# Let's Think in Two Steps: Mitigating Agreement Bias in MLLMs with Self-Grounded Verification

DevFeed: [Let's Think in Two Steps: Mitigating Agreement Bias in MLLMs with Self-Grounded Verification](<https://devfeed.tech/articles/let-s-think-in-two-steps-mitigating-agreement-bias-in-mllms-with-self-grounded-verification-131846.md>)

Original publisher: [Read original article](<https://arxiv.org/abs/2507.11662>)

Author: Moises Andrade, Joonhyuk Cha, Brandon Ho, Vriksha Srihari, Karmesh Yadav, Zsolt Kira

Published: 2026-11-13T05:00:00Z

Content type: article

Language: en

Sources: [arXiv: Artificial Intelligence](<https://devfeed.tech/sources/arxiv-artificial-intelligence.md>)

Topics: [llm-reasoning](<https://devfeed.tech/topics/llm-reasoning.md>), [AI Chat](<https://devfeed.tech/topics/ai-chat.md>), [Machine learning](<https://devfeed.tech/topics/machine-learning.md>), [ai alignment](<https://devfeed.tech/topics/ai-alignment.md>), [ReAct agent Python](<https://devfeed.tech/topics/react-agent-python.md>)

Tags: [agent](<https://devfeed.tech/tags/agent.md>), [agent-behavior](<https://devfeed.tech/tags/agent-behavior.md>), [ai](<https://devfeed.tech/tags/ai.md>), [alignment](<https://devfeed.tech/tags/alignment.md>), [applications](<https://devfeed.tech/tags/applications.md>), [bias](<https://devfeed.tech/tags/bias.md>), [code](<https://devfeed.tech/tags/code.md>), [completion](<https://devfeed.tech/tags/completion.md>), [computer](<https://devfeed.tech/tags/computer.md>), [computer-use](<https://devfeed.tech/tags/computer-use.md>), [conditional](<https://devfeed.tech/tags/conditional.md>), [cs-ai](<https://devfeed.tech/tags/cs-ai.md>), [cs-cl](<https://devfeed.tech/tags/cs-cl.md>), [cs-lg](<https://devfeed.tech/tags/cs-lg.md>), [cs-ma](<https://devfeed.tech/tags/cs-ma.md>), [cs-ro](<https://devfeed.tech/tags/cs-ro.md>), [detection](<https://devfeed.tech/tags/detection.md>), [diffusion](<https://devfeed.tech/tags/diffusion.md>), [environment](<https://devfeed.tech/tags/environment.md>), [evaluation](<https://devfeed.tech/tags/evaluation.md>), [failure](<https://devfeed.tech/tags/failure.md>), [games](<https://devfeed.tech/tags/games.md>), [generation](<https://devfeed.tech/tags/generation.md>), [gui](<https://devfeed.tech/tags/gui.md>), [verification](<https://devfeed.tech/tags/verification.md>)

## AI overview

Across 13 multimodal language models and 28 verifier designs for web navigation, computer use, and robotics, the study finds agreement bias: models tend to over-validate agent behavior, and test-time scaling does not reliably remove it. The proposed Self-Grounded Verification (SGV) method first generates general priors about desired behavior independently of the trajectory, then uses those priors to assess a candidate trajectory. The authors report gains of up to 25 percentage points in failure detection and 14 points in accuracy. In self-improvement and online supervision, SGV-based verifiers improved task completion for agents in OSWorld, robomimic, and VisualWebArena, surpassing prior results by 20 points. The abstract also reports an updated VisualWebArena with faster environment parallelism and a smaller VWA-Lite subset; it does not provide details of the verifier designs or experimental conditions.

## Source excerpt

arXiv:2507.11662v4 Announce Type: replace 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.