o3 v.s. qwen2.5vl-3b

[Glupayy]

Hi, thank you for your interesting work! It's an impressive finding that simply using BEV + object information to prompt O3 can achieve such high results. I noticed that training qwen3b with the collected data doesn't yield as good results. I'd like to ask whether the prompts given to the models are exactly the same?

[fangruizhu]

Yes, they are the same! But the results of prompting O3 are reported on the subset, not full set. The results with qwen3b are on the full set, so not comparable with results of O3. I think we will add a comparison here based on same testing set. Thank you for pointing it out!