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.
- 1218Research Agents
MLE-STAR: Machine Learning Engineering Agent via Search and Targeted Refinement
- Problem: Existing machine learning engineering (MLE) agents lean on what the LLM already knows, so they drift toward familiar methods such as scikit-learn and, for example, propose logistic regression even for a text classification competition. They also rewrite the whole code structure in each iteration, so they move on to other steps before exploring one component such as feature engineering in depth.
- Input and output: This is an agent system for Kaggle-style ML tasks. The input is a task description (task type, data modality, score function) plus datasets that may span several files, and the output is a single Python script that trains a model and prints its hold-out validation score. At the end a separate agent adds test-data loading and submission file creation to that script.
- Method: A retriever agent first uses Google Search to find task-specific models with example code, writes a simple script for each, and merges them in order of validation score into an initial solution. In each outer loop an ablation-study script picks the code block with the largest effect on performance, and in the inner loop a planner proposes several block-only refinements using earlier scores and keeps the best one. Solutions built in parallel are merged with ensemble strategies that the agent proposes and refines itself, while a debugging agent, a data leakage checker and a data usage checker fix runtime errors, test-information leaks and unused data.
- Implementation: The default LLM is Gemini-2.0-Flash, with extra runs on Gemini-2.5-Pro and Claude-Sonnet-4; the setup retrieves 4 models, runs 4 inner loops and 4 outer loops, builds 2 parallel solutions, explores ensembles for 5 rounds, and caps each task at 24 hours. On 96 vCPUs, 360 GB of memory and 8 V100 GPUs with 16 GB each, one solution took 14.1 hours on average, against 15.4 hours for AIDE. The appendix states that example outputs such as the retriever output, the ablation code and the final solution code are placed in the authors' GitHub repository.
- Results: On the 22 MLE-bench Lite competitions with 3 seeds, MLE-STAR with Gemini-2.5-Pro won a medal in 63.6% of competitions, with gold in 36.4%. With the same Gemini-2.0-Flash, the medal rate rose from 25.8% for AIDE to 43.9% and the above-median rate from 39.4% to 63.6%. Under the same setting the medal rate was 37.9% without ensembling and 43.9% with the proposed ensemble, whose gold rate of 30.3% beat the 25.8% of a simple average ensemble.
- Limits: Kaggle competitions are public, so the LLM may have been trained on related discussions; as evidence against this the authors report that a Gemini-2.5-Pro judge rated all 75 pairs formed from 25 top discussions in 7 competitions and the final solutions as novel. Evaluation uses the 22-competition Lite split, the low-complexity part of the 75-competition MLE-bench, and the Claude-Sonnet-4 runs and the DS-Agent comparison each cover only 4 tasks. The AIDE results with o1-preview and GPT-4o are taken from the MLE-bench repository and use 16 and 36 seeds, against 3 for MLE-STAR.
- 1217Research Agents
MOOSE-Chem3: Toward Experiment-Guided Hypothesis Ranking via Simulated Experimental Feedback
- Problem: Wet-lab experiments in the natural sciences are costly and low-throughput, so choosing which hypothesis to test first matters, yet existing hypothesis ranking relies only on LLM internal reasoning without experimental outcomes. Ranking strategies that use earlier results need repeated real experiments, which made them hard to develop and evaluate.
- Input and output: This is a method paper that introduces a new task, a simulator and a policy. The policy receives a research question and 64 candidate hypotheses (1 ground truth and 63 generated by MOOSE-Chem), tests one hypothesis at a time, and is scored by the number of trials until it finds the ground truth. The simulator takes the research question, a candidate hypothesis and the ground-truth hypothesis hidden from the policy, and returns a performance score between 0 and 1.
- Method: CSX-Sim uses an LLM to break the ground-truth and candidate hypotheses into key functional components and mechanisms, sets a weight for each component, marks critical components, and scores a candidate by the weighted sum of per-component mechanism similarity; a candidate missing any critical component scores 0. CSX-Rank is an ICRL (in-context reinforcement learning) policy that learns inside the context of a frozen LLM. It classifies components as effective, uncertain or ineffective, drops the ineffective ones, clusters the rest by function, tests a hypothesis from a promising cluster and adds the result analysis to a running summary.
- Implementation: All experiments use GPT-4o-mini. For simulator validation the authors collected 30 research questions and 124 hypotheses with experimentally reported outcomes (3 to 6 per question) from the literature, and each simulated result is the average of three runs. Ranking is evaluated on the 51 problems of TOMATO-chem, and the official repository holds Data, Method and Simulator folders with scripts for per-problem ranking, ablations and simulator validation.
- Results: In simulator validation, CSX-Sim reached a Spearman correlation of 0.960, perfect agreement on 26/30 questions and an RMSE of 0.213, better than the Matched Score baseline (0.843, 12/30, 0.232). The average number of trials to find the ground truth was 15.196 for CSX-Rank, fewer than 32.500 for random search, 28.608 for pre-experiment ranking and 27.980 without clustering. Under the most complex noise, CSX-Rank needed 32.706 trials against 36.471 and 40.451 for its ablations, and with 128 candidates it needed 30.7 trials.
- Limits: Simulator scores deviate from real experimental outcomes, and the authors note that the assumption (A1) that the ground truth is the best hypothesis in its neighborhood is mostly valid, though a better hypothesis may exist there. All ranking evaluations use simulator feedback, and more complex noise increased the number of trials needed. The authors also point to possible biases from the literature-based datasets and the simulator's assumptions.
- 1216AI Agents
AI Agents vs. Agentic AI: A Conceptual Taxonomy, Applications and Challenges
- Problem: The authors argue that existing surveys fold AI Agents and Agentic AI into one umbrella of intelligent agents, which blurs the concepts and misaligns system design. Without a clear line, developers under-engineer tasks that need agent coordination and over-engineer tasks a single agent could solve.
- Scope: This is a literature review (survey). The authors searched twelve platforms, academic databases such as Google Scholar, IEEE Xplore and arXiv plus AI tools such as ChatGPT, Perplexity.ai and Grok, with terms like AI Agents and Agentic AI. Papers were included for novelty, empirical evaluation, architectural contribution and citation impact.
- Taxonomy: An AI Agent is a single system that uses LLMs and LIMs (Large Image Models such as CLIP) and calls tools for a narrow task, defined by autonomy, task specificity and reactivity. Agentic AI is a system of specialized agents that split a goal, communicate, and are coordinated through persistent memory and an orchestrator (meta-agent). The comparison tables (Tables 1 to 9) contrast the two types against Generative AI as the baseline and add Generative Agents as an inferred fourth type.
- Applications: For AI Agents the review covers customer support with enterprise search, email triage, personalized recommendation with basic reporting, and scheduling assistants. For Agentic AI it covers multi-agent research assistants, robotics coordination, collaborative medical decision support, and game AI with adaptive workflow automation. Tables 10 and 11 list representative AI Agent and Agentic AI systems from 2023 to 2025 by application area and mode of operation.
- Challenges and remedies: The review names four main limits of AI Agents: missing causal reasoning, hallucination and shallow reasoning inherited from LLMs, incomplete agentic properties such as autonomy and proactivity, and failures in long-horizon planning and recovery. For Agentic AI it names eight bottlenecks, including inter-agent error cascades, coordination breakdowns, emergent instability, scalability limits and poor explainability. As remedies it proposes ten design strategies such as RAG, function calling, ReAct loops, memory architectures, role-based orchestration and causal modeling.
- Limits: The authors themselves cite critiques that agent benchmarks rely on artificial environments and that Agentic AI work favors architectural novelty over rigorous empirical validation. The review rests on literature and conceptual comparison, and its detailed application examples are hypothetical scenarios. For medical decision support, the authors state that real-world implementations are still lacking.
- 1215Agents / Memory
Librarian Agents: Training LLMs to Curate Long-Context Workspaces that Anticipate Future Queries
- The authors built the Librarian Agents framework, in which a builder agent transforms raw documents into a hierarchical filesystem of synthetic files and reference links.
- They formalize filesystem workspace construction as an MDP and optimize the builder policy directly with reinforcement learning to raise answering accuracy.
- On BrowseComp-Plus it outperforms the strongest baseline by 33 percentage points on average and reaches 0.77 accuracy with the Gemini 3.1 Flash Lite answering agent.
- On the unseen BRIGHT domains it scores pass@1 52% and pass@3 67% zero-shot.
- Construction cost is amortized across repeated queries, with accuracy of 0.64 versus 0.24 for the base builder on reuse.
- As a limitation, compression may discard details needed for needle-in-a-haystack tasks, so the system keeps reference links for fallback access.
- 1214Reasoning / RL
X-RIC: An Agentic RAN-Intelligent Controller Benchmark for Enhancing Reinforcement Learning
- The authors propose eXplainable RIC (X-RIC), a framework that integrates agentic AI into O-RAN architectures.
- The BARX evaluation framework assesses agentic reasoning candidates in closed-loop control across four dimensions: latency, compliance, determinism, and policy impact.
- The Telecom Language Model (TLM) augments LLaMA3-8B with a trainable RAN-state encoder that maps network telemetry into 256-dimensional embeddings.
- On SLA-aware network slicing it achieves 89% operational cost reduction, converging 75% faster with 50% fewer training steps and 46% lower variance.
- The non-fine-tuned LLM increases cost by 90.3% due to hallucinations that violate resource constraints.
- On the OpenAirInterface over-the-air testbed, X-RIC500 maintains margins of 17 Mbps for eMBB and 5 ms for URLLC.
- 1213Robotics
CarryOn: A Dual-Loop Harness for Personalized and Reliable Robot Assistance
- The authors built CarryOn, a dual-loop harness that augments a frozen VLA policy.
- The planning loop maintains user-level and scene-level context and revises it through dialogue and observation to dispatch personalized subtasks.
- The execution loop filters actions against predicted hazardous contact as a proactive shield.
- After failure, reactive recovery skills restore task progress, and offline reflection refines the recovery skills.
- Across 900 LIBERO-Safety episodes it reduces the constraint-violation rate from 52.6% to 6.1%.
- Under a 12 cm object displacement, the pure VLA completes 26/200 episodes while reuse of refined skills raises this to 160/200.
- 1212Research Agents
MolHop: Scaffold-Hopping Molecular Design with Agentic Reinforcement Learning
- The authors extracted approximately 1.6M scaffolds from the Enamine commercial molecular library and organized them into a coarse-to-fine hierarchical action space from H1 to H7.
- The agent chooses scaffold navigation, molecular editing, and undo within a single harness, while the environment maintains trajectory memory as a structured state to support long-horizon search.
- Without training, connecting Grok-4.6 to MolHop raised drug-task joint@H1 from 41.0% to 64.0% and the commercial-compound task from 36.5% to 69.0%.
- MolHop-9B, trained with SFT followed by GRPO, achieved joint@H1 of 55.5% on the drug dataset and 73.5% on the commercial dataset, beating GPT-5.6-sol, Claude-5-Opus, and Gemini-3.1-Pro.
- Candidates from MolHop-9B showed scaffold Tanimoto medians of 0.123 and 0.115, indicating larger structural differences than baseline models.
- Removing the scaffold library or trajectory memory caused substantial performance drops, confirming that the hierarchical search space and memory are key to long-horizon search.
- 1211RAG / Agents
Retrieve What Helps, Only When It Helps: Agentic Reinforcement Learning for Selective Retrieval-Augmented Machine Translation
- The model learns a policy that chooses reasoning, retrieval query generation, and translation within a single trajectory.
- It is trained with GRPO using final translation quality as reward, so the model itself decides when to retrieve for each input.
- To align the retriever with translation quality, the authors built joint Plackett–Luce optimization and translation-quality contrastive training.
- On IdiomKB Zh→En, SRAT achieved I5 of 0.731 and MQM of -2.74, beating the strongest baselines.
- On RAGtrans, contrastive training raised retrieval accuracy from 56.3% to 70.7% on En→Zh and from 72.6% to 81.4% on En→De.
- With a frozen black-box Claude Sonnet 4.6 translator, SRAT-Claude improved RAGtrans En→De MQM by 16.8% over direct translation.
- 1210Memory / Agents
Seeing, Maintaining, and Learning: The Evolution of Multimodal Agent Memory
- The survey organizes memory along two axes: Axis I covers what representation retained experience takes, and Axis II covers how memory state and management policy change during operation.
- The management axis has three tiers: T1 reuses fixed memory, T2 updates memory through a stable policy, and T3 evolves the management policy itself using feedback.
- Analysis of 297 architectures showed that T2 Dynamic Memory is the most common with 204 architectures (68.7%), while T3 Self-Evolving Memory accounts for only 17 (5.7%).
- Hybrid representations form the largest category with 139 architectures (46.8%), and designs keeping explicit records together with source evidence are widely used.
- Of the 17 T3 Self-Evolving Memory architectures, 13 are cross-represented systems, and no memory-policy evolution case exists with only latent or only parametric representations.
- Across 82 benchmarks, cross-modal and long-range evidence are well evaluated, but only 7 and 10 benchmarks directly evaluate conflicting-memory handling and unanswerable queries, respectively.
- 1209AI / Agents
Autonomous Science Bench: Evaluating AI Agents for Autonomous Scientific Discovery
- Autonomous Science Bench is a science-agent benchmark comprising six budgeted design-build-test-learn campaigns grounded in published experiments across biology, chemistry, and materials science.
- It evaluates an active experimentation loop where agents select design batches, receive measurements, and plan subsequent experiments before delivering a predictive model.
- In a pilot of eight agent-model configurations and 144 trials, the best configuration succeeded in 67% of trials and the hardest campaign was solved in only three of 24 attempts.
- Budgets cover at most 11% of each candidate pool, so covering the pool evenly is not enough to pass.
- Opus 5.5 in Claude Code resolved 67% of campaign trials, with GPT-5.6 Terra and GPT-5.6 Luna following at 44% and 28%.
- 1208Agents / Memory
Can Agents Forget When Asked? A Benchmark for Memory Unlearning in LLM Agents
- Bemula is the first benchmark dedicated to memory unlearning in LLM agents, the ability to delete exactly what the user requests.
- Its tasks jointly measure removal of requested facts and preservation of the rest of memory.
- Across all evaluated systems and backbones, no system passed more than 1.5% of forget requests end to end, and most passed none at all.
- A human validation of the LLM judge on 100 sampled items yielded 98% agreement, supporting evaluation reliability.
- 1207Robotics
Agentic Robot Creation through Generation, Verification, and Revision
- RoboForge is an agentic framework covering both embodiment and behavior creation, chaining motion generation, policy learning, and execution-grounded revision.
- Motion stays in the loop: rollout feedback can revise the reference motion, the reward, or both.
- Computational metrics exposed recurring mechanical failures in generated designs before fabrication.
- Against Eureka-style reward-only refinement it improved the overall rating from 2.10 to 3.97, showing the value of structured motion guidance.
- The team fabricated an agent-generated Gorilla robot and directly deployed learned behaviors on the physical machine.
- 1206Research
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.
- 1205Research Agents
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.
- 1204Robotics
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.
- 1203Robotics
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.
- 1202AI / Agents
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%.
- 1201AI / Agents
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.
- 1200Machine Learning
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.
- 1199AI / Agents
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.
- 1198Software Engineering
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%.
- 1197Research Agents
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.
- 1196Agents
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.
- 1195Agents
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.
- 1194Agents
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.
- 1193Computer Vision
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.
- 1192LLMs / NLP
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.
- 1191Research Agents
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.
- 1190Human-Computer Interaction
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.
DOI:10.1007/s10055-026-01497-x
- 1189Machine Learning / Security
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.
DOI:10.1186/s42400-026-00679-z
From research to product use
Operating capabilities, pilots, and technologies in development are identified separately.
View technology