The paper introduces LRE (Learned Relevance Eviction), a lightweight, CPU‑only, language‑model‑free scorer that learns which parts of an agent’s interaction history are task‑critical and preserves them verbatim. In experiments, LRE matches or surpasses baseline eviction policies on accuracy‑cost trade‑offs, recovers 93% of full‑history accuracy, reduces worst‑case prompt size by 52%, and outperforms dense and token‑pruning encoders in conversational memory while being 295–1569× smaller. The method also achieves superior budgeted answer quality on LoCoMo reading and can be trained annotation‑free, recovering 95% of supervised scorer performance.
By Nusrat Jahan Lia, Aritra Mazumder
WorldReward introduces a vision‑language model–based reward system for camera‑conditioned world models, combining action consistency and visual quality evaluation. It processes paired videos by splitting them into action‑aligned chunks, structuring visual evidence, and aggregating decisions through voting. The model is trained on a large, reasoning‑augmented preference dataset and outperforms GPT‑5.5 on a human‑annotated benchmark, improving both action execution and visual quality when applied to RL post‑training.
By Yibin Wang, Zehan Wang, Junshu Tang, Zhimin Li, Yujie Zhou, Jiazi Bu, Pengyang Ling, Feng Han, Zhixiong Zhang, Long Xing, Shengyuan Ding, Ziang Li, Cheng Jin, Yuhang Zang, Jiaqi Wang, Tianyu Pang
The paper introduces HARNESSEVO, a method that decomposes a large language model’s harness into four independently evolvable components—role, task‑strategy, tool/format‑rules, and reflection/control. Experiments on ALFWorld show that overall success rates are similar to flat‑string evolution, but the reflection/control component alone accounts for most of the performance gains. The study also finds that evenly distributing optimization budget across all slots can be detrimental; concentrating resources on the high‑credit control slot recovers lost performance, while on WebShop all slots remain ineffective, suggesting task‑specific differences in harness value.
By Michael Nguyen, Wei Chen Tan, Nurul Aisyah Hassan, Arvind Raman, Li Hua Lim, Ahmad Faiz Razak
The paper introduces CONFLICTGUI, a benchmark that tests GUI agents on instruction-internal and instruction‑GUI context conflicts, revealing that many agents over‑comply with infeasible instructions. To address this, the authors propose CONFLICTGUARD, an inference‑time framework that couples a feasibility verification protocol with a conditional action modulation mechanism, enabling agents to assess instruction logic and GUI evidence before acting. Experiments on five agents show that CONFLICTGUARD significantly improves conflict‑task success while maintaining normal task performance.
By Zhaoyuan Huang, Tianjie Ju, Pengzhou Cheng, Zheng Wu, Yansi Li, Chuanbiao Song, Jun Lan, Huijia Zhu, Weiqiang Wang, Zhuosheng Zhang
The Bioinfoysis Technical Report introduces a multi‑agent harness designed to improve long‑horizon bioinformatics tasks by maintaining persistent, artifact‑grounded analysis runs. It combines global planning with step‑wise, evidence‑driven replanning, ensuring intermediate results are tied to responsible agents and preventing stale evidence reuse. The system was evaluated on BixBench and LAB‑Bench 2, achieving state‑of‑the‑art accuracy and demonstrating that reliable bioinformatics automation relies on robust planning, execution, memory, and evidence flow.
By Qingyang Shao, Xin Zhang, Zhouyang Yuan, Xianying Chen, Yujia Xiang, Zihao Yang, Tong Ye, Yangqi Zhang, Jiakang Xu, Xiaoqing Yan, Xuan Luo, Keyi Li, Enci Fan, Kai Kang, Zhuohan Liu, Xingyu Jin, Chunran Teng, Tao Li, Xinyu Lv, Minghui Wang, Wenfeng Li, Yidan Gao, Siyu Liu, Mingrui Luo, Zhu Liang, Guanren Qiao, Zhiping Xu
RealCADBench is a new benchmark for evaluating intent‑to‑program parametric CAD modeling, featuring 12,632 tasks drawn from 19 factory‑automation categories and covering text, 2D drawings, product photos, and rendered images for both Part and Assembly modeling. The study reports results on a 1,770‑task evaluation slice, using metrics such as executability, Solid IoU, Surface IoU, and a rubric‑based visual‑semantic identity Judge. Across nine standalone and six frontier‑scale large models, no single model dominates all four metrics, highlighting diverse strengths and failure modes like missing fine structures and incorrect assembly placement.
By JoyIndustrial VisCAD Team, Linxin Cai, Qiuhe Hong, Zhichao Huang, Guanlin Li, Zongzhen Li, Hongsen Liu, Yichen Long, Wei Wang, Yuchen Wang, Dongyue Yang, Huimu Yu, Xianwen Zhong
The paper presents a framework that uses saliency maps to create hierarchical attention profiles, tracking how deep reinforcement learning agents allocate attention over time. By comparing these attention trajectories across different conditions and linking them to behavioral metrics, the study reveals algorithm‑specific biases, unintended reward‑driven strategies, and overfitting to redundant sensory inputs. Experiments on Atari 2600 games, custom Pong environments, and biomechanical visuomotor simulations demonstrate that these attention patterns correspond to measurable behavioral differences, establishing attention trajectories as a diagnostic tool beyond traditional performance metrics.
By Charlotte Beylier, Hannah Selder, Arthur Fleig, Simon M. Hofmann, Nico Scherf
CauseCollab is a causal unified and modality‑agnostic network designed to improve collaborative perception across heterogeneous sensor modalities. It disentangles semantic factors from modality‑specific confounders using causal metric learning and employs a context‑guided Unified Converter to maintain cross‑modal semantic consistency. The approach requires only minimal adapter training when adding new modalities and achieves state‑of‑the‑art results on the OPV2V and DAIR‑V2X datasets, especially in scenarios with large modality gaps.
By Weize Li, Yang Li, Quan Yuan, Xiaoyuan Fu, Guiyang Luo, Jinglin Li
Semantic Bayesian World Models (SBWMs) propose a shift from static knowledge graphs to a dynamic, probabilistic fabric of beliefs that can be updated via Bayesian conditioning and influenced by actions. The approach aims to bridge the gap between crisp factual assertions and the probabilistic reasoning of foundation models and autonomous agents, enabling richer inference in scenarios such as home‑security decisions, actuarial estimates, and planning tasks. Realizing SBWMs requires new tools for belief annotation, probabilistic entailment, semantic calibration, and protocols for belief exchange among agents.
By Tommaso Soru
The paper presents a graph‑constrained agentic framework that enables large language models to adapt retail supply‑chain decision modules to evolving requirements. It jointly selects intervention routes and admissible module changes, validating candidates against downstream KPIs. Experiments with 100 warehouse requirements and three LLMs show the framework improves end‑to‑end success from 72–76% to 79–83%.
By Lei Zheng, Liping Yang, Zihao Li, Guodong Lyu, Chaik Ming Koh, Chung-Piaw Teo
DRACO introduces a method for fine‑grained credit assignment in long‑horizon reinforcement learning tasks that lack verifiable rewards. It dynamically generates multi‑criteria rubrics during training, scores them once per trajectory, and redistributes the resulting judgment over the steps responsible for each rubric to produce differentiated per‑step advantages. Experiments on AppWorld and Tau‑Bench show significant performance gains over baseline models and other rubric‑based approaches, even without using any verifiers.
By Shubham Gandhi, Saurabh Goyal, Kiran Kate, Yara Rizk
The paper introduces PACE, a dataset that tests whether personalized assistants can detect hidden conflicts in user requests by integrating implicit, egocentric knowledge from a knowledge base. PACE pairs user requests with persona-specific facts, requiring models to retrieve contextual evidence before deciding if a request is inappropriate. The authors also propose PaceMaker, a multi-agent system that improves conflict detection by coordinating query reformulation, multi-hop graph traversal, and conflict-aware filtering, outperforming existing methods on the PACE benchmark.
By Yoojin Kim, Jihyoung Jang, Hyounghun Kim
The paper introduces a computable representation of the physical laboratory that combines typed research objects, capability‑bound operations, and a compositional workflow algebra. This framework allows scientific workflows to be expressed as programs that track evolving laboratory states, including explicit dependencies, decisions, iteration, and concurrency. Implemented in a modular agentic robotic laboratory, the system binds formal operations to executable Function Skills, generates capability‑relative workflows for diverse scientific goals, and uses stateful simulation to verify operation preconditions and laboratory constraints before execution.
By Xiaobo Li, Luyao Ge, Xiaohui Li, Lulu Guo, Ming Mao, Jiwang Zheng, Wenting Guan, Xin Yang, Yi Luo, Jun Jiang, Linjiang Chen
The paper introduces Test-Time Adaptation through Human‑Agent Interaction (TAHI), a method that uses iterative human feedback to adapt AI agents to individual users’ criteria. By integrating cross‑session interaction data into agent context and weights, and building an evolving rubric module, the authors demonstrate that agents can improve task success by 4.5–20.9% after only a few interactions. The evolving rubric also serves as a scalable annotation tool, detecting 16.0–22.3% more failures than language models or humans alone, and personalized agents can even generalize improvements up to 8.8% across users.
By Zora Zhiruo Wang, Apurva Gandhi, Rulin Shao, Aspen Chen, Jonas Mueller, Zhiqi Liang, Jett Chen, Michael Ryan, Qianou Ma, Luxi He, Zhoujun Cheng, Andre He, Seungone Kim, Jiayi Geng, Mingqian Zheng, Weiwei Sun, Zheyuan Zhang, Xinran Zhao, Yike Wang, Abe Hou, Liwei Jiang, Pang Wei Koh, Diyi Yang, Graham Neubig, Daniel Fried
The paper introduces a method to detect which web scrapers feed data to large language models (LLMs) by deploying dynamic websites that issue unique canary tokens to each scraper. By querying LLMs for information about these sites, the authors can identify when an LLM consistently outputs the unique tokens, indicating exposure to a specific scraper. Experiments on 22 production LLM systems show the technique reliably uncovers both known and undisclosed scrapers, offering a tool for third parties to monitor and control unwanted web scraping.
By Steven Seiden, Triss Ren, Caroline Zhang, Taein Kim, Enze Liu, Emily Wenger
RuleMem is a rule-based memory framework designed for long-term conversational agents. It generates natural-language Horn clauses from dialogue histories and validates them with a Rule Perplexity Consistency mechanism, enabling active guidance of evidence retrieval and reasoning. In evaluations on LoCoMo and LongMemEval_s*, RuleMem outperformed 14 baselines, achieving the highest accuracy on LoCoMo with a 27.47‑point absolute gain (54.3% relative improvement).
By Xingyuan Zeng, Zuohan Wu, Quanming Yao, Yue Wang, Wei Liu, Libin Zheng, Jiuke Wang, Jian Yin
SLIDEFORGE is a new AI agent designed for controllable editing of presentation slides. It constructs a Deck State Graph that links visual decomposition, native PowerPoint object structure, and perceptual organization, enabling theme‑preserving reconstruction through slide‑native operations and rendered‑state verification. The authors also propose an evaluation framework that jointly measures component recovery, preservation, restyling consistency, visual quality, and native editability, and demonstrate that SLIDEFORGE outperforms existing prompting, screenshot‑based, and generic code‑agent baselines.
By Haozhen Zheng, Fulin Wang, Tianhu Xiong, Yingjie Yu, Shengyi Qian, Hanchao Yu, Alex Schwing, Klara Nahrstedt, Mingyuan Wu
VeriPhy is an auditable physical‑verification system that transforms a text prompt into typed physical obligations and a statically validated execution plan before any video frames are generated. During execution, it gates calls to frozen low‑level experts (segmentation, tracking, counting, depth, OCR, audio‑event detection, etc.) and records provenance‑carrying evidence for each action. The system maps these records to a three‑valued state—supported, contradicted, or unknown—providing traceable verdicts that can be used to refine generation models.
By Wenzhuo Xu, Yuchen Zhu, Chongjian Ge, Xuan Shen, Jing Shi, Jason Kuen, Yongxin Chen, Molei Tao, Christopher McComb, Noelia Grande Guti\'errez, Jiuxiang Gu
KC-Bench is a dynamic interactive benchmark designed to evaluate how large language model agents reconcile user instructions, internal knowledge, and real‑time environmental observations. It contains 238 manually curated multi‑turn tasks that test world‑knowledge conflicts, input inconsistencies, and multi‑source temporal conflicts, using a user simulator, stateful tools, deterministic environment assertions, an open‑source natural‑language evaluator, and human trajectory verification. Evaluation of nine models—including DeepSeek‑V4‑Flash, GLM‑5.2, and MiniMax‑M3—reveals significant cross‑domain variation, with no model reliably handling factual correction, identity consistency checking, and temporal conflict resolution across all settings, and shows that missed conflicts can propagate to tool calls or synthetic protected‑data flows.
By Yaxing Lyu, Shengjie Zhou, Binbin Toh, Pengyu Zhu, Lijun Li
Terminal-Universe is a framework that transforms large collections of terminal‑based agent trajectories into reusable, executable environments. By replaying recorded file operations and completing missing files, it reconstructs the original workspace and task, then synthesizes new tasks and multi‑round interactions. The resulting 37.3k task‑sufficient environments enable significant performance gains when fine‑tuning language models on terminal‑based benchmarks.
By Jie Wu, Zhenru Zhang, Beichen Zhang, Xuwu Wang, Yuhui Su, Mouxiang Chen, Peng Wang, Zhihai Wang, Que Shen, Hao Zhou, An Yang, Fei Huang, Yujiu Yang, Dayiheng Liu