The question of model quality

[jjunsss]

Hi authors,

First of all, thank you very much for your outstanding contributions to the vision-language model community! Your work, especially regarding the Chain-of-Manipulations (CoM) process introduced in your paper, is very inspiring.

I am currently exploring your code to reproduce and perform inference. However, I'm encountering some difficulties obtaining the correct intermediate manipulation results during inference, even though the final answers are interestingly correct in some cases.

For your reference and better understanding, I have attached some example images below:

Specifically, in these examples, I am unsure why the OCR manipulation is being triggered or used to generate the answers.

Could you please let me know if my inference process might be incorrect or if these images are particularly challenging? Any guidance or tips you could provide to help achieve the correct intermediate manipulation processes would be greatly appreciated.

Thank you very much for your valuable time and assistance!
Best regards,