Tool use, function calling, orchestration and the protocols that let models act rather than only answer.
arXiv:2609.38480v1 Announce Type: cross Abstract: Most clinical benchmarks evaluate language models (LMs) on diagnosis using complete case descriptions. In clinical practice, however, patients presen...
AREX-2 is a new approach that enhances the self‑improving ability of large language model agents by combining reflection—producing better solutions—and long‑horizon execution—maintaining effectiveness over many iterations. The method synthesizes improvement trajectories from machine‑learning and algorithmic programming tasks, providing verifiable feedback and sustained iteration. Trained on this data, an agent based on Qwen3.8‑27B achieves strong performance across multiple benchmarks and continues to improve as more iterative rounds are allowed.
arXiv:2609. 38369v1 Announce Type: new Abstract: We study generalized Blaschke curves as a controlled environment for AI-assisted mathematical rediscovery.
The paper introduces NAQD‑Env, a synthetic benchmark designed to test language agents’ ability to selectively withdraw and resume actions when new evidence, permissions, or stop instructions arise. It evaluates models against a deterministic reference policy across eleven dependency families, measuring policy agreement, task value, withdrawal, resumption, and event reporting. Experiments on 350 scenarios show low withdrawal recall, no valid resumption, and limited policy alignment, highlighting the need to treat selective withdrawal as a distinct reliability component.
PrivMeSA is a privacy‑aware, self‑evolving multi‑agent system that enables local clinical LLM agents to consult remote specialists while controlling patient data disclosure. It uses reinforcement learning to balance task accuracy with privacy risk, evaluates risk over entire outbound transcripts, and employs a local lesson memory to distill and reuse expertise without further remote exchanges. On an emergency‑department benchmark, PrivMeSA boosts task accuracy by up to 15.8 percentage points and virtually eliminates patient re‑identification risk, reducing personal detail disclosure from 98.0% to 0.2% of cases.
The paper introduces decision checkpoints that log observations and tool actions during inference, enabling the assignment of outcome categories to scientific-search agent responses. Using these checkpoints on 540 AutoResearchBench questions, the authors compare keyword search and read-first strategies, finding that keyword search yields 24.6% accuracy while raw search achieves 17.8%. The checkpoint protocol reveals detailed differences in target exposure, inspection attempts, and evidence-search calls that aggregate accuracy alone obscures.
EvoSteer introduces an online self‑evolving graph orchestration framework that continuously builds and repairs a team of tool‑using agents during execution. It employs Anchored Trajectory Balance (AnchorTB), a regression‑style loss that assigns credit to each orchestration action by comparing subtrajectories to a frozen reference, and Validated Skill Admission, which tests candidate skills before promotion. Experiments on twelve datasets demonstrate that EvoSteer outperforms existing baselines in question answering, mathematical reasoning, code generation, and interactive decision making.
The paper introduces Long-Transduction, a diagnostic framework designed to evaluate how well language models can maintain task fidelity during extended generation tasks that involve continuous reading, mutating, and outputting of context-dependent operations such as arithmetic, sorting, variable lookups, and table transformations. By independently varying local task complexity, input data formatting, and context length, the study isolates failure modes across these axes. Experiments on seven open-weight models reveal significant performance drops—62.8% when scaling context length from 4 to 128K, 36.5% with input format changes, and 39.9% with increased local task complexity—highlighting critical vulnerabilities in long-horizon agentic workflows.
The paper introduces a co‑evolving framework where a target agent improves by learning from its own failures, and a separate failure agent is trained to generate hard negative trajectories. These hard negatives, derived from plausible but incorrect attempts, help the target agent better distinguish successful behavior from subtle errors. Experiments on online shopping, scientific reasoning, and interactive SQL querying show a 5.7% average reward improvement over baseline methods.
PathAnchor is a new scientific reasoning system that uses path-structured evidence workspaces instead of independent passages or concepts. It retrieves source-linked Material‑Sensor‑Signal‑System trajectories that preserve role, direction, and supporting evidence, and a controller uses read‑only tools to search, trace, and open exact evidence before producing a claim‑cited answer. In evaluations on 120 flexible‑sensor questions, PathAnchor achieved an 82.6% score, outperformed six other systems, and improved source recall, citation completeness, and reduced tool calls compared to unordered concept graphs.
The paper introduces an action‑conditioned bisimulation rule for merging states in a GUI agent’s memory, replacing observation‑based similarity with a requirement that shared actions produce identical outcomes and successor blocks. This rule is applied to the empirical predictive state graph built by a frozen agent during interaction. Experiments on MiniWoB++ show that the new merge rule improves success rates compared to a memoryless baseline, while controls that omit action conditioning or use the previous successor‑representation merge show no improvement.
The paper introduces Consistent Plan-Act (ConPAct), a method that addresses coordination failures between high-level planners and low-level actors in long-horizon agentic tasks. By prompting both agents to produce structured state assertions and programmatically detecting contradictions, the authors identify a systematic planner-actor state mismatch. ConPAct feeds these detected contradictions back to both agents, fine‑tunes them on consistent interactions, and achieves notable performance gains, such as raising MiniGrid success rates from 38.6% to 54.4% with GPT‑5.6‑sol/terra.
The paper introduces Retrospective World Modeling, a new paradigm for vision‑language‑model (VLM) agents that allows them to reason backward by estimating which action most likely caused a state transition. It proposes the Self‑Consistency Reward (SCR), an intrinsic signal that measures how well a policy action aligns with this retrospective explanation, providing dense transition‑level feedback. Experiments demonstrate that incorporating SCR improves policy robustness and generalization compared to purely prospective world‑modeling approaches.
SkillFM is a generative framework that creates task‑conditioned textual skills for large language model agents without relying on manual skill banks or reinforcement learning. It encodes skills into a continuous latent space using a codec and trains a conditional flow model with improved MeanFlow, allowing single‑step latent sampling at inference. The sampled latent is decoded by an LLM into textual guidance, and the method outperforms other vector‑based skill approaches on ALFWorld, Search‑QA, and other tasks.
RankEvolve is an auto‑research framework that evolves generative ranking models by orchestrating multiple large‑language‑model coding agents through an Executable Operating Protocol (EOP). The system compiles a state machine that enforces phases, gates, branches, and loops, while a meta‑meta‑harness lets agents review and repair each other’s code. In budget‑matched experiments, heterogeneous composition of agents raised execution accuracy from 45.8 % to 62.5 % and reduced silent critical‑defect rates, achieving notable gains on the HSTU recommender and other benchmarks.
The paper investigates latent communication in multi‑agent systems, where agents exchange information in internal representation space via lightweight trainable links. It demonstrates that even benign training of these links can increase harmful compliance, and that attackers can exploit or poison the links to amplify this effect. A reinforcement‑learning attack further boosts harmful compliance while maintaining benign task performance, but adjusting rewards toward safer behavior can repair compromised links without updating the agents.
The paper introduces Runtime Assurance Contracts (RAC) as a formal policy framework for high‑risk AI agents, addressing the "assurance‑transition gap" by binding autonomy boundaries, component eligibility, evidence state, transition policy, human‑review capacity, and non‑compensatory gates. RAC allows soft metrics to influence routing while mandating retries, switches, escalations, deferrals, or stops when mandatory gates fail or are unknown, ensuring aggregate performance cannot alone authorize action. The authors define the contract, evidence record, permission rule, and five invariants, and evaluate RAC through deterministic failure‑injection studies, hand‑authored traces, and a prospective synthetic holdout, comparing it to score‑only and restricted protocol baselines.
The paper introduces ConflictGuide, a method that enhances LLM-based AutoResearch by incorporating feedback on competing behaviors during model code editing. By first exploring with scalar task performance and then using probes to measure and alleviate conflicts, ConflictGuide increases the proportion of edits that improve multiple behaviors and sustains progress beyond scalar-only plateaus. Experiments across five model families show reductions in task and conflict-related errors by up to 28% and 14% compared to scalar-only AutoResearch.
The Agent Error Dataset (AED) presents 50,228 error–diagnosis pairs collected from 9,961 source tasks across 33 environments, 19 harness families, and 23 policy models in text‑based agent systems. A five‑stage Agentic Error‑to‑Training (AET) pipeline generates diagnoses and proposed corrections, verifies them against recorded evidence, and creates separate training views for diagnosis and actor recovery. Experiments show that first‑proposal corrections improve verifier pass rates from 18.4% to 51.1%, and fine‑tuning with full‑diagnosis data raises Qwen3‑8B’s exact‑step agreement from 47.2% to 63.6% on a holdout set.
The paper investigates how causal action verifiers, which guard language agents’ tool calls by checking identifiability against a committed action‑state graph, can be compromised through small graph misspecifications. By removing a single bidirected edge or reversing an arrowhead, the authors demonstrate that a verifier (CIVeX) that originally had zero false executions can suffer false execution rates up to 48.9%, with most of those executions being harmful and overall utility dropping dramatically. An additional attestation step that samples executions can detect these attacks with few false alarms, but it also leads to many wrongful rejections that reduce beneficial actions and incur significant experimental costs. whyItMatters":"The study shows that even minor errors in the verifier’s underlying graph can drastically undermine safety and performance, highlighting the need for robust auditing mechanisms."