BAMI: Training-Free Bias Mitigation Fixes GUI Agent Grounding Errors Without Retraining

arXiv·medium signal

Identifies two primary error sources in GUI agent grounding on complex interfaces: precision bias from high image resolution and ambiguity bias from intricate interface elements. Using a novel Masked Prediction Distribution attribution method, BAMI mitigates both biases at inference time without any additional training. Relevant to anyone building computer-use agents — addresses the practical gap between demo-quality and production-quality GUI interaction.