The paper introduces SetwiseEvalKit, a benchmark with nine dimensions and 28K rubrics for evaluating document sets, highlighting that current rerankers only achieve up to 45% coverage and struggle with cross‑document coordination. It proposes Rubric4Setwise, a training‑free method that transforms rubric criteria into selection signals, outperforming other methods in downstream generation tasks across both short‑form and long‑form scenarios. The study demonstrates the importance of evaluating and optimizing document sets as a whole rather than independently.
By Kailin Jiang, Lei Liu, Jian Xi, Hui Xu, Junlin Liu, Baochen Fu, Bin Li, Vichwang, Yu Lu, Haibo Shi
The paper introduces ORACLE‑SWE, a study that systematically evaluates how different intermediate information signals—such as Reproduction Test, Regression Test, Edit Location, Execution Context, and API Usage—contribute to the success of language‑model agents on software engineering tasks. It quantifies the impact of both oracle‑provided signals and those extracted by agents, and measures the performance gap between current agent‑found signals and their oracle counterparts. The authors release code and a dataset to support future signal‑analysis research in this area.
By Kenan Li, Qirui Jin, Liao Zhu, Xiaosong Huang, Yijia Wu, Yikai Zhang, Xin Zhang, Zijian Jin, Yufan Huang, Elsie Nallipogu, Chaoyun Zhang, Yu Kang, Saravan Rajmohan, Qingwei Lin, Wenke Lee, Dongmei Zhang
CertEx introduces a GPU‑based exact k‑nearest‑neighbor search that relies on certified exclusion, a technique that conservatively eliminates vectors using low‑precision bounds before computing exact distances. By storing a half‑precision lower‑bounding projection and embedding a rounding certificate into the tensor‑core dot product, CertEx only evaluates distances for a small subset of candidates, guaranteeing exact results. Experiments on sixteen datasets with up to 100 million vectors show that CertEx achieves 7.1× to 18.0× the throughput of the fastest cuVS search on an H100 and outperforms quantized and graph indexes when exactness is required.
By Dongfang Zhao
GRASE is a new agentic framework for converting real-world video into editable 3D scenes while preserving geometry and layout. It treats captured multi‑view video as a persistent reference, revising intermediate hypotheses whenever rendered evidence conflicts with the capture. By iteratively refining architectural shells, asset placement, and geometry with cross‑view feedback, GRASE produces faithful, consistent scenes that outperform existing state‑of‑the‑art methods.
By Zhaorui Lin, Haohan Liu, Zhen Chen, Xiangru Huang
OpenGCUA is an open‑stack project that aims to make generalist computer‑use agents (GCUs) fully accessible to the research community. The first release focuses on code execution, allowing agents to use intermediate results and programming skills directly in computer interaction, and introduces harness‑to‑model decision transfer to shift observation and task‑state decisions from fixed rules to the model. On the OSWorld‑v2 benchmark, the full configuration boosts task scores for all eight evaluated models—including GPT‑6 Astra and Claude Opus 5—by an average of +19.6 partial‑score points and cuts steps per task by 53 % compared to batched GUI‑only execution, while also providing a real‑device harness and cross‑device SDK.
whyItMatters":"By releasing the full stack, OpenGCUA enables systematic study and broader experimentation with generalist computer‑use agents, addressing the current limitation of closed implementations and demonstrating significant performance gains across diverse models."
By Hanzhang Zhou, Long Li, Quyu Kong, Lei Wang, Panrong Tong, Long Chen, Tianyu Xia, Gongjie Zhang, Xu Zhang, Chen Liu, Jianan Zhang, Chenglin Cai, Zhengwen Zeng, Steven Hoi, Yue Wang
Line-Augmented cuVSLAM extends the CUDA-accelerated visual‑SLAM system cuVSLAM by incorporating line features to improve visual odometry in low‑texture environments such as warehouses and production facilities. Lines are represented using Plücker coordinates for transport and a minimal 4‑DOF orthonormal form for optimisation, and a Cauchy‑robust point‑to‑line residual is integrated into the per‑frame Gauss‑Newton pose solve and sliding‑window bundle adjustment. The adaptive weighting allows line features to take on more of the estimation burden when point features are scarce, and the extension adds residual blocks to existing solvers rather than creating a separate estimator.
By Dara Molloy, Terry Wu, Wonju Lee, Abhilash Nand Kumar, David Adley, Conor Ferguson, Hesam Rabeti
GRID‑Planner introduces a graph‑structured relational reward shaping framework for diffusion‑based autonomous driving planning. By integrating graph‑aware importance weighting, a time‑indexed relational reward surrogate, and an advantage‑weighted diffusion objective with on‑policy self‑distillation, it forms a rollout‑return‑advantage‑update loop that eliminates explicit likelihoods. Experiments on nuPlan and interPlan demonstrate significant reductions in collision rate, lane deviation, and jerk, along with a 3.4% improvement in overall planning score.
By Zhao Yang, Zezhong Qian, Yinan Shi, Longjun Liu, Yawei Jueluo
SceneReVis is a vision‑grounded, self‑reflective framework for 3D indoor scene synthesis that uses an iterative diagnose‑and‑act loop to detect and fix spatial conflicts such as collisions. It introduces SceneChain‑12k, a large dataset of causal construction trajectories generated via a reverse engineering pipeline, and employs a two‑stage training recipe that moves from supervised fine‑tuning to agentic reinforcement learning, turning the model into an active spatial planner. Experiments show that SceneReVis achieves state‑of‑the‑art performance in high‑fidelity generation and goal‑oriented optimization, with better transfer to underrepresented room categories and potential for out‑of‑distribution room types.
By Yang Zhao, Shizhao Sun, Meisheng Zhang, Yingdong Shi, Xubo Yang, Jiang Bian
RSIgen is a task‑centric reciprocal self‑improvement framework for agentic visual generation that coordinates learning allocation, supervision refresh, and recursive curriculum construction. It employs Dynamic Responsibility Routing (DRR) to match interventions on evidence, inputs, and actions, selecting learning recipients and building compatible training targets. After generator updates, RSIgen re‑evaluates action utility with fixed branch inputs to refresh agent supervision, then constructs the next curriculum, yielding measurable gains on benchmarks and real‑user evaluations.
By Yidi Liu, Haozhe Wang, Jie Xiao, Kai Zhu, Xueyang Fu, Zheng-Jun Zha
The paper introduces TPMDM, a diffusion model that decouples and aligns heterogeneous geometric constraints—user-defined global trajectories and local image-derived poses—through a parallel control strategy. Building on TPMDM, the IKMo framework integrates a VLM-powered multi-agent system to translate user-friendly inputs into explicit motion constraints, enabling precise animation from unaligned, in-the-wild 2D reference images. Experiments on HumanML3D and KIT-ML show that IKMo achieves state‑of‑the‑art performance for motion generation under mixed trajectory and keyframe pose constraints.
By Yang Zhao, Yan Zhang, Zixuan Guo, Hangyu Zhou, Haotian Mao, Hui Wang, Xubo Yang
StepSafe introduces a training framework called ASPO that focuses on step‑level timely intervention in agentic systems. It aligns decisions with their respective prefixes, rewards actions that fit the current stage, and uses adaptive feedback self‑distillation to credit reasoning. The resulting StepSafe guardrail achieves state‑of‑the‑art performance on five agent‑safety benchmarks, improving the safety‑utility Pareto curve with a Youden’s J of 53.0 for a 4B model.
By Siyi Gu, Kunhan Gao, Chao Peter Yang, Sukanta Ganguly, Rex Ying, Arman Cohan
The paper investigates who can cancel an AI assistant’s promise in a multi‑party chat setting, defining a concept called standing. Experiments across six language models show that only the creditor or an authorized party should release a commitment, yet many models either incorrectly release promises (unauthorized release) or continue to execute them after release (override). The study finds that larger models are more compliant, that oracles can reduce unauthorized releases, and that adding a commitment ledger or a concise release policy significantly improves compliance.
By Lan Wei, Bingyu Shen, Jiaqi Tang, Qing Lu, Kuan Huang, Meng Xu, Miles Q. Li, Bin Hu, Boyang Li
MetaExp is a framework that trains language‑model agents to learn reusable experience across tasks by using verifiable cross‑task environment scaling. It introduces two designs—Shared‑Tool families that keep black‑box tools constant and Shared‑Workflow families that preserve latent procedures—allowing agents to acquire operational and procedural experience respectively. After fine‑tuning Qwen3.5‑122B‑A10B on thousands of trajectories from 1,000 task families, MetaExp achieves significant performance gains on benchmarks such as SpreadsheetBench, APEX‑Agents, and AutomationBench, with improvements persisting even when domain‑excluded training is applied.
By Renhao Li, Xinjie Shen, Yang Su, Jianhong Tu, Derek F. Wong, Dayiheng Liu, Min Yang
The paper introduces a three‑level shortcut‑auditing framework to assess AI guard reliability in agent‑safety evaluations. It demonstrates that shortcut cues—such as shorter shell commands or exact‑tool lookups—can predict trajectory labels and influence guard decisions, leading to significant changes in measured performance when benchmark representations are altered. Experiments on LinuxArena, AgentDojo‑Traj, and AgentHarm‑Traj datasets reveal that guard accuracy can drop dramatically under domain holdouts and that editing safe‑row markers or tool aliases materially affects false‑positive rates.
"whyItMatters":"The study shows that benchmark design can directly alter guard decisions and evaluation outcomes, underscoring the need for careful evaluation protocols to ensure reliable safety assessments."
By Hasan Demirkiran, Jens Ernstberger
Matoi is a system that separates reusable consequence evidence from current execution authority and durable occurrence history, enabling consumers to reuse only their own validated evidence while keeping signed requests, live grants, quotas, and uncertain outcomes as distinct obligations. Its finite query‑aware contract handles branches, ordered collections, and negative lookups, and it distinguishes locally serialized claims from non‑transactional remote effects. Empirical studies show that reuse reduces lifecycle cost by up to 3.49 % in JavaScript and 1.79 % in Python, and that the system can preserve benign transfers while rejecting reused attack‑plan/policy crosses.
By Zexun Wang
The paper introduces Empirical Research Assistance (ERA), an autonomous system that employs Large Language Model‑guided tree search to create, evaluate, and refine executable forecasting software for infectious diseases. In a real‑time evaluation during the 2025‑26 U.S. respiratory season, ERA produced diverse models for influenza, COVID‑19, and RSV, and an ensemble of these models outperformed the CDC’s gold‑standard forecast ensembles across all three pathogens. ERA also demonstrated robustness in data‑scarce RSV scenarios and showed that higher‑tier language models and specific optimization metrics can improve forecast accuracy.
By Sarah Martinson, Michael P. Brenner, Martyna Plomecka, Brian P. Williams, Nicholas G. Reich, Zahra Shamsi
The study reports a mixed‑methods pilot evaluation of an AI‑orchestrated research workflow in which a human researcher coordinated multiple LLM‑based agents to perform data extraction, corpus construction, artifact generation, and artifact evaluation. Using the generation and assessment of multiple‑choice questions (MCQs) as a testbed, the team collected 1,071 SAT Math MCQs and employed LLM agents to extract questions from PDFs, retrieve and convert open textbooks into structured representations, align each MCQ with relevant textbook content, generate new MCQs under specified difficulty and cognitive levels, and evaluate both original and generated MCQs using a 24‑criterion quality framework. Across all evaluations, average MCQ quality was high, but criterion‑level analysis and equivalence testing show that generated MCQs are not fully comparable to expert‑vetted baseline questions, with persistent gaps in skill depth, cognitive engagement, difficulty calibration, and metadata alignment.
"whyItMatters":"The findings highlight how AI‑orchestrated workflows shift researcher labor toward specification, orchestration, verification, and governance, underscoring emerging AI research operations skills needed for AI‑empowered research pipelines."
By Yuan An
The paper presents a taxonomy of 20 failure modes—called $A^2$-Failures—in agent-to-agent (A2A) systems where agents operate for different principals and have private states. Analyzing over 2,500 runs across 488 scenarios, the authors identify four categories of failures: discovery and delegation, grounding and record consistency, verification, and conduct and fidelity. Key findings include that task phrasing drives interaction, agents rarely verify peers’ claims, and overall success metrics obscure underlying coordination and verification issues.
By Sarath Shekkizhar, Prathyusha Jwalapuram, Zixuan Ke, Vaidehi Patil, Jiayu Wang, Srijan Bansal, Trong Huy Nguyen, Dingfeng Jiang, Dongqi Liu, Ruochen Zhao, Wei Lu, Giuseppe Carenini, Mohit Bansal, Silvio Savarese, Shafiq Joty
The study examines how temporal resampling (binning) of irregular clinical data affects the safety and performance of offline reinforcement learning (ORL) models in healthcare. By analyzing MIMIC‑IV insulin infusion records, the authors show that binning can reverse causal order and obscure critical events. In simulated trials with virtual type‑1 diabetes patients, ORL models trained on binned data performed up to 60% worse than those trained on unprocessed data, yet retrospective evaluation on resampled data overestimated their returns.
By Thomas Frost, Hrisheekesh Vaidya, Steve Harris
NeoHorse-1 is a family of agent-native models that pursue recursive self‑improvement through agentic post‑training. The system uses a heterogeneous model pool with intelligent routing to record capability demands, service tiers, and interactions, converting these records into training examples that preserve reasoning, tool calls, and harness context. By structuring supervised fine‑tuning into a three‑stage curriculum and employing routing‑guided on‑policy distillation, NeoHorse-1 closes an evaluation‑selection‑update loop, raising macro‑average performance from 58.94 to 64.87 on 4B models and from 65.60 to 69.04 on 9B models across eleven benchmarks, while also producing a prefill‑only variant that outputs structured decisions and probabilities for agent workflows.
By NeoHorse Team, Guoliang Cao, Guohao Dai, Tianyu Guo, Kai Han, Hailin Hu, Zihan Jiang, Xiang Kuang, Boxun Li, Yulong Li, Zehua Pei, Yuchuan Tian, Jiamin Wang, Yu Wang, Yunhe Wang, Yihong Wu, Haiyang Xu, Shuo Zhang, Hang Zhou, Siyang Cheng, Jiayu Fan, Wei He, Qingrui Jiao, Hongguang Li, Zhiyuan Li, Runke Liu, Xi Liu, Xinchen Liu, Sinno Jialin Pan, Yi Ren, Liuyang Song, Chenyu Wang, Bei Yu, Quanlu Zhang, Xiangyu Zhang, Mengyu Zheng, Yingjie Zong