WebVoyager Benchmark Evaluation Method: Text-only or Multimodal?

[jimodeniuren]

Thank you for your excellent work on the UI-tars model! I have a couple of questions regarding the evaluation setup for the WebVoyager benchmark mentioned in your paper/repository.

Question 1: Evaluation Modality
When evaluating UI-tars on the WebVoyager benchmark, did you use a text-only evaluation approach or a multimodal evaluation method? Specifically, I'd like to understand whether the model processes screenshots/visual information or relies solely on textual representations (e.g., HTML, accessibility trees) during the evaluation.

Question 2: Evaluation Prompts
Could you share the exact prompts you used during the WebVoyager benchmark evaluation? Understanding the prompt template would help me better reproduce your results and compare with other approaches.