The paper introduces Gated-Memory Routing, a method for efficient collaboration in multi‑agent large language model systems. It uses a learned execution memory with write and retrieval gates to keep only non‑redundant reasoning steps, and an adaptive halting controller to stop execution when enough evidence is gathered. Experiments on five reasoning and code‑generation benchmarks show the approach achieves higher accuracy and reduces inference cost by 31.9% compared to the strongest baseline.
By Rakibul Hasan Rajib, Mengxing Zheng, Qian Lou
CoBRA is a counterfactual boundary‑learning framework designed to improve when a tool‑augmented language model should call an external tool. It builds internal and external experts from the same base model, collects paired trajectories, and estimates the reward margin between answering with and without tools. Using these margins, CoBRA partitions data into internal‑favored, external‑favored, and ambiguous cases, then applies Boundary‑Aware Cold‑Start SFT and MARS‑RL to optimize boundary decisions, leading to more efficient tool use and better accuracy on tool‑dependent out‑of‑distribution questions.
By Wenhao Zou, Xianglong Liu, Wendong Bi, Hanjie Wang, Simin Zhao, Gong Zhi
The paper introduces the Prospective Intention Store (PIS), a method that places lifecycle logic in code and confines language tasks to a typed action space, enabling small models to perform prospective memory tasks more effectively. Using PIS, a small model (DeepSeek-Chat) achieves 82.9% Set‑F1 on PM‑Bench, surpassing the previous best of 65.1%. On Gemma‑E2B, PIS boosts Set‑F1 from 4.2% (without a store) to 66.2%, and reaches 70.1% Set‑F1, outperforming retrospective memory approaches that max out at 54.4%.
By Jinqing Zhao, Chengcan Wu
The paper introduces the Mass‑Conserving Perceptron (MCP), a physics‑aware AI framework that enforces conservation laws while learning hydrological process relationships from data. By progressively adding physically meaningful components—such as bounded soil storage, state‑dependent conductivity, variable porosity, infiltration capacity, surface ponding, vertical drainage, and nonlinear water‑table dynamics—to a single MCP storage unit, the authors demonstrate that predictive skill for daily streamflow improves across 15 U.S. catchments. The study finds that the impact of each process representation varies with hydroclimate, with vertical drainage boosting performance in arid and snow‑dominated basins but hindering it in rainfall‑dominated ones, while surface ponding has minimal effect; the best MCP configurations rival LSTM benchmarks while retaining explicit physical interpretability.
By Mohammad A. Farmani, Hoshin V. Gupta, Ali Behrangi, Muhammad Jawad, Sadaf Moghisi, Guo-Yue Niu
PEARL is a framework that trains Socratic tutoring agents using pedagogically aligned reinforcement learning. It introduces a controllable student simulator to model diverse cognitive states, a reward model that jointly evaluates pedagogical quality and correctness, and a stable multi‑objective RL approach to balance competing tutoring goals. Experiments demonstrate that PEARL competes with both open‑source tutoring systems and leading proprietary LLMs.
By Qikai Chang, Zhenrong Zhang, Linbo Chen, Pengfei Hu, Jianshu Zhang, Youhui Guo, Jun Du
Qwen-Drive-1.0 is a vision‑language foundation model tailored for autonomous driving that builds on a pretrained VLM architecture. It incorporates a bird’s‑eye‑view perception head for 3D object detection, semantic occupancy prediction, and BEV map segmentation, and a Planning Expert that generates future ego trajectories from shared representations. Experiments show strong 3D perception, driving scene understanding, and competitive motion‑planning performance while largely preserving general vision‑language capabilities.
By Xin Zhou, Zongchuang Zhao, Zhibo Yang, Mingsheng Li, Humen Zhong, Shuai Bai, Du Chu, Ruizhe Chen, Zhaohai Li, Jun Tang, Qiuyue Wang, Mingkun Yang, Jiazhao Zhang, Dayiheng Liu, Dingkang Liang, Xiang Bai
The paper introduces LEGO, a benchmark dataset pairing user text descriptions with human‑annotated fine‑grained constraints and reference 3D scenes, and LEGO‑Eval, an evaluation framework that decomposes descriptions into atomic constraints and verifies each using grounding and spatial reasoning tools. It demonstrates that LEGO‑Eval detects misalignment more accurately than existing methods and that current 3D scene synthesis approaches achieve at most a 10% success rate on this benchmark.
By Minseok Kang, Dongwook Choi, Gyeom Hwangbo, Seungwon Lim, Kai Tzu-iunn Ong, Jinyoung Yeo
UniACE is a unified framework that standardizes the evaluation of large language model (LLM) agents by representing each benchmark as an instruction–tool–environment triplet and running models through a shared, task‑agnostic harness in isolated runtimes. It preserves native success criteria, offers an offline mode for dynamic‑resource tasks, and standardizes efficiency metrics, execution records, and failure attribution. Applying UniACE to 7 benchmarks across 24 domains and 15 models revealed significant score shifts, ranking reversals, and sensitivity to evidence representation, highlighting the impact of evaluation configuration on reported agent performance.
By Pengyu Zhu, Lijun Li, Yaxing Lyu, Qianxin Luo, Jingyi Yang, Yi Liu, Tingfeng Hui, Xinyu Yuan, Li Sun, Sen Su, Jing Shao
MemoryWalker addresses the conditioning problem that arises when training agents with compressed context during rollout. It introduces two exact, gradient‑equivalent corrections—LogitTree, a segmented K‑forward traversal, and a packed 4D attention mask—alongside SDCC, a self‑distillation method that reduces the train‑deployment gap by minimizing KL divergence at each eviction. Experiments on seven web‑search benchmarks show that SDCC significantly lowers logit drift and boosts rollout rewards compared to naive training.
By Zinco J, Xunjie Zhu, Shen Huang, Zhenyi Wang, Pengjun Xie, Jieping Ye
UI‑Venus‑2 is a general‑purpose foundation GUI agent that operates across mobile, web, and desktop environments using a unified closed‑loop reasoning‑action framework. The report details how the system expands environment coverage to over 170 multilingual mobile apps and native desktop OSes, scales task generation through a deep‑research pipeline, and enhances verification with trace‑level and sample‑level evaluators that use visual keypoints and multi‑model voting. Safety‑aware mechanisms are also incorporated to control consequential actions, positioning UI‑Venus‑2 as an efficient, open‑source tool for more generalizable, verifiable, and self‑reflective agents in real‑world applications.
By Venus Team, Zhuohan Cai, Haoxing Chen, Jiaxuan Chen, Weizhi Chen, Changlong Gao, Zhangxuan Gu, Yuan Guo, Yusong Hu, Jianrong Jiang, Jianguo Li, Runze Li, Jinzhen Lin, Zhenyu Ma, Changhua Meng, Han Peng, Xinyu Qiu, Shuheng Shen, Zhongyi Shui, Weiqiang Wang, Ming Wen, Zhuoer Xu, Hang Yan, Kaiwen Yang, Ruilin Yao, Nanjun Yu, Zhengwen Zeng, Lianrui Zhang, Yunzhu Zhang, Zhe Zhao, Beitong Zhou
ReDeck introduces a step‑level render‑grounded refinement framework for document‑to‑slide generation, breaking slide revision into atomic edit actions with immediate renderer‑derived observations. It employs multi‑granular feedback—step‑level spatial checks, turn‑level adaptive critique, and a submission‑level layout gate—to balance local repair with overall quality. The authors also present DeckQuiz, a benchmark that separates content fidelity, spatial correctness, and design quality, and demonstrate ReDeck’s superior performance across GPT‑5.4, Claude‑4.6, and Gemini‑3.1.
By Muzhao Tian, Zezi Zeng, Yifan Yang, Xin Gao, Yan Li, Zisu Huang, Xiaohua Wang, Changze Lv, Mingxi Cheng, Bei Liu, Kai Qiu, Qi Dai, Dong Chen, Yue Dong, Xiaoqing Zheng, Ji Li, Chong Luo
The paper introduces a novel framework for assessing second‑order bias in large language models (LLMs), defined as bias in how an LLM judges the acceptability of biased content. Using principles from entitlement epistemology, the authors design a reasoning task that asks LLMs to determine whether a biased text is acceptable for specific demographic groups, and propose two metrics to quantify biased judgments. Experiments on both open‑source and closed‑source models reveal that the task bypasses safety guardrails, uncovers systematic variations across target groups, and demonstrates that models still rely on demographic labels when evaluating bias.
By Ramaravind Kommiya Mothilal, Terry Jingchen Zhang, Raiyan Ahmed, Zhijing Jin, Shion Guha, Syed Ishtiaque Ahmed
The paper introduces DroneCATS-Agent, a modular framework that places a multimodal large language model (MLLM) at the core of a drone’s control loop, allowing the model to decide actions solely from prompts. It presents the DroneCATS benchmark, evaluating MLLMs on four tasks—approaching, tracking, searching, and multi‑drone commanding—without fine‑tuning or function‑calling. Results show that while small open models can navigate reliably, they often fail by mismanaging protocol termination, highlighting a gap between perception and action planning in current MLLMs.
By Jaewoo Park, Minyoung Lee, Sukmin Seo, Moonbin Yim, Hyunwook Yoon, Dohoon Ryu, Daehee Kim, Myungseo Song, Jihyuk Byun, Seunggyu Chang, Taeho Kil, Jiseob Kim, Bado Lee, Geewook Kim
TRIAGE introduces a three-level routing framework for Large Language Model agents that reduces token consumption by reusing historical execution trajectories. The system classifies queries into direct reuse, skill substitution, and full ReAct levels, achieving significant token savings in large-scale security monitoring and cross-domain benchmarks. An automatic skill extraction mechanism further refines reusable patterns, creating a positive feedback loop that improves efficiency over time.
By Ruocan Wei
The paper introduces a benchmark for evaluating large language models (LLMs) on long‑horizon state tracking by having them compute the MD5 hash through 196 dependent tool calls across 64 rounds, carrying four 32‑bit words in context. It shows that a mixture‑of‑experts LLM can maintain the full state and produce correct digests in most runs, even when all primitive tools are replaced by another LLM. The study isolates state‑tracking difficulty from instruction interpretation and identifies key factors—contextual reasoning and worker voting—that enable success.
By Dheeraj Mohandas Pai, Lu Xian
EULER is a multi‑agent system designed to facilitate the transfer of mathematical problems across different domains by constructing and evaluating ‘bridges’ that connect a source representation to a target representation. The system employs direct, adjacent‑domain, and distant‑domain search strategies, and uses six ordered stress tests to filter out invalid bridges before costly search. In experiments on 120 recent conjectures from combinatorics, EULER produced 10 proofs, 3 refutations, and 45 partial results, demonstrating that bridge‑specific stress tests and combined bridge‑material with target‑native operations significantly improve performance.
By Ren Zhenzhuo
The paper introduces FAIRY, a full-stack smart‑agriculture agent system deployed on a soybean research farm at Harbin Institute of Technology. FAIRY executes and evaluates end‑to‑end agronomic operations—from ridge preparation to storage—using an event‑driven world model that integrates machinery, sensors, drones, satellite data, weather, crop models, and historical yields. The system implements a comprehensive agentic stack and is used to benchmark nine state‑of‑the‑art agent controllers across 100 full‑season soybean scenarios, assessing success, spatiotemporal correctness, token cost, and edge‑device runtime.
By Ao Qu, Panagiotis Michelakis, Linyuan Han, Yiannis Hadjiyianni, Kun Ouyang, Konstantinos Siskos, Feng Li, Ran Meng, Jingchi Jiang, Dimitrios Stamoulis, Jie Liu
The paper investigates how large language and visual‑language models used in autonomous vehicles inherit human driver biases when deciding whether to yield to pedestrians. It introduces two new bias‑testing methods—All Else Being Equal and Self‑Consistency tests—to evaluate these models. Results reveal that the models’ yielding decisions are influenced by pedestrian attributes such as gender, ethnicity, religion, disability, age, skin tone, and socio‑economic status, with varying patterns across models.
By Irem Yoldas, Martim Brand\~ao, Jie Zhang, Odinaldo Rodrigues
The paper introduces the concept of an irreversibility budget for managing risk in fleets of large language model agents. It proposes a cumulative accounting system that tracks residual value‑at‑risk across agents, workflows, and tenants, charging each effect its residual loss and denying any marginal effect that would exceed the budget. A controlled study shows that per‑effect gates can allow fleet‑level overdraws up to 48 times the tenant’s risk limit while the budget keeps all correctly charged runs within limits, highlighting the challenge of designing conservative, dependency‑aware pricing.
By Bardia Mohammadi, Laurent Bindschaedler
SlideBank is a training‑free framework that turns each whole‑slide image into a persistent, concept‑indexed evidence bank. It performs coarse‑to‑fine exploration to locate informative regions and multi‑scale views, converts them into explicit morphological observations, and anchors pathology signals to the supporting patches and slide coordinates. During inference, questions are routed to relevant signals and evidence scales, and a confidence‑based cross‑level consensus integrates global, regional, and patch evidence, achieving high accuracy on WSI‑VQA and SlideBench‑BCNB while enabling consistent re‑phrasing and reduced inference cost.
By Beidi Zhao, Gexin Huang, Ciro Zhang, Anqi Li, Yusheng Tan, Chen Zhou, Gang Wang, Zu-hua Gao, Xiaoxiao Li