arXiv cs.AI· Moises Andrade, Joonhyuk Cha, Brandon Ho, Vriksha Srihari, Karmesh Yadav, Zsolt Kira·· 3 小时前精选AI 评分65
ICLR 2026 论文提出 Self-Grounded Verification 缓解 MLLM 验证器的 agreement bias
Let's Think in Two Steps: Mitigating Agreement Bias in MLLMs with Self-Grounded Verification
AI 导读
arXiv:2507.11662 论文评估MLLM作为智能体验证器，发现其普遍倾向过度认可智能体行为的agreement bias，且测试时扩展难以缓解。
推荐理由
论文基于13个模型和28种评审设计识别出MLLM验证器的agreement bias，并给出可复用的SGV两步验证方法与量化收益。
正文 · 原文
Abstract:Verifiers--functions assigning rewards to agent behavior--have been key to AI progress in math, code, and games. However, extending gains to domains without clear-cut success criteria (e.g., computer use) remains a challenge: while humans can recognize desired outcomes, translating this intuition into scalable rules is nontrivial. Multimodal LLMs (MLLMs) offer a promising solution, given their world knowledge, human-preference alignment, and reasoning skills. We evaluate MLLM verifiers across web navigation, computer use, and robotics, spanning 13 models, 28 judge designs, diverse agent trajectories, and distinct downstream applications. We identify a critical limitation: a strong tendency for MLLMs to over-validate agent behavior, a phenomenon we term agreement bias. This bias is pervasive across models, is resilient to test-time scaling, and harms pipelines relying on MLLM judgments (e.g., self-improvement). We discuss principles for design and evaluation of verifiers, and introduce SGV, a method that improves verification by harnessing MLLMs' own capabilities via modulation of (un)conditional generation. First, an MLLM generates broad priors about desired behavior, independent of the data under evaluation. Then, conditioned on generated priors, it reasons over and judges a candidate trajectory. Our methods yield more human-aligned verifiers, with gains of up to 25pp in failure detection, 14pp in accuracy, and benefits extending to downstream applications. In self-improvement and online supervision, our verifiers boost task completion of a GUI specialist in OSWorld, a diffusion policy in robomimic, and a ReAct agent in VisualWebArena--surpassing the previous SOTA by 20pp. Finally, we release an updated version of VisualWebArena, featuring improved evaluators, high-fidelity environment parallelism with >10x speedups, and VWA-Lite, a 1/3-scale subset with comparable evaluation fidelity.
| Comments: | ICLR 2026. Code, models, and data publicly available at this https URL |
| Subjects: | Artificial Intelligence (cs.AI); Computation and Language (cs.CL); Machine Learning (cs.LG); Multiagent Systems (cs.MA); Robotics (cs.RO) |
| Cite as: | arXiv:2507.11662 [cs.AI] |
| (or arXiv:2507.11662v4 [cs.AI] for this version) | |
| https://doi.org/10.48550/arXiv.2507.11662 arXiv-issued DOI via DataCite |
|
| Journal reference: | Proceedings of the Fourteenth International Conference on Learning Representations (ICLR 2026) |
Submission history
From: Moises Andrade [view email]
[v1]
Tue, 15 Jul 2025 18:50:29 UTC (37,947 KB)
[v2]
Tue, 23 Dec 2025 11:29:24 UTC (45,504 KB)
[v3]
Sun, 8 Mar 2026 00:45:24 UTC (13,237 KB)
[v4]
Wed, 11 Nov 2026 02:30:46 UTC (13,237 KB)
来源：arXiv cs.AI · arxiv.org