The paper proposes an adaptive multi‑discriminator Wasserstein GAN (MD‑WGAN) framework for Internet of Vehicles (IoV) that combines reinforcement learning and game‑theoretic coordination to manage resource allocation for distributed training. Roadside units host generators with Deep Q‑Network agents that select discriminator subsets and coordinate training epochs, while a unified objective links adversarial quality to resource efficiency, communication overhead, and latency. Experiments on real‑world NGSIM trajectory data show comparable prediction accuracy to state‑of‑the‑art GANs, with a 28% reduction in CPU usage and 6% reduction in memory usage, while maintaining competitive communication overhead and latency.
By Farhoud Jafari Kaleibar, Amr M. Zaki, Marin Litoiu
The paper introduces MARS-Gov, a standpoint-aware multi-agent system designed to detect and mitigate bias in Dutch government documents. It combines legal evidence retrieval, open-set target screening, specialized jurors, conservative routing, and rewrite verification, and adds a dynamic "10th juror" when new bias groups are identified. On the DGDB benchmark, MARS-Gov achieves state‑of‑the‑art performance with an F1 score of 0.880, surpassing zero‑shot LLM detectors and supervised Dutch encoders while reducing unnecessary interventions to 2.5%.
By Yuchen Miao, Zijun Wang, Chang Han, Yurui Shi, Mingtai Zhang, Siyang Xu
The paper proposes using language models as Research World Models (RWMs) to predict outcomes of experimental interventions across diverse research environments. By training on over 2,600 experimental records and 171,000 GPU‑hours, RWMs improve prediction accuracy within and across environments, reducing selection regret by up to 78% and boosting gains in multi‑round autoresearch. Ablation studies show that incorporating research knowledge yields greater benefits than merely changing the language model or increasing reasoning effort.
By Zijun Wang, Zewen Liu, Minhua Lin, Zhaotian Weng, Zhan Shi, Bing He, Yisi Sang, Dakuo Wang, Benoit Dumoulin, Wei Jin, Yuyin Zhou, Cihang Xie, Hanqing Lu
REMORY is a neural memory network that augments a textual summary of an agent’s history with a small set of soft memory tokens. These tokens, conditioned on the summary, enable a frozen large language model to approximate the continuation it would generate using the full history, effectively acting as a residual connection along the sequence dimension. Experiments on SummHay and long‑horizon agent benchmarks show that REMORY improves source attribution, maintains insight coverage, and achieves near full‑context performance while using only 5.2% of the input positions, while also reducing repeated tool outputs and errors on BrowseComp and Terminal‑Bench 2.1.
By Hanchen Xia, Baoyou Chen, Yutang Ge, Naihao Deng, Senqiao Yang, Zilong Dong, Weihao Yuan, Siyu Zhu
TaReD introduces a tool-aware recursive decomposition strategy for long‑horizon tasks, organizing tools into a hierarchy of capabilities and allowing agents to discover and use tools on demand. By aligning subtask decomposition with the required capabilities, the method reduces the unreliability of long execution chains and improves task success rates. Experiments on complex real‑world tasks demonstrate up to a 40‑percentage‑point improvement over baseline approaches.
By Wei-Xiang Mao, Zhi-Kai Chen, De-Chuan Zhan, Han-Jia Ye
Mine Odyssey is a new benchmark that tests agentic spatial intelligence by having agents navigate Minecraft reconstructions of real-world locations. The benchmark includes 180 tasks across 30 sites in 20 countries, featuring 20 outdoor and 10 indoor settings with varied scales, layouts, and terrains. Each task gives a natural‑language instruction to visit specific waypoints in order, requiring agents to find routes, open doors, and manage navigation errors.
By Yuxuan Cao, Junlong Li, Hao Li, Junxian He
The paper introduces the concept of "empty commitments," promises made by AI agents that cannot be fulfilled due to the absence of necessary tools or runtime capabilities. It defines three failure types, proposes a detection checker, and evaluates it on 400 human-labeled examples and a benchmark of 293 follow‑up requests. The study shows that many open‑weight models frequently issue such empty commitments, and that simple interventions—informing the model of its runtime, adding a capability card, or running a checker loop—can significantly reduce these failures.
By Jiaqi Tang, Bingyu Shen, Lan Wei, Qing Lu, Bethel Ololade, Danny Galvis, Miles Q. Li, Bin Hu, Boyang Li
The fourth Perception Test challenge was held as a workshop at ECCV 2026 in Malmö, Sweden, focusing on spatial intelligence. It included four tracks: unified multiple-choice videoQA, grounded videoQA from the original benchmark, and two new city-scale walking-tour tracks—KilometerAudio and KilometerVision. The report details the new benchmarks and summarizes the winning solutions, noting that while expensive agentic pipelines succeeded in the city-scale tracks, standalone multimodal models still struggle with complex spatial and multimodal reasoning.
By Fedor Kitashov, Jo\~ao Carreira, Shiry Ginosar, Dima Damen, Andrew Zisserman, Viorica P\u{a}tr\u{a}ucean
The paper introduces a distribution‑free framework for assessing the quality of external data before acquisition. By applying novel conformal two‑sample tests, it identifies data agents whose data exceed a contamination threshold, enabling selective data acquisition for personalized machine learning. The approach guarantees finite‑sample false discovery rate control and demonstrates robustness across various collaborative learning scenarios.
By Martin V. Vejling, Shashi Raj Pandey, Christophe A. N. Biscio, Petar Popovski
VGGTWorld-VLA extends the VGGT-World model by adding intention-conditioned control to 3D world evolution for autonomous driving. It introduces an action–semantic conditioning mechanism that injects driving semantics and ego‑motion into the future‑token stream, allowing different future geometry predictions for the same observed scene under alternative actions. A geometry–language–action bridge further integrates historical geometry, VLA semantic features, and maneuver/trajectory representations to jointly condition future geometry prediction, achieving competitive performance on NAVSIM and demonstrating the benefits of semantic and action conditioning.
By Zhaoyang Liu, Kun Jiang, Ziying Song, Diange Yang
The paper investigates how scaling ego‑centric human video data affects world model performance. Using 30,000 hours of video across diverse scenes, the authors find that while agent‑level fidelity improves markedly with more data, object‑level fidelity remains low and improves slowly. They demonstrate that careful visual conditioning and a new supervision scheme can shift model capacity toward object dynamics, yet a substantial gap persists, indicating that data quantity alone is insufficient for full world modeling.
By Jiahua Dong, Anurag Bagchi, Yash Jangir, Muhammad Zubair Irshad, Sergey Zakharov, Martial Hebert, Homanga Bharadhwaj, Yu-Xiong Wang, Vitor Campagnolo Guizilini, Pavel Tokmakov
The paper investigates how a coding agent learns across a sequence of abstract reasoning tasks in the ARC-AGI-3 benchmark. The agent, operating on a frozen foundation model, writes and runs Python and shell scripts, retaining only its written artifacts as state. The authors develop a measurement protocol that traces each thought—any belief, rule, or plan committed to a file—through its lifecycle, revealing that most task-specific scripts are discarded at task boundaries and that the agent selectively forgets by rewriting knowledge into new scripts while accumulating unedited notes.
By Chen Wu, Josh Passenger, Yin Song
The paper investigates when to evolve a harness versus training model weights for long‑horizon LLM agents. By labeling failed trajectories with the first signal that fires, the authors separate process failures from content failures, showing that harness evolution fixes the former while weight training addresses the latter. Experiments on DeepPlanning and WebArena‑Lite demonstrate significant performance gains, with LoRA adapters trained on evolved harness trajectories further improving held‑out scores and reducing content failures.
By Yuan Tian, Bing Hu, Hao Wang, Binghang Lu, Fang Wu
The paper introduces two inference‑time designs—factor‑compiled admission (FC) and permission‑semantic admission (PS)—to govern persistent memory in large language model agents without retraining. FC evaluates entire memory entries, while PS decomposes entries into typed units, both using deterministic policies to decide what recalled information enters the working context and how it is expressed. Experiments on a development suite and an external benchmark show that FC and PS reduce failure rates and cross‑domain leakage compared to verbatim injection, though they increase personalization failures and do not meet all preregistered criteria.
By Chang Liu, Deliang Ding
TokenBank is a financial infrastructure that uses structured contracts to manage the financial relationship between AI service operators and their users. The contracts cover service‑consumption rights, cash settlements for API‑price differences, financing based on future revenue, and protection claims for specified service failures. Evaluation on nearly 900,000 API requests shows that forwards can reduce average expenditure while increasing variability, and that financing can yield positive contributions under certain capital conditions.
By Cary Chang, Jialin Zhou
The paper studies how to curate the fixed context files that large language model agents load at the start of each session. It models this as a capacitated assortment problem, proving bounds on optimal file size and showing that simply appending all positive‑value instructions can be far suboptimal. The authors analyze censored feedback from past sessions, derive evidence‑collection thresholds for adding instructions, bound regret under limited human review, and empirically demonstrate that irrelevant rules in real context files hurt model compliance.
By Zexuan Liu, Yuning Yang, Tiancheng Zhao
AgentHorizon is a benchmark comprising 1,373 computer‑use tasks with instruction‑trajectory pairs collected from 166 hours of human‑recorded activity across three operating systems. The benchmark uses a paired design that swaps instructions to create negative tasks, testing judges’ ability to distinguish successful trajectories from those that completed a similar but incompatible request. Eleven judges were evaluated, with GPT‑5.5 achieving the highest balanced accuracy of 80.9% on the full trajectory subset, while tool‑use benefits some models but harms others, revealing large variability in judges’ performance.
By Xing Han L\`u, Dheeraj Vattikonda, Sina Hajimiri, Fatemeh Pesaran Zadeh, Parishad BehnamGhader, Ghazwa Darwiche, Amirhossein Kazemnejad, Christopher Pal, Alexandre Drouin, Siva Reddy
StoreBench is a live‑commerce environment where an autonomous agent runs a mid‑size online apparel store on a production‑grade backend, facing continuous customer orders, supplier price changes, and market shocks. The agent uses the same 29 merchant tools as a human operator, with a windowed operation budget that ties simulated time to actions, ensuring that model latency does not affect the simulation. Seven large language models were evaluated across 11 scenarios of 30–45 days and a full simulated year, but none matched the scripted smart‑triage policy, and human experts outperformed all models.
By Daksh Raghuvanshi, Ved Vedere, Yifan Wang
The paper introduces E‑Ledger, a multi‑agent harness that ensures safe and persistent execution of large language model agents in partially observable environments. It adds a code approval layer to enforce policy compliance and maintains a world ledger of verified hidden rules and dynamic state. To discover unknown rules, the authors propose WorldAbduct, an abductive framework that diagnoses execution trajectories, hypothesizes latent rules, and validates them before integrating them into the ledger, improving safe task completion on enterprise and scientific benchmarks.
By Yisen Gao, Yue Guo, Qing Zong, Yiwen Guo, Yangqiu Song
ReCast is a new method for learning step representations in large language model (LLM) based agent systems, aimed at improving failure attribution by transforming hidden states into attribution‑oriented representations. The approach selects relevant layers, builds pattern and deviation features, and trains an encoder with contrastive and ranking objectives. In experiments on four benchmarks, ReCast outperforms the strongest baseline, achieving higher Hit@1 scores on the Who&When Algorithm and Handcrafted datasets.
By Weilin Jin, Mingyu Wang, Taiyu Zhu, Ziqi Zhou, Wenbo Li, Haoyang Huang, Nan Duan, Yifan Wu, Ying Li, Zhonghai Wu