The paper introduces a minimal‑data adaptation regime where a pretrained robot policy learns a new task from a single demonstration followed by autonomous online interaction. It shows that while zero‑shot performance is poor, one demonstration can guide the policy toward coarse task‑following, and identifies key design principles—pretrained VLA representations for sample‑efficient value learning, residual policy learning for corrective behavior, and data‑balancing to avoid exploration collapse. The proposed MiDAS method, combining few‑demonstration behavior cloning with value‑based residual learning over frozen VLA representations, significantly boosts task success on LIBERO, RoboCasa, and a bimanual YAM platform after only a few hours of online interaction.
By Shreyas Kowshik, Sreyas Venkataraman, Leo Wang, Niharika Pant, Max Simchowitz, Aviral Kumar
MemPlan is a memory‑conditioned PDDL planning framework designed for partially observable interactive text environments. It keeps the PDDL domain fixed and reconstructs the problem at each step using parsed observations, candidate hypotheses, and interaction memory, employing positive graph memory to assign relevance‑based action costs and negative memory to record refuted assumptions as exclusion constraints. Experiments on TextWorld, ALFWorld, and Robotouille demonstrate that MemPlan improves task completion and step efficiency over language‑only and PDDL‑aware baselines while significantly reducing online token cost compared with LLM‑planning baselines such as ReAct, and its fine‑tuned interface generalizes across task types without task‑specific fine‑tuning.
By Tengwei Song, Robert Hoehndorf
The paper identifies a "specification-oracle gap" in coding‑agent benchmarks, where tests rely on details (constants, module paths, exact messages) that are not stated in the natural‑language task description. The authors formalise this gap, develop a static audit tool that classifies such dependencies across five languages, and demonstrate that many benchmarks (e.g., SWE‑Milestone) contain tasks whose tests depend on unstated information, leading to higher failure rates for agents that do not reproduce upstream implementation details. The study also shows that the prevalence of these gaps varies with benchmark construction and that removing affected tests does not alter leaderboard rankings but shifts scores within repositories, favoring agents that replicate upstream names.
whyItMatters":"The findings reveal that benchmark tests can inadvertently measure an agent’s ability to mimic upstream implementation details rather than its true problem‑solving skills, potentially skewing evaluations and comparisons across agents."
By Fei Ding
UniGuard introduces a training framework that learns interface‑invariant safety representations by leveraging multiple semantically aligned views of the same safety case. The authors created the UniGuard‑FT dataset, transforming 1,000 public safety cases into prompt, response, multi‑turn conversation, agent‑trace, and multi‑agent‑trace formats, and trained a 0.6B model using classification, supervised contrastive learning, and asymmetric confidence calibration losses. On unseen‑interface benchmarks (ATBench, PolicyGuardBench, DynaBench), UniGuard‑0.6B outperforms its pretrained guardrail by an average of 45.12 F1 points and reduces expected calibration error by 35.41 points, while remaining competitive with much larger guardrails such as Claude‑Sonnet‑4.5 and GPT‑5.2.
By Viet-Thanh Pham, Thuy-Trang Vu, Junae Kim, Rollin Omari, Dinh Phung, Gholamreza Haffari, Lizhen Qu
ReflectPO introduces a reinforcement learning framework that enables large language model agents to recover from errors during long-horizon tasks involving external tools. The method uses Reflection-Guided Sampling to generate recovery hints from branch differences and repeated visits, and Tree-Structured Credit Assignment to allocate action-level credit based on parent–child reward differences and recovery gains. Experiments on two tool-use benchmarks demonstrate that ReflectPO outperforms existing baselines and improves agents’ ability to detect and correct deviations, thereby advancing autonomous error recovery in complex real-world environments.
By Rumei Li, Xunliang Cai
IQuest‑Q1 is a 320 B sparse Mixture‑of‑Experts foundation model that activates 15 B parameters, designed to enhance agentic command‑line interface systems. Its development follows a three‑stage pipeline: stable pre‑training with advanced normalization and curriculum techniques, fine‑grained mid‑training that expands data domains and context length, and multi‑scaffold post‑training that blends supervised fine‑tuning, domain‑specialized reinforcement learning, and multi‑objective policy distillation. The model merges experts from different stages to reduce parameter conflicts, achieving strong coding and general task performance, and is released as open‑source to aid community research.
By A. Jian Yang, An Huang, Bo Dan, Chen Mu, Chengran Yang, Chuan Shi, Feng Chang, Hongxin Li, Hongyuan Tao, Jinxing Zhang, Jujie He, Kun Wang, Lin Jing, Ming Yang, Qicheng Hu, Renyuan Li, Shark Liu, Shawn Guo, Shi Qing, Shi Zhong, Tianyu Zheng, Wei Liu, Wei Zhang, Wenrui Cai, Wenxiao Chen, Xuxing Chen, Yan Xing, Yi Xing, Yifan Luo, Yiran Xing, Yizhi Li, Yongsheng Kang, Yucheng Ding, Yue Song, Yun Qing, Yunqi Liu, Zelong Huang, Zhengmao Ye, Zhentao Fan, Zhiteng Li, Ziling Huang, Bihong Tang, Bingli Wang, Ce Jun, Changhao Chen, Chen Hu, Chen Xiao, Chi Liu, Chunjiang Ge, Danyang Li, Deyang Kong, Dian Qi, Ding Tan, Evan Yang, Fanglin Xu, Fei Bai, Gang Liu, Guochang Li, Hao Tang, Haoran Feng, Haowen Wang, Huahao Hu, Huanjin Yao, Huatong Song, Jia Deng, Jiajun Wu, Jialiang Cheng, Jiayi Liang, Jiayi Wu, Jie Deng, Jiejun Tan, Jing Yang, Jingda Xu, Junhang Cheng, Junxiao Liu, Kai Lv, Kaixiang Pan, Kecheng Lu, Kun Zeng, Kun Zheng, Kunbo Ding, Liang Qiu, Lingyi Sun, Luan Zhang, Mengxiao Lv, Ning Dong, Pei Wang, Peng Zhang, Qingsong Cai, Qixiang Yin, Qiyao Wang, Renbiao Liu, Renfei Dang, Ruihang Li, Ruirui Huang, Shanhui Yin, Shen Jian, Shengjie Fang, Shikai Li, Shinan Zhang, Shiqi Chen, Shuai Liu, Shuangqi Li, Shuo Tang, Shuqing Deng, Sijin Hao, Siwei Wu, Siyue Wang, Tao Hai, Tao Ren, TingYang Chen, Wei Shao, Weicheng Gu, Weihan Peng, Wenlong Miao, Xavier Yu, Xiangyuan Ji, Xiaocheng Li, Xiaodong Huang, Xiaohan Ge, Xiaoxiang Shi, Xiaoyang Zhang, Xiaoyu Yang, Xidong Wang, Xin Li, Xinyu Tang, Xiyuan Yang, Xudong Ma, Ya Tong, Yachao Wang, Yang Tian, Yaxin Du, Yeli Ruan, Yi Wang, Yi Yu, Yibin Feng, Yibo Tan, Yihang Wang, Yiming Ju, Yinda Chen, Yinda Chen, Yingjie Du, Yiwei Liu, Yizhi Zhou, Yuan Wei, Yuli Chen, Yuqi Zhu, Yuwen Li, Yuxi Dong, Yuxuan Zhang, Yuyang Song, Zhange Zhang, Zhaohan Zhang, Zhaoyi Li, Zhen Zhang, Zhengbo Zhang, Zhihang Hu, Zhiwei Yu, Zhixin Xie, Zhiyu Chen, Zhizhao Luo, Zhongyuan Wu, Zhou Fang, Zhu Lin, Zhuocheng Gong, Zixi Zhao, Ziyang Zeng, Ziyi Wang, Ziyu Zhao, Ran Tao, Mingjie Tang, Chuan Hao, Peihao Wu, Jianzhou Wang, Nan Jiang, Yuan Zhi, Bryan Dai
CoWriter is a proactive agentic system that adapts expert human co‑authoring methodologies from cognitive science to guide the entire writing workflow and actively encourage user engagement. In an exploratory user study, authors using CoWriter expressed their perspectives more clearly, felt greater ownership of the document, and engaged with the agent 2.2× more often than a standard LLM baseline. External reader judges also rated CoWriter‑generated texts higher in human‑touch qualities such as storytelling and personal experience integration.
By Shmuel Amar, Benjamin Lieber, Ori Shapira, Debanjana Kar, Sebastian Maximilian Dennerlein, Yufang Hou, Ido Dagan
The paper investigates how large language model agents can inadvertently leak confidential information by combining data from multiple documents—a phenomenon termed the "mosaic effect." Using 30 synthetic cybersecurity, medical, and military scenarios, the authors generate 10,800 simulated and 1,440 live answers from three LLM agents, finding that even the best model discloses forbidden composites in over 7% of simulated responses. While visible classification headers and aggregation rules reduce disclosure rates, unsafe disclosures remain in roughly one third of simulated answers, and increased refusals do not fully prevent leakage of prohibited conclusions.
By Daniel Ruiz, Luca Luceri, Emilio Ferrara
The paper introduces Dependency‑Tethered Memory (Tether), a memory system for tool‑using LLM agents that records the exact versions of premises each verdict depends on. By withdrawing any verdict whose support has failed anywhere in the store, Tether reduces unsafe publications to 0.3% and cuts verification calls by 37% compared to naive fact memory, while maintaining similar recall. The study demonstrates that capturing dependencies is crucial for safe memory use, as weaker writers or delayed relation arrivals increase unsafe publication rates.
By Yiqi Wang, Jinqian Ju, Jiaqi Zhang, Zhangkai Wu, Yiqun Duan, Zequn Sun, Taotao Cai, Mingkai Zheng
HiSTEVE 1.0 is a hierarchical VLA agent designed to operate in open‑world games like Minecraft, combining long‑horizon autonomy with native interface control. The system uses a capability‑aware harness with structured memory, a VLM planner, and a policy‑coupled task completion detector to orchestrate subtasks, while a single VLA handles vision‑only perception and native keyboard‑mouse actions. HiSTEVE achieves state‑of‑the‑art performance on atomic tasks and successfully completes long‑horizon tasks such as crafting an iron pickaxe from scratch.
By Zecong Tang, Zian Guan, Yuncheng Guo, Zhanqiu Zhang, Yiwen Guo, Bo Dai
The paper introduces a sparse topometric graph for indoor navigation that encodes local 3D geometry in nodes and distance‑weighted edges for traversable connectivity. A learned graph‑matching network aligns subgraph structures across different sessions, enabling robust point‑goal navigation even with scene changes, sensor variations, or limited overlap. Experiments on Gibson and 3D‑FRONT datasets show higher success rates and a memory reduction of over two orders of magnitude compared to point‑level maps.
By Eun Sun Lee, Minkyun Seo, Gwangtak Bae, Junho Kim, Seunggu Kang, Jaesik Park, Young Min Kim
Anthropic has announced that it will donate the Agent Skills standard to the Agentic AI Foundation (AAIF), a Linux Foundation‑backed organization co‑founded with Block and OpenAI. The move follows Anthropic’s earlier declaration that skills would be an open standard and aligns with the AAIF’s recent role as the home of the Model Context Protocol (MCP). The donation is part of Anthropic’s commitment to keeping Agent Skills community‑driven and vendor‑neutral.
The paper investigates how a large language model can act as a corpus debugger for the Slovak portion of the multilingual FineWeb2 dataset. By using a tool‑using agent that explores the corpus via a read‑only, budgeted API, the authors identified seven issues across four phenomena, including a large machine‑generated content farm that slipped through the pipeline’s filters. The agent’s compiled test suite can scan the entire 26.5 M‑document corpus in 0.7 s without a GPU, providing precise prevalence estimates and actionable insights into pipeline failures.
The paper presents a method for training small language models to select appropriate tools for complex, multi-step user instructions, addressing the challenge of handling large tool corpora. By using reinforcement learning, a Qwen3-8B model is trained to perform tool selection, achieving performance that surpasses a 402‑billion‑parameter Llama‑4‑Maverick model on the ToolBench benchmark. The trained model also demonstrates transferability to different retrievers and generalization across new tool domains.
Flint‑interactive is a language developed by Microsoft Research that enables AI agents to generate interactive data visualizations. It allows users to click on charts, filter data, highlight related content, and explore beyond the initially chosen view. The system is designed to enhance user interaction with visual data representations.
By Chenglong Wang, Yuwei Xiao, Alper Sarikaya, Michel Galley, Chris Quirk, Jianfeng Gao
The article reports on an experiment where an AI agent was allowed to search for evidence whenever it chose. The agent discovered some papers that a standard search missed, but also lost some that the plain search had found, and overall the gains were insufficient to be trusted. The post examines whether agentic Retrieval-Augmented Generation (RAG) truly retrieves better evidence across 298 questions.
By Abdullahi Dattijo
RIDAR is a framework designed to detect and mitigate research integrity violations (RIVs) in autonomous research agents. It introduces RIVT, a taxonomy of 14 RIV types, and offers two components: RIDAR-post-hoc, which identifies RIVs in agent trajectories with 85.2% F1 on a benchmark, and RIDAR-runtime, which dynamically corrects RIVs during execution, reducing remaining violations by 36.8%. The authors applied RIDAR to large datasets, uncovering 3,964 RIV instances in 99.7% of 300 agent trajectories across 50 tasks, highlighting the prevalence of RIVs in current agents.
Software agents are being trained to exceed human engineers; we argue that capability alone does not produce better software engineering. We define superintelligent software engineering as the discipl...
The study addresses the challenge of limited tumour visibility in intra‑operative CT images for liver hepatocellular carcinoma (HCC) surgery. It proposes a weakly supervised approach that transfers tumour appearance knowledge from pre‑operative MR images to guide localisation on intra‑operative CT, using a classifier and reinforcement learning. Experiments on an external HCC cohort show that synthetic transfer—converting CT to synthetic MR before localisation—outperforms other methods and CT‑only baselines, enabling cross‑modality tumour localisation without paired MR‑CT training data.