News without the noise
New AI papers from arXiv, picked daily and explained with their claims, benchmark results and code.
2566 articles · page 40 of 54 · The Research front page
Jiang et al. developed RuleEvolve, a framework that automatically refines the instruction files used to guide AI coding agents. Across two agent frameworks, four backbone language models, and three coding benchmarks, the evolved rules matched or improved functional correctness while often producing shorter code and using fewer generation resources than manual rules and prompt-optimization baselines.
Park et al. treat scenario selection as part of agent harness optimization rather than as a fixed input to it. Their system, ActiveSaddler, adaptively revisits unresolved failure types while searching for new weaknesses, raising test Pass@1 by 4.4 percentage points on GAIA2 and 7.5 points on Terminal-Bench 2.0 relative to the same optimizer using a predetermined scenario order.
Tiffany et al. present Sapien, a policy engine that generates a task-specific security policy before an AI agent begins work, then checks each tool call against that policy. On two agent-security benchmarks, it ruled out 62 to 95 percent of attacks under a fully hijacked-agent threat model, while keeping task utility within a few percentage points of an unconstrained agent.
Paliskara, Namchittai and Lampinen test whether AI agents can coordinate when they represent different users competing over a shared resource. Across five frontier models and 77 scenarios, multi-agent teams underperformed a single coordinating agent in every environment, even when teammates could communicate.
Du et al. introduce ComputerSD, a training method that learns from feedback generated after each action a computer-use agent takes. On OSWorld-Verified, it improved task success over outcome-only reinforcement learning by 1.9 percentage points for Qwen3-VL-8B-Thinking and 4.1 points for EvoCUA-8B.
Zhang et al. test whether a fixed language model can verify its own long-horizon work without reference answers, grading rubrics or a stronger judge. Across five workspace benchmarks and two frontier models, their harness outperformed the evaluated selection baselines, while evidence-backed revision improved average scores over a single rollout by more than six points.
The authors formalize the interaction between an LLM reporting factor importances and a budget-limited auditor as a verification game. They show that report-dependent auditing creates a suppression incentive, but a mixed rule with a small report-independent floor makes truthful reporting optimal. Results are instantiated with the Counterfactual Brier Score and validated on four NLP benchmarks.
The authors present a framework for generating multi-step GUI interaction data without installing or executing the software being depicted. Fine-tuning Qwen3.5-35B-A3B on the resulting synthetic trajectories raised its OSWorld mean task score from 33.0% to 40.8% and more than doubled its ScienceBoard success rate.
The authors propose RPTune, a framework that learns how to curate product catalogs for long-context LLMs and fine-tunes the LLM to better use the curated input. Evaluated on 7 real merchants, context curation alone improved search accuracy by up to 31.4 percentage points, and post-training added a further 10.3 points on average.
Nisioti and colleagues test whether neuroevolution, which optimizes populations of neural networks through weight perturbation and selection, can address the tension between learning new tasks and retaining old ones. They report that evolution strategies produced the most consistent stability-plasticity balance, while genetic algorithms adapted faster but forgot more.
The authors compare 14 base and post-trained model pairs on three interactive agent benchmarks, using a lightweight harness to let base language models call tools and act across multiple turns. Post-training generally raises pass@1, the chance of succeeding once, but often reduces pass@K, the chance that at least one of several attempts succeeds.
Geng et al. present an inference-time decoding method that separates high-level idea selection from token-level generation. Across chat, creative writing, image-generation planning and a protein-design pilot, the authors report substantially greater diversity at comparable quality than existing sampling and prompting methods.
Li and colleagues address prompt injection in tool-using agents, where poisoned webpages, tool descriptions, memories or reusable skills can induce consequential actions. Their PACE system mediates calls immediately before execution, blocking effects that cannot be justified by the authenticated request while largely preserving normal task performance.
Li et al. address a weakness in language-model alignment: harmful goals that are refused when stated directly can succeed when distributed across a conversation. They derive conditions connecting single-turn safety training to multi-turn risk, then introduce TRACE, an objective that assigns safety credit across response tokens and selectively erases unsafe behavior.
Finley and colleagues introduce DAYJOB, a benchmark of 130 long-horizon assignments designed by professionals in healthcare and finance. Across 30 model configurations, even the strongest system passed fewer than one-quarter of attempts, while typical configurations rarely produced deliverables satisfying every required criterion.
Luo and colleagues study dynamic task routing, where an agent must identify which version of a document applied at a query time and then answer the associated problem. Across 373,428 queries derived from six established benchmarks, their Revision-Aware Independent Agent Graph achieved higher joint routing-and-answer accuracy than comparison methods while using substantially fewer model calls.
Deng et al. present DeFA, a framework for diagnosing failures whose causes and visible consequences occur many steps apart. Rather than judging each step independently, DeFA maps dependencies across an execution and traces how suspicious events influence later actions. It outperformed seven baselines on text benchmarks and produced feedback that improved a downstream skill-learning pipeline.
Zhu and colleagues address a weakness in agentic reinforcement learning: a final success or failure score does not reveal which action caused the outcome, and provides no comparative signal when every rollout receives the same result. Their FAULT method diagnoses intermediate errors, estimates their relative costs from observed outcomes, and redistributes terminal credit across individual steps.
Wiedmann and colleagues evaluate coding agents as configurable systems rather than treating the underlying language model as the sole unit of analysis. They find that task information has a larger effect than either model size or time budget, while repeated runs of identical configurations still produce substantial variation.
Zhang et al. test whether tool-calling agents limit themselves to the private data explicitly needed for a user’s request. In a controlled 720-episode benchmark, every evaluated model exceeded the reference scope, while a zero-shot filtering method substantially reduced unnecessary calls.
Peng et al. identify a mismatch in joint video and action models: training perturbs recorded actions while leaving the corresponding video tied to what originally happened. CtrlWAM instead executes perturbed actions in a simulator and trains on the resulting counterfactual visuals, producing futures that better reflect predicted or commanded behavior.
The authors present FaV-A, a multimodal agent that generates positive-negative space images through three connected stages. It first creates a base object, then analyzes its contours to propose secondary semantics for the negative space, and finally issues compositional instructions for the image generation model. In user studies and ablations, FaV-A produced more visually coherent and semantically aligned compositions than zero-shot MLLM baselines.
Han et al. introduce VISTA, a model-agnostic harness designed to extend multimodal agents beyond their immediate visual context. Rather than compressing each observation once and relying on that representation throughout a task, VISTA stores past images in their original form and lets the model actively retrieve and arrange them as reasoning proceeds. With Claude Opus 5.0, this approach completed all 25 public ARC-AGI-3 games and achieved the benchmark’s maximum Relative Human Action Efficiency score.
Behnam and Wang identify a blind spot in utility-based memory management: a stored memory cannot be evaluated if the retrieval system never places it in the model’s context. Their Causal Memory Policy reserves context slots for randomized memory exposure, improving the ability to distinguish required from non-required memories, while also showing that query-specific utility does not reliably predict future value.
Afifi et al. trained an agent to research and forecast resolved real-world events, rather than supplying it with evidence collected in advance. On 265 held-out Polymarket questions, the resulting policy improved calibration by 30 to 40 percent, searched less often, and narrowly outscored four tested frontier models under the same evidence-based forecasting harness.
Ye and colleagues test whether reinforcement learning with verifiable rewards improves visual perception or merely teaches models to exploit nonvisual shortcuts. They find that benchmark gains can occur without visual training data, while tasks designed to make looking necessary substantially improve target discovery and transfer to independently constructed grounding evaluations.
Wu, Ding and Huang test whether GUI agents should learn from prior trajectories through fine-tuning or by retrieving notes into the prompt. Rather than choosing one destination for an entire trajectory, they split experience into four component types and route each using measurable properties, improving held-out success over either destination alone.
Wang and colleagues turn an omni-modal language model into an agent that chooses when to listen, when to look, and which temporal window to inspect. Trained on 170,000 synthetic reasoning trajectories and refined with reinforcement learning, the system reportedly improves performance across audio-visual benchmarks while reducing reliance on single-modality shortcuts.
Zhou and colleagues present DuoMind, a hierarchical system in which each robot independently plans, communicates with peers and executes detailed actions. Across the new RoboPoly benchmark and the existing RoboTwin benchmark, the authors report that this combination improves coordinated task performance, while ablations indicate that both semantic communication and hierarchical control contribute to the gains.
The authors introduce an 18-task benchmark that evaluates humanoid tool use as an end-to-end process, rather than treating recognition, manipulation and locomotion separately. Across seven policies in simulation and three on a Unitree G1 robot, they find that selecting an appropriate tool does not reliably translate into successful task completion.
Zhu et al. introduce Selection-based Structured Reasoning (SSR), which asks a multimodal search agent to choose among reusable natural-language reasoning templates rather than generate a new rationale token by token before every action. Across seven benchmarks with 2B and 4B parameter models, the authors report performance competitive with similarly sized search agents, alongside more than 90% lower per-turn reasoning latency and 28% to 54% lower total model inference latency per question.
Luo and colleagues introduce PoS, an inference-time framework that replaces reliance on accumulated interaction history with an explicit, continually maintained belief state. Across four execution and diagnosis benchmarks and three language-model backbones, the authors report that PoS achieved the best overall performance for every benchmark and backbone combination.
Wang and colleagues present Reconstruct, Practice, Go Real, a framework that improves a robot execution system through simulated practice without changing the underlying model weights. Across 22 manipulation tasks, success on held-out initializations rose from 28.6% after one practice round to 95.0% after 15 rounds, while a frozen version completed 30 of 30 reported physical trials across three tasks.
This week’s clearest signal is that stronger reasoning alone does not make agents dependable. The best systems instead constrain side effects, preserve recoverable state, select context deliberately and verify behavior through infrastructure the agent cannot rewrite.
Liu, Liu and Hu tested an LLM-agent framework that links company news into issuer-specific event sequences while restricting every forecast to information available at that time. Across daily Nasdaq-100 and Hang Seng Index panels from 2024 to 2026, their method outperformed the evaluated LLM agents and momentum baselines on predictive and portfolio metrics.
Kong et al. demonstrate that a conditional energy-based model trained purely on random exploration trajectories, without action or reward labels, can support long-range planning. The model learns temporal log-density ratios at multiple horizons, which a separate planner queries to evaluate candidate actions via a one-step dynamics model. The method achieves long-range maze planning from states and images, and extends to egocentric navigation and manipulation planning from suboptimal data.
The authors fine-tuned LLMs on aligned responses for specific contexts (e.g., recommending a researcher preserve data for reproducibility). They found this induced misaligned behavior in other contexts (e.g., recommending a mobile app developer save users' sensitive data). The effect, termed 'context confusion,' is narrow and not fixed by general safety data but can be mitigated by including domain-specific alignment data or providing in-context examples at inference time.
Jung et al. introduce SCLATE, an execution substrate for evaluating and post-training agents across sessions, background jobs and memory consolidation events. Across seven ported benchmarks, the system enables controlled comparisons of unmodified agent configurations, while post-training Qwen3.5-4B improved both software-engineering performance and held-out continual-learning accuracy.
Abouzahra introduces NAQD-Env, a benchmark for testing whether language agents revise plans selectively when evidence, authorization, or constraints change. Across 3,150 model-prompt episodes, withdrawal recall never exceeded 0.06, no valid resumption occurred when eligible, and only one episode fully matched the benchmark’s reference policy.
Wang and colleagues developed a multi-agent clinical system in which a local model selectively consults remote specialists while controlling what patient information leaves the institution. On an emergency-department benchmark derived from MIMIC-IV-ED, the system improved task accuracy while sharply reducing both direct disclosure and re-identification from combinations of otherwise indirect clues.
Di and colleagues formulate external-state checking for stored intentions as a resource-allocation problem under a hard episode budget. Across two benchmarks and three model backbones, their static policy preserved nearly all unconstrained task quality while making substantially fewer observations, while a future-aware policy performed better when observation capacity was scarce.
Fang et al. introduce a clinician-authored benchmark that evaluates the process of clinical assessment rather than diagnosis from a complete case summary. Across 31 models, diagnostic accuracy reached 90.7%, yet even the strongest systems passed fewer than 30% of tasks when required to question a virtual patient, request appropriate examinations, respect constraints and provide the correct diagnosis.
Liang, Wu and Dong automate the design of agent harnesses, programs that control which video frames a frozen vision-language model observes and when it answers. Their system combines Monte Carlo tree search, staged candidate evaluation and budget-specific routing, outperforming the strongest hand-crafted video agent by as much as 11.2 points across long-video question-answering benchmarks.
Jana et al. introduce MILO, a framework that automatically rewrites complete agent harnesses while also adapting how it searches for better designs. Using both a frontier model and an open-weight model, the authors report larger performance gains than prior search methods across software engineering, terminal-use and scientific-paper replication tasks.
Li and colleagues study contextual reasoning, the ability to allocate multiple context windows and decide what information should pass between them during inference. Their Hermes harness gives models varying degrees of control over those decisions, while Hermes-Learn uses supervised fine-tuning followed by reinforcement learning to teach smaller models how to use that control. The authors report that training produces adaptive strategies and performance gains that persist across benchmarks, models, and inference budgets.
Ma et al. present a search harness that lets an agent decide when to preserve useful evidence, discard a polluted context, or pursue another search path. Their 35-billion-parameter model scores 72.83 on BrowseComp and improves over its base model across several multilingual and domain-specific search benchmarks.
Tao et al. address a recurring failure in high-resolution visual question answering: crucial evidence can disappear when an image is resized, yet a model may not know where to zoom from an initial overview. Their Visual Parallel Search framework first examines a grid of tiles concurrently, then lets a main agent combine the tile reports, zoom into selected or merged regions, and answer the question. Across five benchmark splits and three model sizes, this scaffold generally outperformed a sequential zoom-only approach.
Das et al. introduce SEABench, a benchmark for testing whether agents that modify their own operating harness develop unsafe behavior over time. Across 48 longitudinal task sequences, self-evolution improved task completion but also produced safety regressions attributable to earlier updates, despite the absence of direct adversarial intervention.