PROBE is an algorithm that trains language‑model auditors to probe deployed tool‑using agents for policy violations through realistic, multi‑turn, tool‑mediated dialogue. By framing auditing as a two‑player Markov game and using a composite reward that balances violation discovery, behavioral plausibility, conversation progress, and policy diversity, PROBE outperforms zero‑shot adversaries in uncovering broader and higher‑yield policy violations. The study compares GRPO, DPO, and GFlowNet optimizations, demonstrating that multi‑turn behavioral auditing is a distinct and necessary axis of agent evaluation, complementing existing capability and safety benchmarks.
By Kshitij Mishra, Abhijith Sharma, Nils Lukas, Salem Lahlou
arXiv:2607. 10814v1 Announce Type: cross Abstract: Evaluating LLM agents in hidden-information multi-agent settings is hard: final outcomes are high-variance and rarely reveal why an agent decided as it did.
By Yuan Gao, Jiangyi Yang, Yao Zhao, Yichi Zhang
The paper studies how to safely delegate action approval to multiple AI reviewers when the reviewers themselves may be misaligned. It introduces a weaker condition—k‑robust coalitional alignment—under which a threshold rule that tolerates up to k disapprovals guarantees that the principal’s expected utility is at least as good as a baseline policy. The authors extend this characterization to sequential decision‑making in discounted MDPs and show that full‑panel coverage of reward functions ensures safety in Nash equilibria, while more permissive thresholds can lead to unsafe outcomes. Experiments demonstrate that collective review can remain sound even when individual reviewers are not fully aligned, provided some disapprovals are allowed.
By Natalie Collina, Surbhi Goel, Aaron Roth, Sikata Bela Sengupta
arXiv:2607. 07435v1 Announce Type: cross Abstract: Agents acting on our behalf in the real world (e.
By Bojie Li, Noah Shi
The paper investigates the problem of sharing a single critic across multiple parallel environments in reinforcement learning. It shows that when environments assign different expected returns to the same state, a shared critic must reconcile conflicting value targets, which can distort advantage estimates and misguide policy updates. The authors propose a simple fix—providing the critic with the environment index—demonstrating through bandit models and experiments on CartPole, MuJoCo, BipedalWalker, and 16 Procgen games that this conditional critic stabilizes learning and boosts returns, achieving a 40.8% improvement in aggregate normalized return on unseen levels.
By Zhenya Liu, Yang Meng, Zhuokai Zhao, Xuefeng Liu, Yuxin Chen
arXiv:2607. 05904v1 Announce Type: new Abstract: Training a language model against its own reference-free judgments (the premise of self-rewarding, self-play, and LLM-as-a-judge pipelines) assumes a model's verdict on a shown answer tracks correctness.
By Chenyu Zhou