MemoWM proposes a memory allocation framework that leverages world models to compress what long‑term agents retain. By using shared predictions, it reconstructs omitted content and balances reconstruction error against storage cost, keeping only information that adds downstream value beyond the predictive prior. On five long‑term agent‑memory benchmarks, MemoWM improves average answer accuracy by 42.42% and cuts experience‑specific storage by 53.9% compared to the most storage‑efficient baseline.
By Bingfan Zeng, Zhisheng Chen, Chenbo Sang, Zhengwei Xie, Jinpeng Wang, Xiangchen Guan, Rui Qian, Zheng Lu, Jingwei Song
The paper introduces ALERT‑BENCH, a benchmark that simulates enterprise security telemetry and evaluates how large‑language‑model agents triage alerts. Five common reasoning strategies—single‑pass tool use, iterative retrieval, sampled investigations, self‑review, and explicit verification—were tested on 1,247 alerts from a multi‑stage attack scenario, each missing at least 40.4% of attack‑related alerts. Building on these findings, the authors propose AIDA, a multi‑agent framework that enforces explicit decision proposals, independent challenges, and stronger evidence requirements, achieving an F1 score of 0.958 and reducing false‑negative alerts from 40.4% to 3.1% while escalating 18.4% to analysts.
By Saimon Amanuel Tsegai (Daphne), Alex Kantchelian (Daphne), Danfeng (Daphne), Yao, Peng Gao
WorldBench is a benchmark and automated judge for evaluating large language models that generate interactive 3D worlds using Three.js. It combines code inspection with dynamic visual analysis, requiring the model’s code to produce observable features such as biomes, physics, and seasonal cycles. The study shows that relying on either code or visual output alone can misjudge a model’s performance, with a mutation test revealing that code-only judging overestimates success by a third.
By Krish Bakshi
SWE-Journey is a new benchmark designed to evaluate coding assistants in more realistic settings, focusing on long-horizon tasks and multi-turn interactions. It introduces a weak-to-strong synthesis pipeline to create extended coding challenges and a user-simulation agent based on four real user personas to mimic authentic code-assistance dialogues. The benchmark reveals that while models succeed on over 75% of tests with software architects, they manage less than 25% with non-coders, highlighting a significant gap in current assistant capabilities.
By Hexuan Deng, Yue Wang, Wenyu Jiang, Cheng Yang, Haolin Yang, Zhaohua Zhang, Chenchen Zhao, Beiduo Chen, Muxi Chen, Sa Zhu, Geyuan Zhu, Jianhuan Zhuo, Qiuyong Xiao, Tianwen Jiang, Jihong Zhang, Xuebo Liu
The paper investigates how a higher‑level controller in a hierarchical latent reasoner decides whether to preserve or replace a lower‑level strategy. Experiments across different training seeds show that learned revision timing can produce varied policies, but none surpasses a fixed, forced timing policy on the same checkpoint. The study highlights that state‑dependent control can alter actions without improving task performance, and that timing decisions are order‑sensitive and largely captured by a strong fixed schedule.
By Ayushi Chadha
FlowState introduces a method for long-horizon LLM agents to treat execution state as a reusable memory, preserving semantically typed state nodes and their relations while separating persistent retention from on-demand access. The system uses Incremental State Update to maintain current state and Progressive State Access to reveal historical states during reasoning, allowing agents to reassess past decisions with new information. Experiments show FlowState improves success rates on MemoryArena and pass rates on τ³‑Bench by 4.55 and 13.95 percentage points, respectively, while cutting token consumption by over 40%.
By Minghao Li, Bangyan Li, Zifan Wang, Yulong Li, Hu Xu, Gan Zhang, Jingtong Wu, Wenqiang Xu
AgentTime is a benchmark that tests whether AI agents can manage their own runtime by working for a specified duration, predicting their wall‑clock time, and estimating elapsed time afterward. It includes 222 tasks from 18 sources across coding, computer use, agentic work, and automated research, with duration‑following experiments that append a single instruction specifying how long to work. Results show significant variability in accuracy, with some agents deviating from requested runtimes by up to 2.9×, and many agents stopping prematurely even after completing a task.
By Michael Ofengenden, Maksym Andriushchenko
The paper introduces an autonomous AI security agent designed to detect multi‑vector fraud and anti‑money‑laundering (AML) threats in both retail and corporate banking. It operates across two parallel event streams—transactions and sessions—using a fusion of LSTM behaviour models, statistical velocity monitors, and graph‑based account‑counterparty analysis. Experiments on a synthetic dataset show the agent outperforms rule‑based and LSTM‑only baselines, achieving F1 scores of 0.787 for transactions and 0.867 for sessions, while also providing rapid, low‑latency responses and supporting customer verification and analyst assistance.
By Joseph Walusimbi, Joshua Benjamin Ssentongo
The paper introduces TRACE, a protocol designed to diagnose whether changes in verifier scores for large language model agents reflect actual changes in agent behavior or merely alterations in the evaluation process. TRACE works by applying targeted changes to evaluation components, running paired experiments, and re‑scoring unchanged trajectories to isolate the source of score variation. Experiments on synthetic tasks and public benchmark tasks demonstrate that seemingly significant score shifts can often be attributed to evaluation artifacts, while TRACE can also detect genuine behavioral changes.
By Radhika Gaonkar
Delphos is a deep reinforcement learning framework designed to aid the specification of discrete choice models. It treats model specification as a sequential decision-making task, where an agent selects modeling actions—such as adding constants, choosing taste parameters, applying transformations, or including interactions—to generate candidate models. Each candidate is estimated and evaluated using a reward function that reflects statistical fit and behavioral expectations, and a Deep Q-Network learns which sequences of decisions lead to high-quality models, as demonstrated on simulated and empirical datasets.
By Gabriel Nova, Stephane Hess, Sander van Cranenburgh
The paper investigates whether small large language model agents can operate within explicit wall‑clock time budgets while using the allotted time productively. Experiments with Qwen3.6‑27B on MLE‑Bench Lite and Qwen3‑4B on Zork I show that agents fail to translate budget statements into controlled time use, largely due to lack of timing feedback and poor anticipation of action durations. Interventions such as harness‑based timing exposure, enforcement hooks, and reinforcement learning with budget‑aware rewards improve budget adherence but do not lead to better task performance, revealing a persistent gap between respecting time limits and effectively allocating extra time.
By Aaron Wang, Neelabh Madan, Vlad Sobal, Matthew Trager, Michael Kleinman, Elman Mansimov, Wei Xia, Stefano Soatto
The paper introduces GenUI‑Harness, a multi‑agent system that pairs a Tool Agent for data retrieval with a GUI Coder Agent that generates structured interfaces to aid human‑agent interaction. It tackles training challenges by using Dynamic UX for lightweight reward collection and Reward Auditor to monitor reward quality. The authors present UI‑TAU Bench, a benchmark of 1,300 database‑backed tasks, and show that GenUI‑Harness improves Pass@3 scores and reduces dialogue rounds compared to existing models.
By Xiaolong Li, Xiaohan Xu, Jinyang Li, Xinnuo Xu, Ge Qu, Nan Huo, Jack Williams, Reynold Cheng
ReTeach is a reflective self‑distillation framework that builds a self‑teacher through multiple rounds of reflection and retry, relying only on self‑generated attempts and outcome‑level verification. Starting from an unsuccessful student rollout, the teacher alternates explicit reflection with new attempts until success or a retry limit is reached, without needing reference answers, external feedback, or cross‑example memory. The method uses outcome‑aware weighting to distinguish correct, reflection‑corrected, and unresolved examples, and on‑policy distillation aligns the student’s token‑level predictions with the teacher’s context‑conditioned distributions, achieving a 1.39‑point accuracy gain over GRPO across six reasoning benchmarks.
By Yafeng Tang, Hao Li, Hongsheng Yu, Qiang Fu
The paper introduces a three‑dimensional framework for comparing AI methods that solve sequential decision‑making puzzles. It maps each approach to a Markov decision process, evaluates autonomy via human ranking, and measures skill against computational cost. Applying this to the Tower of Hanoi, the authors compare graph‑based, reinforcement learning, and large language model techniques, finding that LLMs shift complexity to inference time, incurring higher memory and runtime costs.
By Sadig Gojayev, Carolina Fortuna
Generative Adversarial Loops (GAL) is a framework that alternates between a discriminator generating adversarial data to expose weaknesses in current AI systems and a generator discovering algorithms to overcome those weaknesses. Applied to tasks such as KV compression, sparse video generation, sparse attention, and context extension, GAL’s discriminator identifies shortcomings in state‑of‑the‑art techniques, while the generator produces new algorithms that improve performance on both adversarial and established benchmarks. The study demonstrates significant gains, for example boosting CompactorPress on KV compression from 0.35 to 0.97 on the discriminator dataset and improving Dual Chunk Attention from 0.20 to 0.90.
By Kislay Aditya Oj, Nidhi Jain, Sri Surya Varma Datla, Priyanka Jayaswal, Kumar Krishna Agrawal, Aditya Desai
SkillMorph is a skill‑evolution method that uses trajectory‑guided fault localization to improve agent skills. It links execution evidence to specific skill contents by comparing failure and success trajectories across repeated runs, identifies suspicious actions, and localizes edit sites for revisions. Experiments on SWE‑Skills‑Bench and CannBot show that skills evolved with SkillMorph achieve higher trial‑level accuracy and execution consistency than the original skills and those produced by four other methods, and it has already generated six accepted skill‑revision pull requests in an automated kernel generation setting.
By Yu Ge, Linna Xie, Zhong Li, Yu Pei, Tian Zhang
The paper introduces a systematic study of how large language models (LLMs) perform social attribution—specifically responsibility and blame judgments—by creating a benchmark with 7,639 questions drawn from classic attribution theory scenarios and real-world narratives. It evaluates 32 LLMs and 5 non-LLM baselines, finding that larger models show greater but still incomplete alignment with human judgments. A probing methodology further reveals that certain attribution dimensions can be decoded from specific hidden-state positions, and their influence on LLM outputs mirrors patterns predicted by human Attribution Theory.
By Zhaoxin Yu, Qingchao Kong, Dajun Zeng, Wenji Mao
The paper introduces DVD, a dynamic vector decoding method that unifies 2D and 3D perception tasks for multimodal large language models. DVD transforms various perceptual representations—such as 2D bounding boxes, masks, and 3D bounding boxes—into 1D vector sequences, maps them to compact discrete tokens, and then decodes them back to the original formats via a lightweight de-tokenizer. Experiments on RefCOCO, SUN‑RGBD, KITTI, Hypersim, and nuScenes show that DVD outperforms existing methods while significantly reducing token overhead and inference latency.
By Jinghua Hou, Zhe Liu, Hengshuang Zhao
LiteGUI is a new framework for training lightweight GUI agents that tackles the long‑horizon, multi‑path nature of GUI interactions. It introduces a two‑stage post‑training pipeline: Guided On‑Policy Distillation, which uses human‑verified multi‑solution action annotations to provide privileged teacher guidance at each state, and Multi‑Solution Dual‑Level GRPO, which combines action‑level supervision with history‑conditioned planning‑quality supervision while handling multiple valid actions. The authors also present a scalable data generation pipeline and the Lite‑Dataset, enabling extensive evaluation across GUI benchmarks and demonstrating substantial performance gains over existing training paradigms.
By Yubin Wu, Zicheng Cai, Liping Ning, Hua Wang, Zhi Chen, Yaohua Tang, Hao Chen
The paper introduces DRL-DCO, a Deep Reinforcement Learning‑governed dynamic orchestration of Differential Evolution and CMA‑ES for high‑dimensional structural model updating. A DDPG actor‑critic agent uses a progression‑aware state and diversity‑informed reward to continuously reallocate resources between exploration and exploitation, while controlling population size, elite preservation, and restarts. The trained actor can operate in a supervision‑free inference mode, enabling fast deployment with maintained performance, and demonstrates superior convergence accuracy and robustness on benchmark optimization and structural health monitoring problems.
By Lechen Li (State Key Laboratory of Internet of Things for Smart City, University of Macau, Macau 519000, China, College of Water Conservancy and Hydropower Engineering, Hohai University, Nanjing 210098, China), Rongye Shi (School of Artificial Intelligence, Beihang University, Beijing 100191, China), Wanhuan Zhou (State Key Laboratory of Internet of Things for Smart City, University of Macau, Macau 519000, China)