APEx is a hierarchical framework that organizes a deep research agent’s interaction history into instance-level trajectory memories and category-level procedural skills. It couples these through an Executor, Distiller, and Planner, trained with a three-stage alternating GRPO paradigm to enable reward-guided skill distillation. At test time, distilled skills act as procedural priors for online Planner adaptation via skill-guided reinforcement learning, achieving state‑of‑the‑art results on seven benchmarks, outperforming GPT‑5.4 by 14.7 points and the best memory‑augmented baseline by 3.0 points.
By Jie Ding, Rui Sun, Xinyuan Zhang, Zeyu Zhang, Xin Liu
The paper introduces AGENTSCOPE, a neuro‑symbolic method for diagnosing failures in large language model agents. It abstracts agent trajectories into structured representations and employs neural invariants to define behavior properties. Using LLM‑guided reasoning on these abstractions, AGENTSCOPE identifies both the failure step and its type, outperforming existing techniques on several datasets.
By Jiayi Bi, Yanjie Gao, Yuanmin Xie, Liqun Li, Tianyin Xu, Fan Yang, Mao Yang
Loom is a generative consensus framework designed for real‑world root‑cause analysis (RCA) that combines open‑form hypotheses from modular heuristics with a lightweight large language model (LLM) synthesis step. It projects hypotheses into a continuous embedding space and uses an iterative centroid‑based reweighting algorithm to resolve conflicts, producing a single consensus that is then synthesized by one LLM call. On the OpenRCA benchmark Loom matches state‑of‑the‑art autonomous agents on some datasets while achieving significantly higher efficiency—about 26× faster and 33× faster with an 8B‑parameter synthesizer.
whyItMatters":"Loom demonstrates how embedding‑space reweighting can bridge the gap between statistical rigor and expressive LLMs, enabling efficient, trustworthy RCA in industrial settings."
By Ron Begleiter, Katya Egert Berg, Gilad Saban, Gil Shabat
SafeEvolve is an experience-driven framework that co‑evolves a harness and policy to align large‑language‑model agents with safety goals. It uses completed on‑policy trajectories to update safety prompts and hierarchical skills, then applies a two‑stage SFT‑RL training loop that bootstraps the policy with the evolved harness and refines it through verifier‑augmented rewards. Experiments on agentic safety benchmarks show that SafeEvolve improves the safety‑utility tradeoff, achieving a three‑fold reduction in ASR on AgentDojo for Qwen3.5‑4B while increasing benign utility from 59.79% to 61.86%.
By Qinghua Mao, Wanying Qu, Dadi Guo, Leitao Yuan, Qingyu Liu, Yu Li, Guanxu Chen, Yanwei Fu, Xi Lin, Xia Hu, Dongrui Liu
The paper introduces Discriminative World Models for Web Agents, proposing a predicted-state matching objective that trains world models to produce representations that can distinguish the true resulting state from those of alternative actions. Using a branching dataset from WebArena Go-Browse, the authors demonstrate that this approach outperforms traditional supervised next-state prediction on a held‑out benchmark and improves action ranking on WebPRMBench. Additionally, employing the discriminative world model for test‑time action selection boosts end‑to‑end task success on WebArena‑Lite.
By Kelvin Li, Dhruv Pendharkar, Anish Pahilajani, Chuyi Shang, Leon Oks, Leonid Karlinsky, Rogerio Feris, Trevor Darrell, Roei Herzig
Git4Data introduces a database-native version‑control layer that treats a database as a repository and each table as a versioned object, exposing Git‑style operations—snapshot/tag, branch, diff, and merge—through SQL extensions. Implemented in MatrixOne, it leverages immutable object storage and MVCC so that operation costs depend on the size of the change rather than the entire dataset. In agentic branching workloads, Git4Data outperforms DoltDB by up to an order of magnitude, demonstrating efficient versioning for AI agents.
By Hongshen Gou, Zuyu Zhang, Yuze Sun, Peng Xu, Feng Tian, Long Wang, Jianguo Wang
The article proposes a foundational ontology, Activity Theory-based foundational ontology (ATFOt), to formally represent contradictions in dialogue-based human‑robot interactions. Using METHONTOLOGY and Activity Theory, the authors provide natural language, set‑theoretic, and first‑order logic definitions of dialogues and contradictions, and introduce three novel principles for guiding human‑robot dialogue. The work aims to create an interoperable framework applicable across HRI and human‑agent interaction domains.
By Maitreyee Tewari, Michele Persiani
The paper introduces the AI Mathematician (AIM) framework, which leverages Large Reasoning Models (LRMs) to tackle frontier mathematical research. AIM addresses the complexity and procedural rigor of research problems through an exploration mechanism for longer solution paths and a pessimistic reasonable verification method for reliability. Early experiments show AIM can autonomously construct significant proof components and uncover non‑trivial insights across real‑world mathematical topics.
By Yuanhang Liu, Yanxing Huang, Yanqiao Wang, Peng Li, Yang Liu
OptSkills is an archetype‑centric agent that learns and reasons about optimization problems using large language models. It clusters problems by underlying archetypes, explores diverse modeling and solver configurations within each cluster, and distills successful trajectories into reusable workflow‑level skills. The system achieves state‑of‑the‑art accuracy on multiple datasets, outperforming prior methods on challenging benchmarks such as MIPLIB‑NL and OOD NLCO.
By Haochen Yang, Ke Zhao, Mengyuan Ma, Xingyu Lu, Xiangfeng Wang, Hong Qian
The paper introduces Fetch.ai, an industrial-strength architecture that blends classical multi-agent system principles with modern AI capabilities. It features a decentralized foundation of on-chain blockchain services for identity, discovery, and transactions, a development framework for secure, interoperable agents, a cloud-based deployment platform, and an agent-native LLM that translates human goals into multi-agent workflows. A decentralized logistics use case demonstrates autonomous agents dynamically discovering, negotiating, and transacting securely.
By Michael J. Wooldridge, Attila Bagoly, Jonathan J. Ward, Emanuele La Malfa, Gabriel Paludo Licks
The paper examines how an agent’s probability report is evaluated twice—once by a strictly proper scoring rule and again by an approval rule that determines a decision. It shows that when the approval rule is welfare‑maximizing, it cannot be affine, yet the resulting distortion is predictable and can be mitigated by a reserve report that neutralizes the cost of pretending to be the marginal type. A Lipschitz rule with a single kink achieves first‑best welfare, while smooth rules cannot, and the key constraint is the steepness of the rule rather than its smoothness.
By Lauri Lov\'en, Sasu Tarkoma
DiDrive introduces a risk‑aware hierarchical diffusion framework for offline reinforcement learning in autonomous driving. It combines a low‑level risk‑gated encoder with a high‑level contextual modulator to filter redundant state information, and a 3DICE policy optimization that reduces out‑of‑distribution overestimation and stabilizes gradients. On the CARLA benchmark, DiDrive outperforms baselines such as IQL, CQL, and Diffusion‑QL, achieving an 85% success rate and a 4295.68 average reward in dense traffic with 60 vehicles.
By Qisong Guo, Jingtang Chen, Zhilin Chen, Pei Xu, Mingjian Fu, Wenxi Liu, Yuanlong Yu
The paper introduces the Private Computation Space (PCS), an open‑source federated learning system designed for agriculture that protects farmer data using asynchronous federated learning, differential privacy, and trusted execution environments. PCS runs on commodity hardware and is resilient to rural infrastructure challenges. In two real‑world deployments—nitrogen monitoring in New York and evapotranspiration prediction in California—PCS achieved a Dice Similarity Coefficient of 0.71 and an $R^2$ of 0.84, improving single‑site model accuracy by 22.4% and 9.1% respectively while preserving privacy.
By Shuangyu Lei, Muhammad Salman Abid, Jacob Belding, Sam Mosher, Manushi B. Trivedi, Shivranjani Baruah, Liam Wickes-Do, Andrew Anderson, Braulio Dumba, Alyssa Whitcraft, Ritvik Sahajpal, Sijin Li, Kelly Robbins, Michael Gore, Margaret Frank, Steven Wolf, Liz Jones, Abraham Stroock, Kaitlin Gold, Hakim Weatherspoon
RideSkill is a hierarchical algorithm for generalized ride sharing that uses large language models (LLMs) to automatically design and train a skill repository, a combiner, and a repositioner. The combiner assigns vehicle-specific skills for adaptive dispatch across varying scenarios and objectives, while the repositioner moves idle vehicles to emerging regions to avoid conflicts. By training all components via an LLM-based evolutionary method, RideSkill eliminates the need for real-time LLM calls, enabling high-performance deployment in large-scale systems.
By Zijian Zhao, Sen Li, Xialiang Tong, Mingxuan Yuan
The paper investigates privacy risks in agentic AI systems that assemble sensitive data into a hidden context before responding. It introduces context‑inference attacks, a security game that evaluates how well attackers can recover this hidden context under varying levels of knowledge and indirect delivery. Experiments show that even with controls such as instructions not to disclose, logit suppression, and context dilution, agents can leak significant contextual information, achieving high success rates across multiple attack settings.
By Prince Jha, Samuele Poppi, Nils Lukas
The paper presents a tri‑agent framework for evaluating large language models’ question‑clarification abilities. It involves a Question Clarifying Agent that identifies ambiguities and asks follow‑up questions, a Respondent Agent that simulates human replies, and an Evaluator Agent that judges the dialogue using metrics such as ambiguity handling, question quality, dialogue efficiency, language appropriateness, and intent alignment. The authors illustrate the approach with synthetic supply‑chain data and discuss validating the evaluator against human judgments.
By Yikai Zhao, Saurabh Pandey, Pradeep Kumar Misra
HyGRAIL is a framework for discovering scientific hypotheses in incomplete knowledge graphs by combining a graph neural network (GNN) triage with large language model (LLM) review. The GNN scores candidate hypotheses and routes only ambiguous cases to the LLM, which receives structured evidence from the graph converted into natural language. Experiments on MatKG show HyGRAIL achieves the highest F1 score, improves over baselines, and cuts LLM calls by over 54%.
By Yihang Sun, Zhihan Zhu, Zhiyuan Jiang, Jingyi Ge, Zixuan Li, Jiaxuan You
The paper introduces RuVerBench, a benchmark with 2,458 instances for evaluating the reliability of Large Language Models acting as judges (LaaJ) in verifying rubric compliance within agentic scenarios such as deep research and agentic coding. It systematically meta‑evaluates frontier LLMs, revealing that even the most advanced models perform well yet still produce substantial noise. The study also examines how prompt design, batching, and majority voting affect verification accuracy, noting that weaker models are more prompt‑sensitive, batched verification trades accuracy for efficiency, and majority voting offers diminishing returns.
By Yangda Peng, Yunjia Qi, Haotian Xia, Guanzhong He, Xintong Shi, Richeng Xuan, Songyuanyi Lu, Yixian Liu, Zhichao Hu, Yuhong Liu, Hao Peng
PGMem introduces a persona‑memory graph that tightly couples user persona signals with the events that support or revise them, using typed provenance and evidence edges. This design addresses gaps in existing memory systems where personas are stored as flat, detached profiles. Experiments on three benchmarks show that PGMem outperforms various baseline memory approaches, especially as dialogue context length increases.
By Wonjun Choi, Yerim Kim, Yukyung Lee, Susik Yoon
The paper investigates Joint-Embedding Predictive World Models (JEPA-WMs), a class of methods that perform planning in a learned representation space rather than raw input space. It systematically studies how model architecture, training objectives, and planning algorithms influence success across simulated and real‑world robotic tasks, and proposes a JEPA-WM variant that surpasses established baselines in navigation and manipulation. The authors provide code, data, and checkpoints for reproducibility.
By Basile Terver, Tsung-Yen Yang, Jean Ponce, Adrien Bardes, Yann LeCun