BAMI: Training-Free Bias Mitigation in GUI Grounding

TL;DR

BAMI introduces a training-free bias mitigation method for GUI grounding tasks, significantly improving model accuracy by addressing image resolution and interface complexity biases.

Contribution

The paper proposes BAMI, a novel bias mitigation technique that enhances GUI grounding accuracy without additional training, using coarse-to-fine focus and candidate selection strategies.

Findings

BAMI improves accuracy of GUI grounding models on ScreenSpot-Pro benchmark.

Applying BAMI to TianXi-Action-7B increases accuracy from 51.9% to 57.8%.

Ablation studies show BAMI's robustness across different parameters.

Abstract

GUI grounding is a critical capability for enabling GUI agents to execute tasks such as clicking and dragging. However, in complex scenarios like the ScreenSpot-Pro benchmark, existing models often suffer from suboptimal performance. Utilizing the proposed \textbf{Masked Prediction Distribution (MPD)} attribution method, we identify that the primary sources of errors are twofold: high image resolution (leading to precision bias) and intricate interface elements (resulting in ambiguity bias). To address these challenges, we introduce \textbf{Bias-Aware Manipulation Inference (BAMI)}, which incorporates two key manipulations, coarse-to-fine focus and candidate selection, to effectively mitigate these biases. Our extensive experimental results demonstrate that BAMI significantly enhances the accuracy of various GUI grounding models in a training-free setting. For instance, applying our method to the TianXi-Action-7B model boosts its accuracy on the ScreenSpot-Pro benchmark from 51.9\% to 57.8\%. Furthermore, ablation studies confirm the robustness of the BAMI approach across diverse parameter configurations, highlighting its stability and effectiveness. Code is available at https://github.com/Neur-IO/BAMI.