Question about the rule-based criteria for GUI-Critic-R1 dataset construction

[ChencongZJU]

Hello,

Thanks for your great work. I have a question regarding the construction of the critic dataset as described in the paper. I would like to understand the specifics of the rule-based criteria used during the Negative Operations Sampling stage.

• Is the evaluation based on a direct match with the ground truth operations?

• For actions formatted as "click + coordinate," how is the correctness judged? Since any coordinate that falls within the correct bounding box is technically a valid action, how is this handled in your evaluation criteria?