OPD-V: Visual On-Policy Self-Distillation with Modality Balance

OPD-V argues that modality imbalance limits on-policy self-distillation in multimodal large language models: when text dominates generation, the model underuses visual information and privileged supervision. The method builds a Positive Teacher from a zoomed-in image and a Negative Teacher from a masked image. Their logit differences define Positive Modality-Balance Logits Margins and a Modality-Balance Trust Region, which selects on-policy tokens for distillation. The abstract reports consistent reasoning gains across 6 benchmarks, 4 MLLM backbones, and 5 post-training methods, alongside reduced training cost, but provides no exact improvement figures.