The paper investigates whether latent world models truly capture the underlying dynamics of their environments. It introduces the idea of faithful latent world models, defined as an equivariance between the environment’s transition map and the learned latent dynamics, and proposes latent rollout consistency (LRC) as a metric for evaluating action‑conditioned latent predictions. Experiments show that none of the state‑of‑the‑art latent world models studied actually faithfully model the environment’s dynamics, despite solving the tasks they were trained on.
The paper investigates how vision–language models (VLMs) diagnose failures in bimanual tool‑use tasks, where failures can arise from tool slip, timing mismatches between hands, or combined causes. It evaluates which sensing channels enable a diagnoser to identify each failure type, finding that tool slip is detectable on the tool hand, timing delays are observable on both arms, and some combined failures remain indistinguishable from passive sensing. The study also shows that a repair policy that escalates when the cause is unclear can match the performance of a policy with full cause knowledge, and that a VLM’s internal state captures the late hand almost perfectly, though the answer depends on question wording.
The paper presents the first strategyproof deep reinforcement learning mechanism for online resource allocation, addressing efficient allocation under uncertain demand and truthful reporting of task valuations. It introduces a monotonic neural network to capture nonlinear interactions, a generalized critical payment rule, and training via RL or imitation of existing rules. Experiments on a four‑resource stream show the neural rule achieves 97.8% of an offline welfare upper bound, outperforming linear learners and first‑come, first‑served, while imitation training reaches similar performance in 261 seconds—about 89 times faster than RL.
M4Bench is a benchmark comprising 28 clinical derivation tasks across 15 families on MIMIC‑IV and eICU datasets, designed to evaluate whether AI coding agents can reproduce complex procedural specifications for electronic health record (EHR) variable derivation. The study shows that incorporating specification completeness into agent skills yields the largest performance gains, with modest additional benefits from delivery format and schema familiarity. The benchmark, along with tasks, skills, evaluation code, and sanitized evidence, is released as open source for clinician‑reviewable use.
The paper investigates how communication affects trading agents that operate independently rather than as a coordinated team. Over seven trading days, 200 agents from four model families were grouped under five communication regimes, from isolation to shared message boards, and their reading logs were recorded before each decision. Results showed that in two model families, communicating groups held more of the same stocks than isolated groups, and a group’s return was largely driven by shared holdings; agents were more likely to reuse phrases from groupmates, and communicating groups tended to sell in step more often than isolated ones, though isolated groups also converged on similar stocks, indicating that herding assessments must compare against a no‑peer baseline.
The article describes an experiment where an AI agent was able to identify incorrect model decisions while interacting with a web page, yet failed to correct the errors. By replaying correct decisions through four intentionally flawed browser actions, the author demonstrates that a simple page check missed half of the erroneous clicks. This illustrates that detecting a fault and successfully recovering from it are distinct capabilities.
By Abdullahi Dattijo
The article describes an experiment where the author intentionally broke a browser agent by replaying correct model decisions through four faulty browser actions. A basic page check failed to detect half of the incorrect clicks and incorrectly flagged every successful click within an embedded page. The piece highlights the challenges of diagnosing and resolving such browser-related faults.
By Abdullahi Dattijo
Retell leverages OpenAI models to support 100 million monthly calls, with GPT‑6 Astra achieving top performance on its voice agent benchmark. Codex contributes to the company’s financial success, helping it reach a $1.6 million annual revenue run rate per employee.
The paper introduces the Preference Dependency Graph Framework (PDG‑F) for narrative‑driven recommendation, where user requests are parsed into a directed acyclic graph that captures how preferences qualify and condition one another. PDG‑A builds this graph from free‑form text, retrieves candidate evidence via directed path retrieval, and ranks recommendations using topological path ranking. Experiments across English and Chinese domains show PDG‑F outperforms baselines, retrieval agents, and state‑of‑the‑art language models, highlighting the importance of modeling preference dependencies.
The article titled "Who Watches the AI?" discusses the concept of a separation of powers for controlling autonomous agents. It explores how governance structures can be applied to ensure accountability and oversight of AI systems. The piece was originally published on Towards Data Science.
By Gadi Singer
The paper investigates how language model agents learn from the harnesses—prompts, tools, and control logic—used during post‑training and deployment. It shows that under KL‑regularized control, a policy’s log odds encode the harness’s action preferences, meaning a trained policy carries those preferences. The study demonstrates that changing the harness alters both the achievable return (the ceiling) and the policy’s mismatch, and that comparing agents across different harnesses conflates these effects. Experiments on MiniWoB++ and Terminal‑Bench Lite reveal that a model’s success can drop dramatically when the harness is altered, while new harnesses that raise the ceiling can improve performance after further training, with ceiling and mismatch predicting these gains.
The paper introduces BiPA, a method to evaluate skill optimizers that rewrite natural‑language procedures for agents performing long‑horizon tasks. BiPA replays the optimizer’s acceptance rule on identical candidates to isolate selection effects and records each episode as a bipartite graph of turns and read text, revealing that learned skills only marginally outperform hand‑written seeds and that many gains vanish on test data. The study finds optimizers add extra input tokens and turns, and proposes four design principles for more reliable skill editing.
The paper introduces EvidenceConnect, a data‑synthesis framework that generates multimodal questions by zooming into related image regions to strengthen the capability of connecting evidence—identifying visual facts, linking them to entities, and composing them into answers. By training vision‑language models on this synthesized data with RLVR, the authors produce Zooming without Zooming v2 (ZwZ2), which improves fine‑grained and spatial perception, embodied spatial reasoning, multimodal long‑context understanding, and general visual reasoning. The results demonstrate that targeting the shared evidence‑connection capability yields more generalizable multimodal perception without tailoring data to specific benchmarks.
RubricAtlas is a framework that extracts type-level analytical requirements from expert reports to evaluate deep research agents. It organizes detailed binary checkpoints around recurring analytical questions, linking them to report passages and separating completeness from argument validity. Experiments on 3,079 Chinese financial reports show that individual reports cover only about half of their type’s questions, and that using different reports for requirements changes evaluation scores and agent rankings, with high agreement between judge pipelines.
The paper introduces Policy-Compatible Reinforcement Learning (PICO), a method that combines distribution-constrained joint decoding with on-policy rollouts to address the learning cliff in reinforcement learning with verifiable rewards (RLVR). PICO guides token-level decoding using expert data while preserving the model’s relative token probabilities, and employs a dynamic top‑k mechanism to keep sampling within a finite divergence budget. This approach yields verified, policy‑compatible trajectories that are jointly optimized with on‑policy rollouts, leading to consistent performance gains over standard RLVR and prior off‑policy baselines across diverse reasoning datasets and agentic benchmarks.
The study examines how large language models (LLMs) predict users’ decisions in human‑agent dialogue by combining users’ statements with past decision records. Using 1,260 rounds of a repeated prisoner's dilemma with 42 users and five agents, the researchers found that decision records alone yielded strong predictive accuracy (AUC = .80), while adding dialogue did not improve overall performance but did increase agreement with users’ statements. However, predictions over‑emphasized statements, especially when users had not yet made a decision, and failed to capture variations in cooperation that depended on the specific agent or users’ trait alignment.
OpenAI and Oracle are expanding their partnership to integrate business applications and AI agents directly into ChatGPT. This collaboration aims to power new Oracle offerings by leveraging OpenAI’s technology within the ChatGPT platform.
Self-improving AI systems can revise the designs that govern their own behavior. Yet existing approaches center improvement on selected parent agents and their evolutionary experience, leaving the bro...
arXiv:2610.33907v1 Announce Type: new
Abstract: Computer-using agents are increasingly capable of performing complex tasks in the digital world. However, high-quality interaction trajectories remain...
By Zhoumianze Liu, Zichen Ding, Bowen Yang, Zehao Li, Zhenyu Wu, Yibo Zhao, Kanzhi Chen, Yian Wang, Yanan Sun, Qiushi Sun, Yu Qiao
arXiv:2610.35219v1 Announce Type: new
Abstract: Existing multi-agent collaboration (MAC) methods have primarily explored a group thinking collaboration topology, where agents independently generate s...
By Xu Liu, Jungang Qiao, Guikun Chen, Xiaojuan Qi, Wenguan Wang