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

Verifiers—functions assigning rewards to agent behavior—have been key to AI progress in domains such as math, code, and games. However, extending these 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) emerge as a promising solution, given vast world knowledge, human-preference alignment, and reasoning capabilities. We evaluate MLLMs as verifiers across web navigation, computer use, and robotics, spanning 13+ model families, 28+ evaluation templates, curated trajectories from diverse agents and of varying lengths, and distinct verifier 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, resilient to test-time scaling, and can harm methods relying on MLLM evaluations, such as filtered behavior cloning and self-improvement. We provide guidance on the design and evaluation of MLLM verifiers, and introduce Self-Grounded Verification (SGV), a lightweight method that harnesses MLLMs' own sampling mechanisms by modulating (un)conditional generation to better leverage their knowledge, alignment, and reasoning. SGV operates in two steps: first, the MLLM is elicited to generate broad priors about desired behavior, independent of the data under evaluation. Then, conditioned on self-generated priors, it reasons over and evaluates a candidate trajectory. Our methods yield gains across models and environments, improving failure detection by up to 25pp and accuracy by 14pp, with benefits extending to downstream applications. In self-improvement and online supervision, SGV boosts task completion of a GUI specialist in OSWorld, a diffusion policy in robomimic, and a ReAct agent in VisualWebArena—setting a new state of the art, surpassing the previous best by 20pp. Finally, we release an updated version of VisualWebArena featuring strong agent baselines, more human-aligned evaluators, high-fidelity environment parallelism, runtime speedups exceeding 10x, and VisualWebArena-Lite, a 1/3-scale subset with comparable evaluation fidelity. Our code, models, and data are publicly available at our project page.