The paper introduces Hippocam, a hierarchical memory and continual learning architecture for large language model agents that mirrors human memory processes. Hippocam organizes an agent’s work into nested intents, keeping the current intent active while consolidating completed intents into task‑relevant outcomes and states, and recursively abstracts earlier history to reduce detail over time. When recalled, past experiences are re‑integrated and further consolidated with new data, reinforcing and updating knowledge without requiring parameter updates.
By Xiangyi Zeng, Baihang Liu, Xutong Wang, Ze Jin, Yunpeng Li, Qixu Liu
DataSense-Bench is a new benchmark that tests whether advanced AI models can choose and rank training data subsets that will improve a small language model’s performance. The benchmark lets agents inspect data, run analysis code, and perform forward passes, but not train the model or access evaluation tasks. After agents select subsets, the base model is fine‑tuned on each subset and evaluated, measuring both the performance of the best subset and the accuracy of the agents’ rankings. Experiments show limited gains over random selection and inconsistent ranking ability across tasks, with agents often using similar data signals but interpreting their value differently.
By Yudi Zhang, Mingyu Cao, Lu Yin, Mykola Pechenizkiy, Shiwei Liu
The paper introduces Adversarial Heuristic Learning (AHL), a framework that uses AI agents to refine game policies without changing model weights. It presents AAArena, a benchmark of 12 adversarial games and 1,920 human programs, where agents analyze replays, select opponents, and revise policies to maximize ranking within limited budgets. Experiments show that certain configurations achieve gold medals, while performance drops in games with complex rules, highlighting challenges in game understanding and long‑horizon policy development.
By Kaisen Yang, Qingle Liu, Kejin Wang, Yicheng Zhao, Jieming Li, Shenghan Zheng, Ruize Yang, Bojun Yang, Heng Gong, Xiang Gao, Lanyue Zhang, Kaiyu Zhong, Zhuo Liu, Shaoxuan Li, Chengxi Li, Yong Yan, Weixuan Zhang, Tianwei Luo, Situ Wang, Youjie Zheng, Sihan Zhao, Shengyuan Wang, Huan-ang Gao, Jiazheng Xu, Xiaohui Xie, Wentao Han, Hongning Wang
The paper introduces a new evaluation for large language model agents called epistemic humility (EH), which measures an agent’s ability to recognize, act on, and communicate uncertainty when evidence conflicts with prior beliefs. EH is operationalized through three trajectory-level behaviors—Identify, Solve, and Escalate (ISE)—and tested in both controlled and naturally occurring knowledge conflict scenarios. Results show that higher task accuracy does not guarantee greater epistemic humility, and that model-level interventions can improve EH but often reduce accuracy, indicating that humility depends on the interplay between the backbone model, agent harness, and environment.
By Kaiser Sun, Bernal Jimenez Gutierrez, Hongjun Liu, Jingyu Zhang, Jie Gao, Mark Dredze, Daniel Khashabi
The paper surveys how large language models (LLMs) are being used in hardware verification, covering tasks such as assertion generation, testbench creation, bug localization, and formal reasoning. It categorizes the literature by methodology, verification goal, tool interaction, benchmarks, and evaluation metrics, and distinguishes between inference-time techniques (prompting, retrieval, structured reasoning, agentic workflows) and training-time adaptation. The authors highlight that LLMs excel as semantic reasoning, search, and orchestration components within verification workflows, yet emphasize the need for semantic alignment, scalable integration, generalization, and rigorous evaluation to ensure true verification correctness.
By Hao Zheng, Jaime Rafael Imperial, Bardia Nadimi, Xiangfei Kong
The paper presents a unified dataset of 350 Lean tasks to evaluate specification generation (SpecGen) for verifiable code, combining metrics of formal validity, reference similarity, equivalence, and behavioural adequacy. It highlights that restricting similarity measures to a subset of tasks can shift rankings, and shows that a perfect postcondition score can still miss critical input coverage, underscoring the need for separate feedback on input and output constraints.
By Srijith Nair, Aditya Vempaty, Jia Liu, Ashish Jagmohan
The study investigates whether large language models (LLMs) have improved in screening systematic reviews (SRs) within software engineering. Using the SESR‑Eval benchmark and a new smaller dataset, the authors evaluated eight recent LLMs and found only marginal performance gains over older models (average MCC rose from 0.347 to 0.365). The research also examined prompt variations, inter‑model agreement, and the impact of refining inclusion/exclusion criteria, concluding that LLMs still fall short of replacing human reviewers and that future work should explore agent‑based methods, prompt engineering, and further criteria refinement.
By Aleksi Huotala, Miikka Kuutila, Mika M\"antyl\"a
The paper investigates whether subliminal learning (SL) can transfer a broader set of traits beyond simple preferences, such as complex capabilities, backdoors, and hacking tendencies. Experiments show that SL can indeed impart a student model with the ability to predict outputs of a random MLP, partially embed a backdoor that triggers French responses to female names, and increase hacking success in a chess environment. The effectiveness of transfer varies with the distillation method, being enhanced by logit distillation or limiting LoRA to attention layers.
By Jan Dubi\'nski, Anna Sztyber-Betley, Jan Betley, Owain Evans
Skill‑V is a verifiable self‑evolving skill library for interactive agents that represents skills as versioned, falsifiable contracts linking semantic intent to observable behavior. The system uses task failures to add new skills and contract‑outcome disagreements to revise existing skill boundaries, ensuring revisions preserve semantic constraints and pass non‑regression tests on historical evidence. An applicability‑aware filter excludes confidently inapplicable skill candidates, leading to high success rates on ALFWorld (95.3%) and WebShop (85.9%) while keeping the library more compact than growth‑only baselines.
By Jie Ma, Zhipeng Qian, Yufei Ma, Zihan Liang, Jiayi Ji, Qingpeng Cai, Ben Chen, Peng Jiang, Xiaoshuai Sun
The paper investigates whether in‑domain training alone can enable multimodal large language models (MLLMs) to perform fine‑grained industrial anomaly understanding (MM‑IAU). It shows that even specialized MLLMs fail to match human experts on defect localization and detection, and that combining different MLLMs does not eliminate this shortfall. To address this, the authors propose SiGMA, a spatially grounded multi‑agent framework that splits tasks among heterogeneous MLLM agents and a dedicated visual defect expert, achieving 85.2% accuracy on the MMAD benchmark—close to human performance and surpassing the best trained specialist.
By Xingwu Zhang, Duanyang Du, Huiling Zhu, Jiayue Dai, Yixiao Liu, Guozhi Liu, Zhihan Zhang, Zijun Long
The paper introduces RISEBench and its expanded version RISEBench++, a benchmark for evaluating reasoning-informed visual editing (RISE). RISEBench++ categorizes tasks into six reasoning dimensions—Temporal, Causal, Spatial, Logical, Counterfactual, and Hybrid—decomposed into 12 subcategories and 65 fine-grained task types, and includes 1,000 human‑annotated test cases in English and Chinese. The authors also present RISE‑Agent, a training‑free framework that integrates reasoning‑driven planning, tool‑augmented execution, and verifier‑guided refinement, outperforming most existing approaches on diverse RISE tasks.
By Xue Yang, Peiyuan Zhang, Yilun Zhu, Qihao Yang, Mingxin Liu, Xiangyu Zhao, Ziqian Fan, Zhaokai Wang, Yan Li, Yifan Yang, Xu Yang, Xiaosong Jia, Yue Zhou, Zhihang Zhong, Junchi Yan
OneSearch‑VL is a unified multimodal deep research agent that handles single‑image, multi‑image, and video tasks by leveraging a Visually Grounded Evidence Graph (VGEG) to encode dependencies among visual anchors, entity relations, facts, and answer‑producing operations. The VGEG‑based data engine builds and verifies questions, filters expert trajectories, and supports supervised fine‑tuning (SFT) and reinforcement learning (RL) with the Evidence‑aware Visual‑Grounded Rubric reward (EVGR). Experiments demonstrate that OneSearch‑VL‑8B outperforms Qwen3‑VL‑8B with tool access by 20.2 and 17.6 percentage points on new multimodal benchmarks, while also achieving significant gains across seven image benchmarks and VideoDR.
By Hongyu Li, Manyuan Zhang, Kaituo Feng, Shu Chen, Dian Zheng, Hao Li, Hao Yu, Zhangquan Chen, Zoey Guo, Ray Zhang, Shaofei Huang, Tianrui Hui, Linjiang Huang, Si Liu
WorldWeave is a framework that separates world-state maintenance from visual rendering to enable persistent geometric memory in video generation. It uses diffusion-based image outpainting to create continuous elevation maps, an agent that plans and stitches scenes hierarchically, and a read-only interface that queries world geometry for depth-guided video synthesis. This design allows the world to expand continually without predefined boundaries while preserving consistent structure across camera trajectories and revisits.
By Yifan Huang, Lifan Jiang, Qingyue Hao, Cheng Chen, Boxi Wu, Xiaoxue Ren, Xiaofei He, Dehai Zhao
The paper explores optimal trading under partial information, modeling the trading signal as an Ornstein‑Uhlenbeck process with regime switching. It proposes three reinforcement‑learning approaches that integrate recurrent neural networks to filter latent market regimes, comparing a one‑step method and two two‑step methods that use posterior regime probabilities or signal forecasts. Simulations and an equity pairs‑trading case study show that using posterior probability estimates of the latent regime yields the best cumulative rewards and more interpretable strategies.
By Andrea Macr\`i, Sebastian Jaimungal, Fabrizio Lillo
The paper introduces BEVPIPE, a framework that enables deployment of bird's-eye-view (BEV) perception models—used in autonomous driving to fuse camera and LiDAR data—on portable GPU compute APIs. By partitioning the model into dense subgraphs managed by standard inference runtimes and handling sparse 3D convolutions and geometric scatter operations through external extensions, BEVPIPE achieves a 19.5× speedup while preserving 98.5% of the reference mean average precision. The approach also demonstrates portability across different GPU backends, breaking the current reliance on CUDA-only, PyTorch‑coupled sparse convolution libraries.
By Rohit Verma, Anand V Bodas
BridgeGuard is a safety‑constrained diffusion planning method for autonomous driving that progressively strengthens a constraint term during denoising, guiding intermediate trajectories toward a scene‑dependent safety domain. It operates in a low‑dimensional curve space and uses a learned DistanceFieldNet to predict a time‑dependent distance field from bird’s‑eye‑view features, providing safety injection while keeping the perception backbone and planner frozen. Experiments on Bench2Drive show that BridgeGuard improves driving scores and success rates across two diffusion‑based models, demonstrating cross‑model generalization.
By Zhenjun Qiu, Jianing Huang, Dongang Liu, Baiyu Du, Yixun Niu, Hao Yang, Xinyu Huang, Chuan Hu, Shu Liu
The paper introduces uncertainty-aware extensions of the X-TRACK physics‑aware trajectory prediction framework, namely X-TRACK-DE and X-TRACK-MCD. These extensions predict future vehicle motion variables while explicitly modeling and propagating both aleatoric and epistemic uncertainties through vehicle dynamics, rather than only in trajectory space. Conformal prediction is then applied to the trajectory‑space covariance to create uncertainty regions that achieve a desired marginal coverage level, and experiments on the highD dataset demonstrate improved prediction accuracy and calibrated uncertainty estimates.
By Aanchal Rajesh Chugh, Sebastian Dorn
The paper investigates how co-installed coding‑agent skills can conflict, causing the model to choose a similar skill over the intended one based solely on name and description. An empirical study of 20,947 repositories uncovered 822,109 candidate skill pairs, with 312 confirmed conflicts tested across three models, revealing that nearly one in four installed skills is overridden by a similar skill. Key findings include that conflicts often involve normative skills, that the installed skill loses core functions in about one‑fifth of runs, and that the first skill read determines the outcome, suggesting benchmarks should evaluate exclusive core functions and platforms should guard the initial read.
By Chaoliang Yan, Zihao Xu, Yuekang Li, Shangzhi Xu, Yi Liu, Gelei Deng, Siqi Ma
The paper introduces study contracts that bind declared experimental choices, run obligations, and claim scope to recorded execution evidence, aiming to distinguish contract-relative verification from scientific truth. A diagnostic using eight self‑authored clean/mutated pairs shows that a deterministic checker can detect all registered mutations when the fault‑specific rule is applied to approved and executed objects. In a broader test involving eighteen judge aliases and 144 mutated evaluation cases, 104 cases received defect flags, while the rest were abstentions or failures, illustrating the challenges of an information‑asymmetric development setting.
By Eitan Waks, Ben Glocker
Large language model agents often reason before every action, but this may be unnecessary if earlier reasoning still supports later actions. The paper introduces RACE, a method that uses likelihood-based signals to detect when reasoning can be omitted, called LoGiC. Experiments on four benchmarks show RACE cuts reasoning cost while keeping or improving performance.
By Yiruo Cheng, Shen Huang, Xiaoshuai Song, Jiejun Tan, Guanting Dong, Pengjun Xie, Ji-Rong Wen, Zhicheng Dou