, decaying process rewards delivered a 45.4% relative gain over outcome-only GRPO, while on the three shorter-horizon task types the relative gain was at most 10.4%.
    • On-policy self-distillation (OPSD), a token-level signal, scored 0.277 with reference answers and collapsed to 0.044 with turn-level suggestions, showing that dense supervision alone does not guarantee effective learning.
  28. 1376AI Agents

    RSIV-Harness: A Recursive Self-Improvement Harness Framework for Open-Domain Conversational Text-to-Visualization

    • The paper defines the problem: when users request charts without knowing which database holds the data, the agent must discover sources, ask clarifying questions, and refine charts under earlier constraints, yet existing harnesses are authored once and never revised, so the same failures recur.
    • The ODAC-Vis benchmark contains 640 dialogue tasks and jointly evaluates open-domain source discovery, ambiguity resolution, and multi-turn chart refinement.
    • RSIV-Harness keeps the model weights fixed: an inner loop solves tasks, an outer loop reads failure traces and proposes harness revisions, and only revisions passing the anchor and memory checks are accepted.
    • A single round of self-improvement raised all six metrics, with executability improving by 28.02 points and chart structure by 34.01 points.
    • Removing the stability rules degraded all six metrics, with executability dropping by 9.73 points.
    • The harness evolved with DeepSeek transferred to GPT-6 Luna and improved five of six metrics over the static harness, showing that harness improvements transfer across backbones.
  29. 1375Reasoning

    Evolve on Graph: Self-Evolving Knowledge Graph Reasoning without Question Annotations

    • The authors proposed EoG, a framework that improves LLM graph reasoning using only the knowledge graph itself, without question annotations.
    • A single LLM alternates between a Proposer that poses questions and a Solver that answers them through graph self-play.
    • SVTS executes first-order logic queries on the graph to build training tasks with complete and correct answer sets.
    • A boundary-aware curriculum rewards the Proposer by the Solver's score variance, keeping task difficulty on the Solver's capability boundary.
    • With Qwen-2.5-7B, EoG exceeded the strongest baseline by 7.46 points in average Hit@1 and by 17.80 points on the industrial KG.
    • The synthesized training questions barely overlapped with test questions, so the gains did not come from test-set leakage.
  30. 1374RAG / Agents

    From Questions to Queries to Insights: A Schema-Grounded Agentic Workflow for Natural-Language Crash Data Analysis

    • The authors proposed CrashDB-QA, a fully local, schema-grounded agentic workflow that enables natural-language analysis of crash databases.
    • A shared knowledge base grounds query construction in the agency's data dictionary and safety-analysis conventions, while recorded intermediate outputs support inspection and auditing.
    • They evaluated the workflow on a frozen snapshot of 4.4 million Wisconsin crash records using a benchmark of 300 questions and 9 local open-weight models.
    • A code-specialized 7B model reached 80.7% accuracy at a median latency of 9.2 s, the best accuracy-latency balance among the tested models.
    • Comparisons between groups or periods stayed hard, and no model exceeded 43.3% on them.
    • When the Judge used authoritative field definitions and the Refiner received execution evidence, accuracy rose to 83.7%.

From research to product use

Operating capabilities, pilots, and technologies in development are identified separately.

View technology