Question about fairness of comparisons with baseline models

[Zhao-Qihao]

Hello, thanks for great work. I have a question about the experimental setup and the fairness of comparisons.
In your paper, UniVG-R1 is trained on the newly constructed CoT dataset, while the baseline models (e.g., Migician, Qwen2-VL, etc.) are directly evaluated without being trained on this dataset. This may introduce a data advantage for UniVG-R1.
Just I wanna know if it is a common practice in this field?

[SuleBai]

Hi. Our main motivation is to verify whether incorporating reasoning-chain data and RL training can improve performance on the multi-image grounding task. To this end, we augment the Migician dataset with reasoning chains and use it as our training data (without introducing any additional external data). We believe this evaluation setting is fair, and our approach demonstrates the effectiveness of reasoning-chain supervision and RL training.