Results on POPE for LLaVA-1.5

[xlchen0205]

Hello,

A pretty novel work! But I have some questions.

I notice that in Table 2 of your paper, the base ACC & F1 on POPE for LLaVA-1.5 is 84.07 and 82.62,

However, in other works like https://github.com/Lackel/AGLA, the corresponding ACC & F1 are 76.03 and 76.26,

Could you tell me why there is a such discrepancy?
Thank you!

[jhkim0911]

Hello! Thank you for your interest in our work.

This may be mainly because of the model param size (ours: 13B, An's work: 7B).
Also, different setting of seed / attn implementation / decoding strategy can affect the model's results.

Hope this is helpful for your concerns.