The paper investigates whether an agent’s own verifier remains reliable when the agent is optimized against it. It introduces a stress‑testing framework that measures the pressure ratio ρ, the influence of the verifier on the agent’s search, and four falsifiable probes targeting common failure modes. Applying this framework to the Rationale Consistency Score (RCS) in an LLM agent that designs machine‑learning pipelines, the authors find that the verifier’s impact is minimal in their baseline setup, that agents can shorten claim lists without affecting search, and that pre‑commit probing can hide declines in faithfulness. The study concludes with a proposed reporting standard to audit verifiers before trusting them.
By Jonghyun Lee, Moonsu Kim, Sunghwan Park
Reward Engine is a method for creating a reward function for open‑ended tool agents by first determining what success looks like and then optimizing toward it. It builds a RewardPack of test cases and rubrics from initial trajectories, which is frozen and used to guide the agent’s output selection and can also serve as a reinforcement‑learning reward. Experiments on WorkBuddy Bench Office and Claw‑Eval show that the constructed reward aligns with benchmark graders on most candidate pairs and improves three‑run consistency from 45.0% to 54.0% and from 78.4% to 83.7%.
By Litong Feng, Ji Qi, Yiran Cheng, Shaowu Pan, Xinyu Zhang, Qijun Li
The paper introduces MAGE, a framework that uses user manuals to enhance computer use agents (CUAs) for new software. MAGE retrieves relevant manual content during execution, generates procedural guidance, and adapts the agent policy through interaction. In training, it converts manuals into executable task-plan pairs to guide rollouts, reducing the need for human‑annotated demonstrations and improving task success rates across six applications.
By Chulhoon Yoon, Kyungri Park, Jinyeong Yun, Daeyun Jeong, Sukmin Yun, Woohwan Jung
The paper introduces Distill Until Divergence, a method for agentic reinforcement learning that uses error-truncated hindsight distillation. It routes two complementary hindsight skills—Positive for successful workflows and Negative for failure corrections—based on trajectory outcomes and error boundaries. The post-error segment is optimized solely by GRPO, preserving outcome-based learning while avoiding mismatched hindsight supervision.
By Yibin Huang, Conghui Zhu
The paper introduces Agentic In-Context Learning (Agentic ICL), a framework where fixed‑parameter Transformers learn from interactions that shape the context they receive. It models an interactive learner with state, update, and decision rules, covering active learning, bandits, and online reinforcement learning, and demonstrates that causal Transformers can approximate these algorithms over a finite horizon. Experiments show that Transformers can learn to execute recursions and that different reward signals produce distinct decision policies, with alignment between process supervision and outcome‑optimal actions influencing learned behavior.
By Shiguang Wu, Yaqing Wang, Quanming Yao
WebCoder is a general frontend development agent that uses artifact‑grounded agentic reinforcement learning to learn how to compose, generate, modify, and repair code over an evolving application state. It evaluates development actions by their impact on a shared executable artifact and a verifier that provides terminal rewards, while intermediate progress redistributes rollout advantage across development segments. The system includes Webdev‑World, an executable environment that integrates code editing, execution, browser inspection, and on‑demand skills, and Webdev‑Bench, a benchmark for composed development tasks, where WebCoder‑27B outperforms several large language models on requirement satisfaction and visual/functional scores.
By Zijun Dou, Zhen Yang, Xijun Liu, Dongzhuoran Zhou, Aohan Zeng, Jie Tang
QuAD is a framework that quantitatively assesses the quality of qualitative response datasets before analysis. It compares an intrinsic cluster distribution derived from the responses with an extrinsic, researcher-defined analytical lens based on metadata, using a four‑quadrant assessment that considers lens‑level relations and within‑cluster variation. The method employs a content‑aware multi‑agent system with two‑stage instruction generation and adaptive anchors to improve scoring consistency, and is validated through synthetic experiments and a real‑world social survey case study.
By Wenshuo Zhang, Leixian Shen, Qiqi Zhou, Shuchang Xu, Feiyang Hao, Huamin Qu, Haobo Li, Zhao Chen, Lin-Ping Yuan
OfficeTown is a generative simulation framework that models professional organizations, where tasks and context evolve from ongoing work. Starting from a natural‑language description of an organization’s purpose, the system initializes members who coordinate, delegate assignments to assistant agents, and review deliverables, while artifacts and decisions persist to influence future tasks. The authors built OfficeTown Bench with 406 tasks and 4,836 evaluation criteria across 23 simulated organizations in eight domains, and show that fine‑tuning and on‑policy distillation on this data improves Qwen3.5‑9B on external professional‑work benchmarks, with a principal–assistant setup outperforming single‑ and multi‑agent baselines.
By Hongchen Wei, Daohong Ying, Zhi Hong, Aidan Wu, Bei Liu, Ruichun Ma, Yifan Yang, Qi Dai, Kai Qiu, Rui Xu, Dongdong Chen, Zhenzhong Chen, Chong Luo
The paper argues that additional computation from the same large language model (LLM) during test time does not necessarily provide independent evidence, as errors are correlated across repeated samples, reasoning depth, and self-evaluation. It introduces DIA (Disagreement-triggered Independent Arbitration), which uses disagreement between intermediate and final predictions as a signal of instability and resolves it by having a small panel of heterogeneous models independently solve the question. Across seven benchmarks and two base models, DIA outperforms full chain-of-thought in most settings, achieving significant gains while using fewer tokens and activating only on a subset of examples.
By Ming Shen, Tianjun Yao
Annotrace is a workbench that converts agent traces from six formats—including OpenTelemetry GenAI, LangSmith, and Langfuse—into an ordered, nested list of execution steps. Users define a labelling codebook, and a frontier model proposes initial labels, enabling a joint human‑AI annotation process. The tool also measures agreement between annotators, reproducing a prior manual effort’s 235 annotation units and achieving 91% agreement (κ = 0.82) between two models on a multi‑agent question‑answering corpus.
By Arun Josephraj Arokiaraj, Chenyi Huang, Sahan Bulathwela, Zekun Wu, Bin Wu
The paper introduces SchemaShield-Lite, a lightweight, training‑free repair layer designed to address interface drift in tool‑augmented language‑model agents. By combining schema validation, candidate‑set retry for missing tool names, and one‑shot constrained repair, the method improves accuracy on Qwen3.5‑9B from 0.717 to 0.813 on DICE‑300 and from 0.580 to 0.805 on BFCL‑200, recovering a majority of drift‑induced failures without observable harms. Ablation studies reveal that candidate‑set drift, particularly due to distractor tools, is the primary source of damage, while description and schema edits alone are less harmful.
By Aniruddha Chattopadhyay, Kanishka Halder, Partha Sarathi Roy
AgentOnto is an ontology-guided large language model agent designed for enterprise data analytics. It structures heterogeneous enterprise knowledge—such as databases, business documents, and example tasks—into an ontology that directs analysis and incorporates confirmed feedback. The system improves its ontology through iterative preparation and demonstrates superior performance over existing methods on public datasets, with a real-world deployment confirming its practical value.
By Dayan Pan, Bincheng Zuo, Xiaojian Wang, Daogang Zhang, Jingyuan Wang
The paper introduces S‑GAP, a stealthy graph attack primitive that unifies textual and structural perturbations into a single coherent attack intent for Graph Large Language Models (GraphLLMs). By iteratively refining the semantic content of the GAP through a multi‑agent framework, S‑GAP achieves higher attack success rates and produces more relationally plausible perturbations compared to existing methods. Experiments on four datasets and three GraphLLMs demonstrate an average 7.02% improvement in attack success rate, revealing significant vulnerabilities in current GraphLLMs.
By Wenda Li, Tongya Zheng, Shunyu Liu, Hanyang Yuan, Kaixuan Chen, Yu Wang, Yuchen Ying, Haoze Jiang, Huan Li, Gang Chen, Zujie Ren
Darwin‑Claw is a new framework for large language model agents that enables environment‑adaptive, self‑evolving behavior. It systematically distills episodic experiences into prototype vectors through a hierarchical experience consolidation system, providing context‑sensitive perception patches and explicit optimization signals for skill refinement. The method is deployable with mainstream agent systems and, according to experiments on WildClawBench, delivers significant performance gains while maintaining controllability and stability.
By Mingze Yin, Xiaohan Wang, Dian Li, Longfei Luo, Jiahong Yan, Yiheng Zhu, Chang-Yu Hsieh
The paper introduces the self‑building laboratory (SBL), an autonomous system that designs, configures, and operates its own laboratory setup based on a high‑level scientific goal. Unlike existing self‑driving laboratories that can only run pre‑built procedures, the SBL starts from any existing environment—whether empty or partially equipped—and builds missing capabilities such as lab layout, digital twin, robot skills, and experimental workflows. The authors demonstrate the system across multiple robots and setups, including a real‑world quantum‑dot synthesis experiment, showing that researchers can shift from manual configuration to simply specifying scientific objectives.
By Jiaru Bai, Yuchi Zhao, Han Hao, Miroslav Bogdanovic, Yang Cao, Varinia Bernales, Al\'an Aspuru-Guzik
The paper investigates using Large Language Models (LLMs) as calibration agents for expensive engineering simulators by conditioning them on physics-labeled prompts that describe simulation errors. In tests on a cylinder-wake Large-Eddy Simulation and a damped-oscillator inverse problem, physics-informed LLM proposals outperform conventional derivative-free optimizers in early trials, but the advantage fades after fifteen evaluations. A hybrid workflow that uses the LLM as a warm-start followed by a conventional optimizer achieves target accuracy in only nine simulator runs, outperforming both the LLM alone and standard optimizers.
By Ruo-Qian Wang
DriveWorld3D is a framework that integrates temporal video learning, multi‑sensor measurement constraints, and offline generative simulation to build a unified 3D world model for autonomous driving. Its online component tracks road surfaces, volumetric occupancy, dynamic entities, and road relations while explicitly managing physical occupancy, observability, and evidence provenance. The offline component reconstructs scenes with neural radiance fields or 3D Gaussian Splatting and expands rare conditions using generative models, but limits generated content from being used as geometric ground truth.
By Wei Wang (Robostreet research team Cambridge, MA, USA), Yiru Veronika Wang (Robostreet research team Cambridge, MA, USA), Yihan Wang (Robostreet research team Cambridge, MA, USA), Zihao Li (Robostreet research team Cambridge, MA, USA), Xiaohui Liang (Robostreet research team Cambridge, MA, USA)
AuditKIT is a source‑available evaluation library that unifies benchmark accuracy, judge‑based assessment, runtime performance, tool‑use behavior, and retrieval quality into a single result representation. It offers direction‑aware metric contracts, configuration fingerprints, and extensible metrics and backends, while providing turn‑aware tool‑call matching and distinguishing verified task outcomes from trajectory‑level diagnostics. Seven case studies demonstrate how answer extraction, runtime behavior, and metric selection can alter evaluation outcomes across model comparison, serving runtimes, retrieval, tool calling, and interoperability.
By Shreeyans Arora, Utsav Avaiya, Zera Lyngkhoi, Hem Gosalia, Ram Mohan Rao Kadiyala, Vinay Kumar Sankarapu, Pratinav Seth
AlphaBet is a reinforcement learning framework that trains a single large language model to search for evidence, reason about it, and report probabilities for unresolved events across multiple pre-resolution dates. By scoring each dated forecast with a proper scoring rule, AlphaBet uses one resolved event to supervise decisions under varying information states, decomposing forecasting regret into evidence-acquisition and probability-reporting gaps. Experiments on GLM-4.5-Air show that AlphaBet improves Brier scores on Forecast-Dojo and FutureX, attributing gains to multi-date training and better information acquisition and probability reporting.
By Liqin Ye, Rongzhi Zhang, Xin Liu, Zheng Li, Yuan He, Qingyu Yin, Priyanka Nigam, Chao Zhang
NoveltyScope is a prototype that evaluates the novelty of a scientific paper by comparing it with prior work and generating a structured report and a five‑level novelty score. It extracts the core task and up to two author‑claimed contributions, retrieves relevant passages from a local corpus, and uses a four‑role multi‑agent system (Prosecutor, Defender, Judge, Critic) to assess novelty, allowing a single revision of the Judge’s score. Evaluation on 50 RINoBench samples shows that multi‑agent scoring with Gemini or Qwen3 achieves lower mean absolute errors (1.26–1.16) than single‑LLM scoring (1.48–1.22) and outperforms a vector‑similarity baseline (1.68).
By Yueran Tian, Huizhi Liang