The paper introduces TRIO, a retrieval‑augmented generation framework that uses three agents—query extension, judge, and answer—to iteratively refine queries and improve retrieval and answer quality. It aligns each agent’s objective with rewards derived from answer quality, employs unanimity voting to label expanded queries, and applies dynamic matching‑based rewards with curriculum learning for the answer agent. Experiments on multiple benchmarks show TRIO outperforms state‑of‑the‑art methods.
By Yingtao Ren, Xiao Luo, Yiwei Fu, Jinzhao Zhou, Yu-Cheng Chang, Chin-Teng Lin
The paper surveys the evolving field of agent skill evaluation and evolution, emphasizing the shift from isolated skill creation to automated, evaluation-driven development. It categorizes skill evolution into four paradigms—execution feedback, trajectory distillation, compression, and reinforcement learning—and analyzes six skill-centric benchmark categories to highlight gaps and trade-offs. The authors outline open research directions for building generalizable, efficient, and verifiably safe skill ecosystems.
By Kexin Ding, Yang Zhou, Can Jin, Feng Tong, Mu Zhou, Dimitris N. Metaxas
TMax is an open reinforcement‑learning recipe that trains terminal‑using agents from small language models, achieving 33% on Terminal‑Bench 2.0 with only 9B parameters and outperforming larger prior models. The approach generates training data via a novel taxonomy that blends difficulty control, personas, and verifier diversification, enabling cheap creation of many terminal environments for both RL and supervised fine‑tuning. The recipe is applicable to models ranging from 2B to 27B across different families, dominating the Pareto curve for open terminal‑agent performance at smaller sizes.
By Hamish Ivison, Junjie Oscar Yin, Rulin Shao, Teng Xiao, Nathan Lambert, Hannaneh Hajishirzi
AgentSnare is a trajectory‑adaptive deception system that expands a decoy environment while autonomous web‑penetration agents explore. It uses an LLM‑driven decoy constructor and a virtual kernel to generate and materialize decoy artifacts that persist across interactions, enabling long‑horizon exploration. In experiments on 15 CVE‑Bench web applications and four attacker models, AgentSnare delays 49.9% of tool calls, diverts 57.2% of post‑entry actions into decoys, and defuses 89.5% of completion reports, preventing any real target exploitation at pass@3 across 60 attacker‑CVE pairs.
By Ruoyu Wang, Heng Zhao, Renjie Wu, Tianhang Zheng, Mengnan Zhao, Wanyu Lin, Kui Ren
SelfImprove-Bench is an evaluation framework for recursively self‑improving (RSI) systems that organizes tasks into three core capabilities—Exploring, Planning, and Reasoning—using game‑based scenarios for controllability and reproducibility. It employs scalable task generation and orchestration driven by templates and difficulty coefficients, and introduces a generalization ladder (G0–G4) to assess how well performance gains transfer beyond training conditions. Experiments show that current RSI methods exhibit inconsistent evolutionary gains across these generalization levels.
By Bangrui Xu, Jie Yang, Sidi Chen, Xuanhe Zhou, Jiantao Qiu, Qianyu Yang, Zijie Wang, Fan Wu, Zhoufutu Wen
MEMENTO is a long‑context agent built on Qwen3‑8B that learns to manage its fixed‑size context window through compaction, replacing its history with self‑generated notes. The model is first fine‑tuned on synthetic trajectories with hindsight notes and then reinforced with a reward combining answer correctness, note quality, and confidence calibration. At test time, MEMENTO can also fold context and backtrack based on confidence, achieving a 67.8% accuracy on book‑level QA (up to 2.5 M tokens) while never holding more than 20 K tokens, outperforming GPT‑5.5 and GLM‑5.3 agents at significantly lower cost.
By Yunfan Zhang, Taehoon Hwang, Kathleen McKeown, Smaranda Muresan
CuriousTogether is a framework that evaluates curiosity-driven interactions among multiple large language models (LLMs) across novelty-seeking, epistemic, and social curiosity dimensions. In eight tasks covering reasoning, creative generation, and social bias mitigation, curiosity-guided multi-agent settings outperform non-curious ones on seven of the eight tasks, achieving up to a 20.6% increase in story semantic diversity and a 14.9 percentage point drop in stereotype rates. The study also shows that country-conditioned cultural grounding can improve story generation and social tasks, though it may reduce precision in reasoning, and that inference-time curiosity-based multi-agent approaches offer better quality-token trade-offs compared to single-agent baselines.
By Angana Borah, Zoha Hayat Bhatti, Zhijing Jin, Rada Mihalcea
IdeaTaste is a benchmark that evaluates large language models (LLMs) on their ability to generate research proposals that align with what human researchers actually publish. The study compares proposals generated by six contemporary LLMs against real papers from four research areas, finding that LLMs achieve only about 26% of the citation matching and 18% of the idea matching of human communities. Even larger LLM populations and richer field contexts do not close this gap, indicating that LLMs struggle more with translating research directions into concrete approaches than with selecting the right topics.
By Shanshan Zhong, Chen Xu, Jingjie Ning, Zhengzhong Liu, Aviral Kumar, Chenyan Xiong
Skill++ addresses lifelong skill routing for large language model agents by rewriting the name and description of each new skill instead of retraining the router. This approach preserves earlier skills while allowing new ones to be selected, achieving 76.3% query success versus 36.4% with original metadata across a library of 731 skills. It outperforms existing methods on three benchmarks and five LLM routers, and its benefits transfer to the Codex CLI agent harness.
By Hanrong Zhang, Minghao Liang, Jiawei Wu, Yankai Chen, Shicheng Fan, Zhiwei Liu, Enze Ma, Junwei Yuan, Tianyi Wang, Dehai Min, Kening Zheng, Xue Liu, Philip S. Yu
The paper investigates whether large language model (LLM) coding agents can function as forward‑deployed engineer (FDE) agents by discovering customer requirements and executing deployment plans in a real enterprise text‑to‑SQL scenario. In a simulated engagement, agents must uncover acceptance criteria, data permissions, and hardware constraints before planning and delivering a post‑training recipe. Across 15 models, the best agents uncovered over 80% of hidden requirements, yet only 25% of runs met all acceptance criteria, revealing bottlenecks in discovery, execution reliability, compute usage, and benchmark relevance.
By Marianne Menglin Liu, Rongguang Wang, Zheng Wang, Yewen Wang, Leonid Boytsov, Tao Sheng, Sujith Ravi, Dan Roth
The paper introduces Agent-Authored World Modeling (AAWM), a training method that replaces traditional next-observation prediction with decision-oriented supervision. AAWM constructs training targets by having the agent identify what it needs to know before acting, then retrieves relevant transition evidence across trajectories to synthesize these needs into learning signals. Experiments across multiple environments demonstrate that this decision-aware approach yields more effective learning than conventional next-observation prediction.
By Guangfeng Cai, Kaibing Yang, Shuo He, Yu Li, Shengtian Yang, Jiaqi Lv, Lei Feng
WebHarbor is a collection of 153 dockerized website environments that faithfully replicate real-world web services, offering 3,518 tasks that include multi-step and post-authentication workflows. Built through task-driven human-agent collaboration, the suite ensures deterministic control, visual fidelity, and functional correctness. The platform enables reliable evaluation and cost-efficient reinforcement learning, with a Qwen 4B-level GUI agent trained in WebHarbor generalizing well to unseen tasks and matching leading open-source agents on four benchmarks.
By Zhaoyang Wang, Qianhui Wu, Boyu Gou, Demi Wang, Jiacheng Zhang, Peichun Hua, Qize Li, Shi Qiu, Yifan Jiang, Yifan Zhang, Zexu Jin, Baolin Peng, Bolin Shen, Chetan Bansal, Dehan Kong, Deren Lei, Dongshen Peng, Fengqing Jiang, Hao Cheng, Haoqing Wang, Jiaming Shan, Jicheng Wang, Jingguang Zhou, Jixuan He, Kaide Zeng, Kangyu Zhu, Lama Moukheiber, Mufan Qiu, Peng Xia, Qihan Sun, Qingwei Lin, Qiushi Sun, Ruichen Zhang, Shuqing Luo, Si Qin, Sicheng Fan, Tao Xiong, Wenlin Yao, Xinyu Zhang, Xinyue Zeng, Xuanrui Li, Xuchao Zhang, Yian Wang, Yichao Liu, Yiming Liang, Yiyang Zhou, Yizhu Jiao, Yuetai Li, Yurun Chen, Yuxuan Zhang, Zheyuan Yang, Zhiyuan Yao, Zhongyang Li, Zihao Li, Zili Wang, Jianfeng Gao, Huaxiu Yao
The paper introduces VIAS, a method for Value-Isolating Activation Steering that aims to improve multi-value compliance in large language model agents. VIAS constructs steering vectors by contrasting responses that include all relevant values with those that omit a single target value, ensuring the target value changes from compliant to non-compliant while maintaining compliance on other values. Evaluations on the PROTECT workplace compliance dataset show that VIAS reduces vector overlap and, when combined with prompting, achieves higher compliance rates across three models compared to prompting alone.
By Avni Mittal, Avinash Anand, Dikshant Kukreja, Erik Cambria, Rajiv Ratn Shah, Sandipan Dandapat
CoworkOS is a general‑purpose harness designed for long‑horizon real‑world automation, separating task strategy from execution management. Its three‑layer architecture allows direct execution or delegation to subagents, while a shared application layer lets humans and agents collaboratively edit documents, designs, and 3D scenes. The system incorporates hierarchical workspaces and persistent semantic memory to maintain continuity across agents and sessions, and demonstrates strong performance across 12 benchmarks and six application categories, outperforming leading agents like Codex and DeepSeek Harness.
By Jiajie Jin, Guanting Dong, Chengjun Pan, Xin Deng, Zhuohui Sheng, Guoqiang Wang, Tao Gui, Yutao Zhu, Yujia Chen, Shihong Yi, Binze Hu, Zhixiong Yang, Zhichao Sun, Rui Wang, Han Han, Mingyu Ma, Junhao Huang, Xiaotao Wang, Huicheng Jiang, Yuyang Hu, Xinyu Yang, Wanghan Xu, Yajie Yang, Zhecan Wang, Honglin Guo, Xiaoyu Xing, Xiaoqian Zhang, Yulun Zhu, Xiaoxi Li, Jiahang Lin, Shaofan Liu, Shihan Dou, Zhiheng Xi, Qi Zhang, Yu-Gang Jiang, See-Kiong Ng, Yang Zhang, Tat-Seng Chua, Ji-Rong Wen, Zhicheng Dou
The paper introduces ARCBench, a configurable benchmark for measuring rule compliance in language‑model agents that compact long histories to fit context windows. Using ARCBench, the authors find that rule violations after compaction rarely stem from rule deletion; instead, agents often ignore the rule while it remains present in the compacted history. The study shows that the source of violations depends on how the rule is delivered and on the compaction strategy, with targeted compactor attacks significantly increasing violations, and that simply retaining a rule in the compacted summary does not guarantee compliance.
By Yiwei Zha, Qingyu Meng, Herbert Bos, Min Chen
The paper introduces PRISM, a framework that separates belief estimation from action selection to handle irregular state transitions in online medical consultations. By maintaining a posterior belief over consultation states, PRISM improves clinical decision accuracy by 22.40 percentage points and response quality by 5.51 BERTScore_F1 points compared to leading LLMs and clinical agents. Ablation studies confirm that oracle state information further enhances response quality across baselines.
By Mingzhe Lu, Xiangru Chen, Jian Qin, Yunpeng Li, Xihan Gao
arXiv:2607.10645v3 Announce Type: replace
Abstract: An LLM agent's public behaviour reveals little about its social reasoning: an agent that votes correctly may be guessing, and its public claims may...
By Ilia Karpov, Maksim Grigorenko, Vadim Ayuyev
AtomMem is a long‑term memory system for large language models that stores high‑value atomic facts extracted from long‑form interactions. It organizes these facts into hierarchical event structures and temporal profiles, enabling coherent episodic context and dynamic tracking of user attributes. During retrieval, an associative memory graph connects fragmented memories, and experiments on the LoCoMo benchmark show state‑of‑the‑art performance across reasoning tasks.
By Yanyu Yao, Shangze Li, Zhi Zheng, Hui Zheng, Qi Liu, Tong Xu, Enhong Chen
The paper introduces Mobius Injection, a new attack that turns autonomous LLM agents into zombie nodes capable of launching Agent-based and -Oriented DDoS (AbO-DDoS) attacks. By combining indirect prompt injection with persistent agent components through a mechanism called Semantic Closure, a single injection event can trigger sustained recursive execution of agent components. Experiments on three claw-style agents and three coding agents across 12 LLMs show that Mobius Injection can amplify single-node calls up to 51×, drain 123k–127k tokens per node on a shared five‑GPU backend, and increase benign request latency by 23–24×, eventually saturating shared capacity and denying service to legitimate traffic.
By Zi Liang, Ronghua Li, Yanyun Wang, Qingqing Ye, Haibo Hu
The paper introduces EDV, an Execute‑Distill‑Verify framework that constructs offline experience for large language model agents without updating model parameters. In EDV, multiple heterogeneous agents execute tasks, a distiller extracts candidate lessons from their trajectories and feedback, and executors review these lessons, keeping only those unanimously approved. Experiments on τ²‑bench, Mind2Web, and MMTB demonstrate improved performance on revisited tasks and within‑domain transfer, while a Retail subset evaluation shows EDV cuts the Self‑Confirmation Trap Rate from 11.0% to 5.0%.
By Shiding Zhu, Yudi Qi, Yajie Wang, Jiaze Li, Chao Song, Yaorui Shi, Yibo Miao, Hanqi Gao, Kai Zhang