Tag
Proposes NOPD, a self-distillation method that improves vision-language models without external supervision by leveraging prediction discrepancies between clean and corrupted inputs. Achieves significant gains on visual reasoning tasks, matching or exceeding RL and distillation from external models.