PersonaCUABench is a benchmark that evaluates computer‑use agents (CUAs) on 74 tasks across 11 domains, varying only the persona of the requester. Each of 44 personas rewrites the instruction in their own words and answers the CUA’s questions, while the CUA sees only the dialogue. The study finds that task completion drops by 36% when personas are introduced, and user satisfaction falls from 84.1% to 62.7% for fully completed requests.
By Haoming Li, Defu Cao, Xu Hu, Yingtao Luo, Yiyang Wang, Xianghe Xu, Jiashu He, Wanting Jiang, Yifan Zhang, Yunhan Zhao, Siyang Wu, Chen Yang, Haixin Wang, Zheyuan Deng, Taoran Li, Jingwen Zhang, Zhejian Peng, Shushu Liang, Yue Zhao, Bingzhe Li, Feng Chen, Yifan Wang, Jintao Huang, Yijun Wang, Yuexing Hao, Xiaomin Li
PanoBench is a new camera‑centric benchmark for panoptic occupancy prediction that unifies 3D geometry, semantics, and instance identities across synthetic and real‑world driving scenes. It combines three datasets—PanoBench‑CarlaOcc, PanoBench‑nuScenes, and PanoBench‑KITTI360—through canonical interfaces and unified evaluation protocols, providing multi‑resolution annotations with persistent instance identities. The benchmark introduces a novel reconstruction pipeline for real‑world data, a unified valid‑mask protocol, and analyses of observation fidelity, geometric completeness, spatial continuity, and temporal consistency, while also benchmarking methods on semantic, panoptic, and 4D occupancy forecasting.
whyItMatters":"PanoBench addresses the scarcity of instance‑aware annotations in panoptic occupancy research by offering comprehensive, high‑quality annotations and evaluation protocols that enable more robust development and comparison of autonomous driving perception methods."
By Yi Feng, Zizhan Guo, Junwu E, Yu Ma, Mengtan Zhang, Jin Wu, Wei Ye, Hanli Wang, Qijun Chen, Danwei Wang, Rui Fan
DLite is a history‑free speculative decoding framework designed for agentic AI workloads. It bypasses the need for a historical KV cache by reusing contextual hidden states from the target model and performing local speculative predictions, supplemented by a lightweight sequential head to maintain causal dependencies. Experiments demonstrate that DLite improves acceptance length by 2–6× and outperforms the state‑of‑the‑art DSpark by up to 70%.
By Jiaming Xu, Hanzhen Wang, Tianlang Zhao, Jiewen Xiao, Chengze Yuan, Jiayi Pan, Yongkang Zhou, Jiancai Ye, Yu Wang, Guohao Dai
NestedVLA proposes a new approach for consolidating skills in Vision‑Language‑Action models by learning skills independently and then generating task‑specific parameters via a nested hypernetwork, rather than aggregating parameters. The method keeps a frozen base VLA and task‑specific compressed VLM residuals, enabling flexible, interference‑mitigated skill reuse. Experiments on MetaWorld, ManiSkill, and CALVIN show that NestedVLA outperforms the best model‑merging baseline by 41 pp on average and approaches per‑task fine‑tuning performance.
By Guodong Du, Wanyu Lin
FRACTAL is a self‑similar multi‑agent architecture that recursively decomposes goals using a five‑component pattern—Orchestrator, Ensemble, Knowledge Ledger, Decision Maker, and Meta‑Observer—at every level of a decomposition tree. The paper presents four formal results: a linear communication bound for intra‑node cost under a two‑phase protocol, a two‑mode orchestration criterion based on diversity‑saturation, a supermajority convergence guarantee with explicit probability bounds, and a Bayesian escalation mechanism that unifies hysteresis and contract‑breach signaling. The authors illustrate the architecture qualitatively and acknowledge that quantitative performance validation is deferred to future work.
By Ohad Dan, Abhinav Taliyan
The paper "From Approval to Abuse: Privilege Escalation Attacks on Harness-Protected Coding Agents" examines how coding agents that use execution harnesses to enforce permission boundaries can be tricked into granting overly broad allow-rules. By planting adversarial context in a repository, attackers can steer an agent into requesting a plausibly justified yet dangerously wide permission, which, if approved, lets the agent bypass the harness and execute malicious actions without further human oversight. Experiments on platforms such as Claude Code and Codex CLI show that 58% of trials succeed in eliciting such rules, which then cover 56.9% of category-specific capability tests and enable silent goal completion in 81% of approved cases, thereby undermining the safety promise of human-in-the-loop oversight.
By Xianglin Yang, Yechao Zhang, Jiahao Liu, Yulin Chen, Yuxin Cao, Bryan Hooi, Jiaheng Zhang
The study investigates whether five large language‑model agents adapt their contributions in a collective‑risk dilemma game. Across ten rounds, most models maintained a roughly fixed contribution level regardless of changing risk or partner behavior, and many continued to contribute even when it offered no payoff advantage. The results reveal that cooperative self‑play does not necessarily indicate strategic adaptability, and that agents’ responses to incentives, partner actions, and population composition vary significantly.
By Trung-Kiet Huynh, Dao-Sy Duy-Minh, Chi-Nguyen Tran, Phu-Hoa Pham, Phu-Quy Nguyen-Lam, Le Hong Trang, The Anh Han
NEXUS is a modular neuro‑symbolic framework that simultaneously learns a PDDL domain model and an LTL safety specification for embodied agents. It repairs action schemas using execution failures and converts detected hazards into reusable LTL constraints via hierarchical semantic and translation sampling. During symbolic search, LTL progression enforces these constraints, allowing the system to differentiate between physically unreachable states and safety‑induced refusals, and experiments on SafeAgentBench demonstrate superior safety‑task completion and low violation rates even under jailbreak attacks.
By Tiehan Cui, Peipei Liu, Yanxu Mao, Gaosheng Wang, Mingzhe Xing, Datao You
SkillBrowser is a Meta‑Skill‑managed framework that allows large language model agents to identify their current capability needs, formulate targeted queries, and selectively load skills on demand. By providing relevant procedural content only when needed, it reduces context pollution and misaligned skill exposure that can disrupt reasoning. Experiments on six SRA‑Bench subsets and five LLMs show that SkillBrowser improves aggregate accuracy over static retrieval and LLM‑based selection, especially for multi‑step tool use, while mitigating accuracy degradation on already‑solvable instances.
By Jin Ye, Yifan Jiang, Lan-zhe Guo
The paper presents an agent‑based model of daily multimodal mobility in Toulouse, comparing fifteen decision‑making approaches—including generative agents (GAs) with minimal and tuned prompts, simple baselines, survey‑fitted models, and a typed classifier—against real household travel survey data. Results show that GAs with tuned prompts approximate the modal split of a simulated day without disruptions but fall short on trip‑by‑trip accuracy, while the Jev‑1.13 TypeSafe classifier performs surprisingly well, indicating that generating output text is not essential. The study highlights the potential and current limitations of generative agent‑based modelling for urban mobility.
By Yves Bru, Kamaldeep Singh Oberoi, Benoit Gaudou
Scene-Designer is an agentic framework that learns design intelligence for generating 3D indoor scenes from natural language. It introduces two datasets—Scene-OTA-35K for design and error correction, and Scene-OTC-15K for quality critique—and a Critic-guided GRPO framework (Scene-GRPO) that refines design decisions. Experiments show that Scene-Designer outperforms existing methods in physical validity, functional correctness, and layout coherence, and that its learned Critic provides effective guidance during test time.
By Xuanyu Liu, Hongyan Feng, Miao Pan, Sunlai Chen, Lizhou Liang, Bing Sun, Wenqi Zhang, Xuhong Zhang
The paper studies how different configurations of privileged information (PI) affect on‑policy self‑distillation (OPSD) for large language models. By evaluating 16 PI setups across three agentic benchmarks and multiple student models, it finds that PI effectiveness varies with content, generation source, and model scale, with information‑dense trajectories and stronger generators yielding larger gains. The authors also propose KL‑based metrics, particularly Δd−_err, to predict PI utility before training, showing that selecting PI with this metric improves post‑training performance and remains beneficial when combined with GRPO.
By Kaijian Zou, Emiliano Penaloza, MohammadHossein Rezaei, Anisha Gunjal, Mohit Raghavendra, Lu Wang, Yunzhong He, Anas Mahmoud
The paper introduces WebMix, a mixture hypernetwork that enables web agents to adapt to unseen interfaces without retraining. WebMix uses a bank of pre‑trained low‑rank adapters and a scout agent that explores a new interface, summarizing observations into a manual. A lightweight router then predicts mixing coefficients to combine the adapters into a new one, achieving performance that matches or surpasses in‑context learning and LoRA baselines on domain and version transfer tasks.
By Md Farhan Ishmam, Minh Pham Dinh, Kenneth Marino
The paper presents a new method for sensitivity analysis of decision rules in an agent‑based agro‑ecological model. It treats the decision structure as a hierarchical design space, using a random‑forest‑supervised HSIC‑ANOVA to rank main effects and interactions across both categorical and numerical variables. The approach is applied to outputs such as nitrogen leaching, soil organic carbon variation, and yield predictions, providing an interpretable, decision‑oriented analysis.
By Benjamin Pianet, Baptiste Lesquoy, Paul Saves, Benoit Gaudou, Jean Villerd, Nicolas Verstaevel, Olivier Therond
ECHOES is a method for long‑horizon GUI agents that compresses visual history into a textual state using a Compression‑Think‑Action format. It employs a privileged teacher model that provides token‑level on‑policy distillation to guide the student’s textual rollouts, while a Perception‑Memory Fidelity Gap‑guided framework balances distillation with task reward optimization. The approach is evaluated on four GUI benchmarks and a new ECHOESBMK diagnostic benchmark, showing improved task performance, memory fidelity, and a favorable trade‑off between performance and inference cost, with positive zero‑shot transfer to unseen planning tasks.
By Tengda Guo (Tsinghua University), Jiawang Hu (StepFun), Yifei Tao (Nanyang Technological University), Yin Wang (StepFun), Haotian Hu (Zhejiang Leapmotor Technology Co., Ltd), Fanyi Wang (Honor), Liujuan Tang (StepFun)
The paper introduces an author‑in‑the‑loop collaborative audit system for reproducibility checks during peer review, where authors explicitly declare reproduction instructions via a structured checklist. An AI agent uses these declarations to investigate and assess reproducibility, while human reviewers approve requests and decide on issue resolution. The system was deployed on 36 conference papers, identifying 20 issues that were subsequently resolved, and the study examined how different checklist conditions affected the agent’s assessment.
By Yukun Dong, Linjun Zhang, Tian Zheng
The paper introduces Harness-Delta Attribution (HDA), a protocol that breaks down the performance gains from harness evolution into overfitting, test-time scaling, and residual generalizable improvement. Applying HDA to 11 executor models and three harness-evolution frameworks across various tasks, the study finds that over half of the search-set gains are due to overfitting or scaling, with limited transfer to held-out tasks. The results highlight the need for attribution-aware evaluation in future harness evolution research.
By Wenxuan Ding, Qin Lu, Changlong Yu, Sha Li, Shuowei Jin, Xin Liu, Zhenyu Shi, Greg Durrett
The paper introduces QEMem, a framework that rethinks how large language model agents use long‑term memory by constructing query‑specific evidence sets rather than relying on individual memory relevance. QEMem combines query‑guided discovery, source‑grounded reification into structured ViewCards, and sufficiency‑guided refinement to assemble comprehensive evidence from multiple memory granularities. Experiments show that QEMem achieves 91.43% LLM‑judge accuracy on LoCoMo with Claude Opus 4.5, surpassing leading baselines by significant margins and improving evidence coverage for cross‑session queries.
By Shixuan Li, Wei Yang, Peiyu Zhang, Anzhe Cheng, Heng Ping, Xiaole Zhang, Jesse Thomason, Xuezhe Ma, Yue Zhao, Paul Bogdan
GlobeBench is a new benchmark for evaluating text-based world models, comprising 1,632 tasks across five domains and nineteen simulation skills. It tests whether models can predict an environment’s response to an action based on the full agent trajectory, requiring understanding of state changes, environment rules, and action effects on related and unrelated objects. The benchmark compares predictions to execution-derived references using a hybrid architecture of deterministic gates and a semantic judge, and finds that GPT‑6 Astra tops the 20 tested models with a 49.57% success rate, while accuracy drops sharply when actions affect many dependent objects, leaving 631 tasks unsolved by any model.
By Nicholas Saban, Charvi Bannur, Mariya Vasileva, Darshan Deshpande
R$^2$Orch introduces a Recursive Self-Play Orchestration framework that uses an inner self‑play loop to improve an orchestrator on self‑selected tasks and an outer loop that refines how the inner loop improves. The method employs Recursive Learnability to reward task selection that yields diverse outcomes as skills are removed, and Skill Withdrawal to prune skills when the orchestrator performs equally well without them. Experiments on twelve datasets demonstrate that R$^2$Orch outperforms existing baselines in question answering, math, code, and interactive tasks.
By Mingda Zhang, Yun Huang, Qiang Huang, Yanjin Li, Rui Qian, Qichao Zhao, Tiesunlong Shen, Na Zhao, Jian Wang, Xiaoying Tang