WCC is a learning‑based framework that decides dynamically when to apply Cooperative Coevolution (CC) versus Non‑Decomposition Algorithms (NDAs) in large‑scale global optimization. It models this decision as a Markov decision process, using a tailored feature set, reward scheme, and network architecture to enable fine‑grained, adaptive switching and conditional budget allocation without expert input. Experiments show WCC competes strongly and generalizes across problems, horizons, and baseline optimizers, supporting a low‑budget "train‑once and reuse" approach while providing new practical insights into CC deployment.
By Wenjie Qiu, Hepei Li, Jianfeng Wang, Jun Zhang, Yue-Jiao Gong
The paper introduces HSC-Bench, a benchmark with 1,039 tasks designed to evaluate how well receiving AI models can take over tasks after handoffs. It measures two key aspects: the ability to recover relevant repository evidence and the ability to continue from established conversational, tool, and file-work states. Experiments with six models show that structured handoff information consistently improves evidence recovery and, for most models, continuation performance, while the best performers differ between the two tracks.
By Wenyuan Li, Guang Li, Dongyuan Li, Keisuke Maeda, Takahiro Ogawa, Miki Haseyama
The paper introduces a linguistic-aware agentic data sampling framework for multilingual automatic speech recognition (ASR). It embeds a data sampling LLM agent into the training loop, guided by a rubric that analyzes ASR errors across phonological, morphological, and syntactic dimensions. Experiments on twelve Common Voice languages with Whisper-medium demonstrate that the method outperforms full-data training while using less data, fewer computational resources, and achieving faster convergence.
By Zhiheng Qian, Junhao Du, Jing Peng, Hanqi Li, Kunyang Peng, Feng Xue, Lu Chen, Shuai Wang, Kai Yu
The study investigates how large language models (LLMs) predict users’ decisions in human-agent dialogue by combining users’ statements with past decision records. Using 1,260 rounds of a repeated prisoner's dilemma with 42 users and five agents, the researchers found that decision records alone yielded strong predictive performance (AUC = 0.80), while adding dialogue did not improve overall accuracy but did increase agreement with users’ statements. However, predictions over‑emphasized statements, especially when users had not yet made a decision, and failed to capture variations in cooperation that depended on the specific agent or users’ trait alignment.
By Baiqiao Zhang, Shuai Ma, Zhifeng Liao, Chao Zhou, Jianshu Wang, Yuhan Zhao, Chaoyue Zhang, Tongjun Si, Yuying Tang, Xiangxian Li, Yong-Jin Liu, Yulong Bian
TrawMem redefines agent memory systems as Task‑Oriented Role‑Adaptive Workspaces, selecting and organizing memories by their roles in task solving rather than mere relevance. It structures interaction histories into local threads and jointly selects complementary memories to build task‑specific workspaces. Experiments on LoCoMo and MemGallery show consistent macro‑F1 gains of 2.23–3.14 points and improved LLM‑Judge scores with fewer context tokens.
By Mingyu Zhang, Zixu Li, Zhiheng Fu, Zhiwei Chen, Jinhe Bi, Yupeng Hu
ETAMem is a multi‑view memory framework designed to improve long‑term conversational agents by decoupling heterogeneous information and aligning memory representations with distinct characteristics. It parses conversational data into atomic units, identifies affective content, and organizes these units into three complementary hierarchical views—Event, Affect, and Time—using adaptively calibrated thresholds. Retrieval is performed via entropy‑based allocation of query budgets across views, enabling hierarchical search within each view. Experiments on LoCoMo and LongMemEval‑S show that ETAMem surpasses state‑of‑the‑art baselines with relative accuracy gains of 12.46% and 7.96%, while remaining token‑efficient, and analyses confirm the complementary benefits of the different memory views.
By Xinying Qian, Ying Zhang, Yu Zhao, Xubo Liu, Baohang Zhou, Jeff Z. Pan
The paper presents an agent designed for the GLEE Competition that engages in bargaining, negotiation, and persuasion games against human and AI opponents. The agent combines a classical game‑theoretic solver, an online Bayesian layer for estimating opponent archetypes, and a language model for generating messages. Experiments include 13,200 Monte Carlo games against 11 synthetic archetypes, an ablation study of the opponent‑modeling layer, and an analysis of 9,337 live‑game messages showing a strong correlation between moral rhetoric framing and buyer purchase decisions.
By Hayden Arnold
EnvSched is an online scheduler that learns from rollout history to make agentic reinforcement learning more efficient. It assigns scheduling value by prioritizing tasks likely to yield informative reward differences and spreading batches across underobserved environments and interaction patterns. By modeling environment–task pairs as a dual‑relation graph and propagating historical rollout statistics, EnvSched predicts reward variance and guides batch allocation, achieving higher scores than random selection within fewer training steps on Qwen3‑8B.
By Sihan Yang, Weichen Bai, Ruichuan An, Yiming Cheng, Zhiyuan Feng, Tianyi Jiang, Renye Yan, Ning Ding, Yifan Zhang, Renrui Zhang, Yehui Tang, Wentao Zhang
HEART-Bench is a benchmark designed to assess whether large language models (LLMs) can make decisions that are consistent with a given persona and past experiences. It features 11 synthetic characters anchored in the Big Five personality traits, each with 1,000 autobiographical memories, and 64 decision scenarios guided by the DIAMONDS framework, resulting in 673 multiple‑choice questions. Models are evaluated with and without memory retrieval, showing that memory‑augmented configurations improve accuracy to 64.9%, while human annotators achieve 80.8% on a matched subset.
By Weihan Peng, Chenxu Zhang, Qianao Wang, Yuling Shi, Heng Lian, Qihong Mao, Jiahao Pang, Chunliang Feng, Bowen Li, Xiaodong Gu
The paper analyzes 2,487 tasks from five public coding‑agent benchmarks and 14,922 runs of six agent configurations, introducing the Spread‑Novelty‑Centrality (SNC) profile to describe how reference solutions are distributed, added, and the code’s dependencies, history, and complexity. It finds that every pair of benchmarks differs on at least two SNC axes, even when sharing the same task label, and that resolved runs tend to involve less distributed changes and lower centrality values, with patch sizes varying between Claude and Qwen configurations. The authors recommend that practitioners choose evaluation tasks based on change characteristics, inspect task descriptions, and assess patch scope alongside test outcomes to better align benchmarks with real‑world workloads.
By Radin Shayanfar, Keheliya Gallaba, Ahmed E. Hassan
BoxWorld introduces a novel approach to real‑time interactive video worlds by decoupling interaction reasoning from generative rendering. A coding agent builds a Box World Engine that applies user‑defined rules to determine interaction outcomes and exposes object states via a lightweight interface of layout boxes, persistent attributes, and dynamic status variables. A separate Box‑State Generative Renderer then autoregressively renders these evolving states into video, enabling online control over object motion, state switches, and conditional events. The system employs a three‑stage training strategy—learning object grounding from open‑domain videos, state transitions from gameplay videos, and rollout distillation on self‑generated histories—to handle real‑time state updates while maintaining visual coherence. BoxWorld outperforms end‑to‑end world models and remains competitive across multiple public benchmarks, with a new PairedStateBench benchmark for fine‑grained interaction control.
By Zike Wu, Ahmed Hatem, Chintan Trivedi, Wallace Michel Pinto Lira, Xuebin Qin, Lele Wang, Renjie Liao
LabWorld is a Real2Sim framework that reconstructs interactive 4D laboratory worlds from monocular RGB video, producing reusable 3D assets, metric scenes, time‑varying object states, and procedure events. By leveraging multi‑frame observations and motion reconstruction that combines visual, geometric, and event constraints, LabWorld outperforms existing baselines on real pipetting and vortexing tasks, achieving higher Mask IoU and PCK‑AUC scores. The framework enables explicit, editable state representation for replay, inspection, and future embodied scientific systems.
By Suozhi Huang, Kelsey Fu, Binxu Li, Luna Lyu, Kaize Wu, Wenbo Lu, David Smerkous, Jiahan Li, Siqi Zhu, Zhizhou Sha, Jiacheng Guo, Zhixuan Liang, Shilong Liu, Le Cong, Mengdi Wang
FeatureArc is a new benchmark that evaluates coding agents on the evolution of software features, requiring agents to extend a feature across multiple stages while preserving earlier functionality. The benchmark uses a pipeline that constructs tasks from real repository histories, validates fail‑to‑pass and pass‑to‑pass tests, and covers four programming languages. In tests on 73 tasks, agents performed well on the first stage (62.7% success) but dropped to 28.5% when the entire feature was evaluated, with inherited defects causing most failures.
By Yunlong Wang, Alyssa Wu, Cuiyun Gao, Jia Feng, Songyuanyi Lu, Zhichao Hu
TeamGym is a new environment that evaluates and trains language‑model agents on team decision‑making tasks drawn from real‑world legal, finance, medical, and software engineering texts. It introduces the concept of pivotal information—a minimal set of facts necessary for a correct decision—and tracks how agents share and retrieve this information. Experiments show that distributing pivotal facts and requiring agents to extract them independently significantly reduces success rates, but reinforcement‑learning training can improve decision quality and foster coordination among team members.
By Chanwoo Park, Jaywoong Jeong, Zachary Wojtowicz, Yubin Kim, Taekyun Lee, Kaiqing Zhang, Asuman Ozdaglar
VibeVideoBench is a benchmark that evaluates multimodal agents capable of generating videos by writing code to render each frame and sound. The benchmark contains 239 self‑contained tasks in English and Chinese across 12 categories, requiring agents to produce a finished MP4 with sound within one hour without using any pre‑trained image or video generation model. Evaluation is hierarchical, using program checks, 13k task‑specific rubric checks, and five quality dimensions scored by three VLM judge panels, and it shows that GPT‑6 Astra and Claude Opus 5.5 perform best among nine frontier agents.
By Muxi Diao, Wenbin Xu, Lele Yang, Xuyang Ma, Yuzhe Chen, Jiankang Chen, Kaiyu Jiang, Changyi Liu, Haoyu Zhang, Hongyi Fu, Jianxiong Wang, Tianke Zhang, Haonan Fan, Yingxin Li, Bingyi Li, Xu Wang, Xuanyu Zheng, Bin Wen, Tingting Gao, Han Li, Wentao Zhang, Weiran Xu, Kongming Liang, Zhanyu Ma
DreamTacVLA is a Vision‑Language‑Action framework that incorporates high‑resolution tactile images alongside wrist‑camera and third‑person vision to enable robots to anticipate and react to contact dynamics. It uses a Hierarchical Spatial Alignment loss to align tactile tokens with visual counterparts and a tactile world model to predict future tactile signals, trained on a hybrid dataset from digital twins and real experiments. The approach achieves up to 95% success on contact‑rich manipulation tasks, outperforming existing VLA baselines.
By Guo Ye, Zexi Zhang, Xu Zhao, Shang Wu, Haoran Lu, Shihan Lu, Han Liu
UnityBench-Mini3D is a benchmark comprising 77 Unity project-level tasks organized into seven expert-defined skill categories and three difficulty tiers. Each task pairs a complete Unity project with natural-language requirements and is evaluated using 1,178 held-out, human-authored milestone tests that assess scene structure, component configuration, and runtime behavior. The benchmark provides deterministic partial credit, localizes failures, and distinguishes partial implementations from complete solutions without relying on subjective judges, achieving an overall pass rate of 57.7% across 21 models and 9 coding harnesses.
By Yijun Zhou, Minghao Shao, Ke Zhang, Renkai Yuan, Hongbo Yuan, Bin Hu, Bozhao Qian, Hongxia Shen, Yanhong Liu, Yuting Chen, Yuntao Shi, Zhongxuan Xu, Shuo Mao, Haijia Zhou, Zhendong Li, Tao Zhang, Chenggang Li, Yuxuan Wang, Zinan Zheng, Calvin Zhang, Jia Li, Heng Chang, Jie Feng
The paper introduces a work-centric framework for understanding and optimizing work agents, which combine foundation models with runtime harnesses to perform tasks in stateful workspaces. It defines two key dimensions—work persistence and work openness—to characterize working conditions and classifies digital work into five forms based on task scope and organization. The authors identify five core capability families necessary for reliable execution, explore optimization strategies for models and harnesses, and evaluate existing metrics from the perspective of verifiable work outcomes.
By Zishan Xu, Enjun Du, Mingqi Gao, Cheng Yang, Junhao Shen, Zirong Chen, Zhiyuan Yao, Yifu Guo, Yuquan Lu, Jinyang Huang, Zhengxi Lu, Siyi Liu, Ziyu Zheng, Yongqi Zhang, Difan Zou, Weiwen Liu
The paper discusses how software engineering (SE) expertise can guide the training of large language models (LLMs) for coding tasks. Using a case study of a 92‑billion‑parameter LLM, the authors created RefactorForge, a pipeline that leverages SE knowledge and tools to generate verified training data from 306 Python repositories. Fine‑tuning the model on this data improved performance on several benchmarks, including RefactorBench and SWE‑Bench ProMax, and the gains extended to other coding‑agent harnesses.
By Amir M. Ebrahimi, Mohammed Mehedi Hasan, Seyed Shayan Daneshvar, Jiahuei Lin, Gopi Krishnan Rajbahadur, Ahmed E. Hassan
Dream-COP introduces a framework that uses large language models to discover algorithms for combinatorial optimization problems by replaying past search trajectories in an experience world. A policy agent learns to guide a discovery agent through dreaming, providing feedback on program quality and search steps. Experiments on five COPs show that Dream-COP outperforms fixed exploration, matching optimal values on the Knapsack Problem and reducing the OBPP gap while cutting search requests by 42.5%, and its policies transfer across different problems.
By Kuan Xu, Jianming Wang, Ni Zhang, Yaoxin Wu, Yong Wang, Zhiguang Cao