Paper disentangles visual vs factual correctness in LVLMs

Arxiv CS.CV (computer vision)

The study reveals that LVLMs' responses often rely on factual priors rather than visual evidence. The authors propose a disentanglement method to isolate visual reasoning from learned knowledge. Evaluations across multiple LVLMs show varying degrees of reliance on factual priors.

lvlmvision-language-model

Arxiv CS.CV (computer vision)·Soohyun LeeJaeyoung KimSeokhyeon ParkSihyeon LeeJiwon SongBohyoung KimHyunjoo SongJinwook Seo

8 days ago