Vision-OPD: On-Policy Self-Distillation Closes the Regional-to-Global Perception Gap in Multimodal LLMs

arXiv·medium signal

Discovers that multimodal LLMs answer fine-grained questions more accurately when shown evidence-centered crops versus full images, revealing failures stem from inability to focus rather than lacking recognition ability. Proposes Vision-OPD: a crop-conditioned teacher and full-image student from the same model, trained via token-level divergence minimization on on-policy rollouts. No external teacher needed.