Enterprise AI Analysis

Understanding Agreement Bias in MLLMs

Our research identifies a critical limitation in MLLMs as verifiers: a strong tendency to over-validate agent behavior, termed agreement bias. This phenomenon is pervasive across various models and evaluation settings, leading to flawed judgments and hindering effective feedback for AI agents. It impacts critical applications like self-improvement and online supervision by validating incorrect actions, even with elaborate chains-of-thought.

This bias persists despite MLLMs exhibiting human-aligned priors, suggesting a bottleneck in knowledge extraction and utilization within current verification paradigms. Addressing this requires a method that better leverages MLLMs' inherent capabilities for reasoning and alignment.

Self-Grounded Verification (SGV): A Novel Approach

To counteract agreement bias, we propose Self-Grounded Verification (SGV), a lightweight, zero-shot method. SGV modulates MLLMs' sampling mechanisms through a two-step process:

1. Prior Generation: The MLLM first generates broad priors about desired behavior, conditioned on partial task information. This allows the model to freely extract pertinent knowledge.
2. Trajectory Evaluation: The MLLM then reasons over and evaluates a candidate trajectory, critically conditioned on its self-generated priors.

This approach significantly improves MLLM-based verification by enabling more effective use of their knowledge, alignment, and reasoning, leading to more balanced and human-aligned judgments.

Transforming Downstream AI Applications

SGV's enhanced verification directly translates to significant improvements in downstream applications:

• Self-Improvement: Stronger SGV-based verifiers lead to gains of up to 10pp (24% relative) on VisualWebArena, by providing accurate corrective signals.
• Online Supervision: SGV boosts task completion rates by 9pp (20%) on VisualWebArena and 5pp (22%) on OSWorld, by encouraging agents to backtrack from greedy strategies and avoid suboptimal behavior.

These results set new state-of-the-art benchmarks, demonstrating SGV's potential to drive more reliable and effective AI agent development across web navigation, computer use, and robotics.