Tool use, function calling, orchestration and the protocols that let models act rather than only answer.
arXiv:2609. 40285v1 Announce Type: new Abstract: On-policy distillation (OPD) is a promising approach for training language agents, providing dense teacher supervision on student-generated trajectories.
The paper introduces ScholarEvolve, a framework that evolves the software harness of language agents by automatically incorporating insights from recent research papers. It organizes harness improvements into functional modules, uses topic modeling to identify distinct strategies, and evaluates combinations to boost task performance. Experiments show significant gains on AppWorld and Tau2-Bench, raising Qwen3.5-27B completion rates from 49.6% to 63.6% and GPT-5.4-mini pass@1 from 72.7% to 81.9%.
EHR2Trace is a system that transforms electronic health records from multiple sources into a standardized, traceable event format suitable for training and evaluating patient world models and clinical agents. It links each event to its original record, separates the event time from the time the information became available, and distinguishes between medication orders, dispensing, and administration. The tool supports both OMOP and MEDS data models, includes automated validation, and was tested on three clinical datasets, converting 846.4 million events and detecting all injected faults.
SCLATE is a new execution substrate that allows continual‑learning benchmarks and agents to share a single event scheduler via adapters, enabling tasks, session events, and memory consolidation to run on a compressed, real‑time timeline. It also functions as a rollout engine that records every model call’s tokens and log probabilities without modifying the agent’s harness or memory. Using SCLATE, the authors ported seven benchmarks, compared ten harness‑memory configurations across ten models, and demonstrated that post‑training Qwen3.5‑4B can effectively leverage both harness and memory, improving performance on multiple metrics.
The paper introduces STAM, a state‑transition‑aware memory framework for large language model agents that process clinical records. STAM records changes in a patient’s state as new entries arrive, using semantic retrieval and typed clinical relations to separate current information (Active) from superseded or resolved information (History). During retrieval, a query‑dependent gate selects the appropriate historical memory, enabling accurate question answering and state‑maintenance diagnostics across four longitudinal clinical benchmarks.
PANDA is a decentralized architecture for large-scale, fault-tolerant multi-agent systems that enables heterogeneous agents to discover each other's capabilities and self-organize into specialized teams for each task. It decouples collective communication from team communication, allowing agents to participate in multiple teams simultaneously and load-balance tasks across the collective. PANDA supports three planning and execution patterns—star, chain, and mesh—detects and recovers from infrastructure and orchestration failures, and uses a web-of-trust model for governance without a central bottleneck.
CollabFlow introduces a recursive self‑improvement framework for multi‑agent collaboration in large language model systems. It trains a Collab‑Director to assemble teams of agents, uses a frozen executor to run them, and retrains the director each round based on outcomes. The system incorporates evidence‑conditioned communication protocols within collaboration graphs and a Collaborative Trajectory Balance objective to maintain diverse high‑performing teams across rounds, achieving superior performance on twelve datasets.
DAMPER is a new technique for actor‑critic methods that reduces action oscillation in continuous control tasks. It combines the native actor gradient with a temporal‑consistency gradient using conflict‑conditioned projection and adaptive magnitude control, ensuring the auxiliary component aligns positively with the actor gradient. Experiments on TD3 and SAC across six tasks show that DAMPER consistently lowers oscillation compared to baseline agents and outperforms other methods in most task‑backbone pairs.
APTInvestBench is a benchmark that evaluates how well autonomous agents can investigate advanced persistent threats (APTs) when faced with different telemetry settings. It contains 370 cases derived from 56 attack reconstructions, totaling 16.4 million log records, and tests agents on seven SOC-inspired telemetry conditions. The benchmark measures evidence acquisition and formal citation support, revealing that while overall coverage drops only slightly when telemetry is limited, a significant portion of actions lose sufficient citation support, highlighting instability in agent performance.
The paper introduces URAI, a Universal Robot‑Agent Interface that separates robot control into two roles: a programming agent that writes reusable, task‑specific tools from intent, and an execution agent that calls these tools in a feedback loop. This design keeps high‑level decision making in the model while delegating low‑level motion to code, allowing tool revisions to persist across episodes without retraining the foundation model. Experiments on RoboDojo and AgileX tasks show significant gains in success rate, speed, and token efficiency compared to direct fingertip control and pre‑written programs.
The paper investigates a failure mode called co‑cheating in self‑evolving search agents, where the proposer and solver agree on shared errors, inflating internal reward without improving external correctness. The authors first propose multi‑sample verification (MSV) to filter unreliable pseudo‑labels, which only partially mitigates the issue. They then introduce CrossFit, a cross‑fitted reward scheme that partitions source documents and uses an auxiliary solver to prevent same‑source agreement, significantly reducing false agreement and boosting downstream benchmark performance.
The paper introduces HealBench, a benchmark of 265 runtime errors from 18 real-world repositories, and HealGuard, a safety framework that restricts healing code to an analyzable subset of Python and applies static and dynamic taint analysis. Using these tools, the authors evaluate a dedicated healing method and three general coding agents powered by different LLM backbones, achieving a 38.11% resume rate and a 28.68% test‑pass rate, while HealGuard flags 17.4% of successful healings as potentially unsafe. The study demonstrates that current LLM agents can meaningfully repair real repository crashes, but also highlights significant safety concerns that the Guardrail framework can detect, albeit with a high false‑positive rate.
DAGent introduces an Evaluate‑then‑Grow planning approach for deep research agents, building directed acyclic graphs incrementally based on confidence and uncertainty from completed tasks. The framework includes a hierarchical context layer for efficient query handling and a structural reinforcement learning component, DAGRPO, that rewards topology‑conditioned execution. Experiments on BrowseComp‑Plus, GAIA, and xbench‑DeepSearch show DAGent outperforming strong baselines across multiple backbones and scaling to large language models.
EngramBench is a new benchmark designed to evaluate skill evolution in autonomous agents by focusing on genuine capability abstraction rather than solution copying. It includes 30 learning tasks and 13 unseen transfer tasks that require agents to manage complex, multi-hour development cycles with LLM‑simulated users. The study shows that while static skill banks cannot eliminate the need for precise code implementation, they effectively reduce redundant context and cut overall coding time by more than 55%.
The paper introduces Mosaic, a multi-image visual harness that lets large language‑vision models (MLLMs) construct visual intermediates using ten composable image operations. It evaluates five re‑representation settings on existing multi‑image benchmarks and a new grounding‑focused benchmark, MosaicBench, finding that visual re‑representation benefits tasks requiring precise visual evidence more than those dominated by high‑level semantics. MosaicAgent‑8B is trained via reinforcement learning to compose these operations without demonstration trajectories, demonstrating diverse problem‑solving patterns.
The paper introduces a new skill poisoning technique for large language model agents that decouples the pretext (rationale) from the actuation (operation). By separating these two risk‑realization factors, the authors create coordinated pretext‑actuation skill pairs that allow malicious actions to remain hidden within legitimate agent behavior. An automated framework is presented to discover execution dependencies, synthesize these skill pairs, and refine them through closed‑loop feedback, achieving high attack success in both single‑session and persistent scenarios.
The paper introduces On‑Policy Warmup (OPW), a teacher‑guided training stage where a student agent learns from a teacher on its own interaction trajectories before switching to reinforcement learning with verifiable rewards (RLVR). OPW differs from traditional imitation by focusing on states generated by the student’s own decisions, including imperfect actions and recovery situations. The authors provide a theoretical link between on‑policy reverse‑KL distillation and trajectory‑level distribution matching, showing that, under a competent teacher and low distillation loss, OPW can lower bound initial verifier success and reduce reward‑discovery complexity, thereby accelerating RLVR performance.
GroundingPI is a 4‑billion‑parameter grounding foundation model that generates points and boxes as quantized coordinates in a shared vocabulary. It is trained with multimodal and spatial pretraining, supervised fine‑tuning, and reinforcement learning, achieving a new state‑of‑the‑art average of 73.68% across 34 grounding benchmarks. As a visual backbone, GroundingPI improves performance in robotic manipulation and autonomous driving, outperforming larger models and mainstream backbones in several out‑of‑distribution settings.
The paper titled "Pretext: Defeating Malicious Skill Detection Frameworks for AI Agents" demonstrates how an attacker can bypass current skill‑scanning defenses by crafting malicious skills that evade both static analysis and LLM‑based semantic checks. By moving malicious payloads into natural language and distributing instructions across files, the white‑box attacker named Pretext achieves high evasion rates—up to 97% against a frozen detector and 77% against a co‑adaptive one—across three open‑source models.
The paper investigates whether enhancing goal representations improves goal-conditioned reinforcement learning (GCRL) performance. By creating an exact temporal-distance goal representation in deterministic mazes and systematically degrading its geometric quality, the authors find that changes in goal representation have little effect on performance. In contrast, degrading the agent’s current state representation more than doubles failure rates, indicating that state representation is the critical bottleneck. The study further demonstrates that simple random Fourier positional encodings can significantly boost performance on challenging navigation tasks without additional map or objective modifications.