ReHarness is a recurrent harness framework designed for agentic continual learning, enabling language model agents to adapt online to distribution shifts. It augments forward read/write operations with backward revision, using feedback from each instance’s trajectory to refine the harness state. Experiments on three continual learning benchmarks and multiple language model backbones show that ReHarness outperforms baseline methods, producing source‑grounded revisions that improve stability and plasticity.
By Yuxin Xiao, Zhiyun Lu, Jianyu Wang, Yangtian Zhang, Yang Song, Qifan Wang, Bailin Wang
The paper presents an unsupervised multimodal reward model built on a frozen Qwen3‑VL backbone. By adding a linear head to an intermediate layer and training it to rank a web caption higher than a perturbed version of itself, the authors avoid human preference labels and achieve competitive performance—outperforming GPT‑4o on several benchmarks with far fewer parameters and compute. The approach demonstrates that a carefully chosen intermediate layer, simple perturbation-based training, and verifier‑driven fine‑tuning can yield effective vision‑language rewards without costly annotation.
By Yueying Li, Jingxuan Fan, Kiant\'e Brantley, Sham M. Kakade, Hanlin Zhang
JoyInside‑VLM‑Agent is a ReAct‑style multimodal agent that learns product‑conditioned policies for embodied robots and smart hardware. It uses a closed‑loop sandbox to generate interaction trajectories, supervised fine‑tuning, multi‑teacher on‑policy distillation, and reinforcement learning with DPER to align token‑level evidence and rewards. The agent achieves a BFCL v4 overall score of 77.44, surpassing baselines on JoyInside‑Devices and JoyInside‑Robotics, and demonstrates visually grounded search, replanning, and failure reporting in two real‑world quadruped case studies.
By Gangyi Li, LuYao Wang, Jingfu Sun, Jingyi Yuan, Hang Liu, Hao Zong, Yujia Hu, Tingting Gao, Jianyong Wang, Xu Yang
AgentSWE is a new benchmark for agent software engineering that evaluates coding agents across three lifecycle stages: creating an agent from a natural‑language requirement, editing existing agent codebases, and optimizing toward a measurable target. The benchmark consists of 25 commission‑style tasks, where builders iterate on a few illustrative cases and are then tested on held‑out cases that cover the full requirement, with hard pass/fail gates or objective‑based grading. Experiments show that current coding agents fail to meet the requirements, revealing common failure modes such as overfitting, incomplete runtimes, and unsupported completion claims.
By Jiahao Wang, Hongjin Qian, Yuyang Hu, Jiajun Zhang, Zheng Liu
The paper presents a method for detecting and mitigating subliminally transferred traits in language model distillation when the teacher and student share base parameters. By analyzing parameter updates across diverse datasets, the authors recover a weight‑space direction that reveals hidden traits such as animal preferences, without requiring trait labels. This direction can then be amplified for interpretation or projected out to suppress the trait with minimal impact on task performance.
By Joseph Egan, Ciar\'{a}n M. Gilligan-Lee, Michael O'Riordan, Yuchen Zhu
The paper introduces the Counterfactual Graph Neural Process (CFGNP), an end‑to‑end method that directly answers counterfactual queries from observational data without first estimating a causal graph. CFGNP employs Graph Neural Networks to leverage a dual‑graph structure, limiting cross‑world message passing to matched factual–counterfactual variables. Experiments on synthetic and image‑based datasets, including the German Loan benchmark, show that CFGNP outperforms existing methods, reducing mean‑squared error from $4.51 imes10^{-2}$ to $0.19 imes10^{-2}$, and generalizes from 10‑node to 30‑node graphs without retraining.
By Hendrik Stumpf, Francesco Quinzan
SpatialASM is a framework that teaches Large Vision‑Language Models (LVLMs) to perform Active Spatial Mentalization (ASM), enabling them to proactively plan and imagine viewpoints during spatial reasoning tasks. The method uses an allocentric novel‑view generator to create explicit ASM trajectories and perspective‑taking tasks, providing two forms of supervision that guide LVLMs in adaptive view planning and scene imagination. Experiments on seven spatial benchmarks show that SpatialASM improves performance across four LVLM backbones while preserving their general vision capabilities.
By Yuhong Liu, Shenxi Wu, Junyi Chi, Beichen Zhang, Xuanlang Dai, Jiaqi Wang, Yuhang Zang, Dahua Lin
The paper introduces QXR, an adaptive framework that repairs inconsistent natural‑language claim sets by treating a large language model as a noisy consistency judge. QXR identifies minimal conflicting subsets and removes a small hitting set of claims, thereby preserving the largest coherent subset of the original set. Experiments on benchmarks derived from VitaminC, FEVER, and conflicting‑evidence RAG show that QXR retains substantially more valid claims than all‑at‑once judging while maintaining high precision.
By Paul He, Elke Kirschbaum, Shiva Kasiviswanathan
The paper introduces ChemCost, a benchmark of 1,427 reactions that tests large language models’ ability to estimate chemical procurement costs by grounding identities, retrieving supplier quotes, selecting purchasable packs, normalizing quantities, and computing costs. It includes a frozen pricing snapshot of 2,261 chemicals and 230,775 supplier quotes, and provides scalar scoring plus stage‑level diagnostics for grounding, retrieval, procurement, and arithmetic errors. Experiments show that even advanced LLM agents achieve only about 50.6% accuracy within 25% relative error on clean inputs, and performance drops sharply with realistic noise, highlighting brittleness in parsing, evidence integration, pack selection, and tool use.
By Yuyang Wu, Yue Huang, Shuaike Shen, Xujian Wang, Shuhao Zhang, Qiyao Xue, Weichen Liu, Runtian Gao, Jian Ma, Xiangliang Zhang, Olexandr Isayev
UxSID is a framework for modeling ultra‑long user sequences that balances efficiency and effectiveness by using semantic‑group shared interest memory. It employs Semantic IDs (SIDs) and a dual‑level attention strategy to capture target‑aware preferences without the heavy cost of item‑specific models. The end‑to‑end architecture achieves state‑of‑the‑art performance and a 0.337% revenue lift in a large‑scale advertising A/B test.
By Hongwei Zhang, Qiqiang Zhong, Junfeng Shu, Jiangxia Cao, Yiyang Lv, Huanjie Wang, Liwei Guan, Jing Yao, Yiyu Wang, Zhaojie Liu, Han Li
MAGE (Multi-Agent Graph-guided Evolution) is a framework that externalizes self-knowledge of language-model agents into a four-subgraph co-evolutionary knowledge graph, allowing a frozen execution model to retrieve task-conditioned guidance from past reasoning traces and teacher-written corrections. During evolution, the graph, a task-level search bandit, and a skill-level routing bandit are updated from the same reward stream while the learner’s backbone remains unchanged. Across nine diverse benchmarks, MAGE outperforms prompt-based frozen-backbone baselines, and ablations show that self-harvested success traces and teacher-written corrections complement each other, with success memories aiding reasoning-template-heavy tasks and corrective memories supporting harder composition and interaction settings.
By Ruiyi Yang, Zechen Li, Hao Xue, Imran Razzak, Flora D. Salim
The paper introduces RobustBench‑TC, a benchmark that tests tool‑use language agents under 22 real‑world perturbations affecting observations, actions, rewards, and transitions. Experiments on 21 models show that observation noise has minimal impact, while reward‑relevant and transition noise drastically reduce accuracy, and model size alone does not mitigate these gaps. The authors propose ToolRL‑DR, a domain‑randomized reinforcement‑learning recipe that improves robustness, retaining most clean accuracy and narrowing performance gaps even on unseen transition perturbations.
By Xiaolin Zhou, Aojie Yuan, Zheng Luo, Zipeng Ling, Xixiao Pan, Yicheng Gao, Haiyue Zhang, Jiate Li, Shuli Jiang, Prince Zizhuang Wang, Zixuan Zhu, Jinbo Liu, Ryan A. Rossi, Hua Wei, Xiyang Hu
The paper introduces BenchAgent, a framework for evaluating multi‑agent large language model (LLM) workflows against a single‑agent baseline, BenchAgent Core. Using GPT‑4.1 across ten reasoning, coding, and tool‑use benchmarks, the study compares six multi‑agent configurations, finding that EvoAgent achieves the highest mean pooled accuracy (79.70%) while other agents lag by 1.71–8.76 percentage points. The analysis also highlights trade‑offs in token consumption and accuracy, offering guidance for selecting and optimizing LLM multi‑agent workflows.
By Yuhang Fu, Ruishan Fang, Jiaqi Shao, Huiyu Zheng, Zhengtao Zhu, Bing Luo, Zhenyu Wen, Tao Lin
UniConFlow is a unified constrained flow‑matching framework that generates robot trajectories while systematically handling both equality and inequality constraints such as collision avoidance, actuation limits, and dynamic consistency. It introduces a prescribed‑time zeroing function to shape a time‑varying guidance field during inference, and employs a violation‑segment extraction protocol and trajectory compression to address computational challenges in long‑horizon, high‑dimensional planning. Experiments on a double inverted pendulum, a real‑to‑sim car racing task, and a sim‑to‑real manipulation task show that UniConFlow outperforms state‑of‑the‑art generative planners and conventional optimization baselines on certified motion planning metrics like safety, kinodynamic consistency, and action feasibility.
By Zewen Yang, Xiaobing Dai, Dian Yu, Zhijun Li, Majid Khadiv, Sandra Hirche, Sami Haddadin
LLMail-Inject is a large-scale benchmark comprising 208,095 human-written prompt injection attacks gathered from a public red‑teaming challenge with 839 participants. The dataset tests the full attack chain—retrieving prompts from an email database, bypassing production defenses, and triggering unauthorized tool calls with correct arguments. Evaluation shows that models performing well on smaller benchmarks struggle on LLMail-Inject, with GPT‑4o achieving 58% attack success and Meta SecAlign 32%.
By Sahar Abdelnabi, Aideen Fay, Ahmed Salem, Egor Zverev, Kai-Chieh Liao, Chi-Huang Liu, Chun-Chih Kuo, Jannis Weigend, Danyael Manlangit, Alex Apostolov, Haris Umair, Jo\~ao Donato, Masayuki Kawakita, Athar Mahboob, Tran Huu Bach, Tsun-Han Chiang, Myeongjin Cho, Hajin Choi, Byeonghyeon Kim, Hyeonjin Lee, Benjamin Pannell, Conor McCauley, Mark Russinovich, Andrew Paverd, Giovanni Cherubin
The paper proposes a model‑free, instance‑based learning framework that uses surprisal—an information‑theoretic measure of uncertainty—to analyze raw data directly, bypassing explicit distribution modeling. By quantifying relevance through uncertainty, the method supports generalizable inference across diverse tasks such as generative modeling, causal discovery, anomaly detection, and time‑series forecasting, while offering traceability and interpretability. The authors claim state‑of‑the‑art performance on anomaly detection and some supervised tasks, and argue that the mathematical foundations create a ‘physics’ of information that can underpin scalable, human‑understandable AI.
By Christopher J. Hazard, Michael Resnick, Jacob Beel, Jack Xia, Cade Mack, Dominic Glennie, Matthew Fulp, David Maze, Andrew Bassett, Martin Koistinen
AnomalyAgent is an agentic framework for industrial anomaly synthesis that uses five tools—Prompt Generation, Image Generation, Quality Evaluation, Knowledge Retrieval, and Mask Generation—to iteratively refine anomaly samples. It employs a two‑stage training process combining supervised fine‑tuning and reinforcement learning, guided by a three‑part reward system that encourages task performance, self‑reflection, and trajectory adherence. On the MVTec‑AD dataset, AnomalyAgent outperforms zero‑shot state‑of‑the‑art methods, achieving superior anomaly generation scores and high classification and detection accuracy.
By Jiaming Su, Tengchao Yang, Ruikang Zhang, Zhengan Yan, Haoyu Sun, Linfeng Zhang
The paper introduces Stochastic MeanFlow Policies (SMFP), a one‑step generative policy class that merges MeanFlow noise‑to‑action mappings with Gaussian reparameterisation. SMFP captures diverse action modes while providing an analytic entropy lower bound through a learnable conditional Gaussian scale, enabling entropy‑regularised mirror‑descent learning. Experiments on seven MuJoCo benchmarks show SMFP outperforming Gaussian and generative baselines, and additional tests demonstrate effective multimodal action modelling and exploration in sparse‑reward navigation.
By Zeyuan Wang, Da Li, Yulin Chen, Yuehu Gong, Yanming Guo, Liang Bai, Tianyuan Yu, Yanwei Fu
The paper introduces a multi-agent peer‑reviewed reasoning approach for medical question answering, where several large language models independently generate chain‑of‑thought explanations and then evaluate each other’s reasoning for factual correctness and logical soundness. The highest‑rated reasoning chain is used to produce the final answer. Experiments with five state‑of‑the‑art LLMs on three benchmark datasets show that this method consistently outperforms single‑model chain‑of‑thought reasoning and majority voting ensembles, achieving an average accuracy of 0.820 versus 0.777 for the best single model.
By Zaifu Zhan, Shuang Zhou, Rui Zhang
The paper introduces First-Order In-Context Shapley, an inference‑only method that evaluates data utility for large language model alignment by computing the first‑order term of an in‑context cooperative game over semantic k‑NN neighborhoods in O(k) forward passes. It projects log‑likelihood shifts onto a directed influence graph, yielding three signals—Exemplifying Advantage, Exemplified Advantage, and One‑of‑a‑Kind Score—to identify high‑utility demonstrations and rare prompts. The approach reduces compute by 19× compared to Monte‑Carlo Shapley, filters 1–5% of records, and improves error detection and reward accuracy in DPO training and benchmark auditing.
whyItMatters":"By enabling scalable, inference‑only data auditing, the method addresses label noise and safety inversions that hinder post‑training alignment of large language models, leading to measurable improvements in model reliability and reward accuracy."
By Yunting Song, Matthew Watson, Moustafa Farid Alzantot, Peter Grabowski, Jun Qin