OPD-V: Visual On-Policy Self-Distillation with Modality Balance
OPD-V selects self-distillation tokens by measuring whether visual input influences generation, addressing text dominance that can blunt multimodal reasoning training.
OPD-V uses a zoomed image as a positive teacher and a masked image as a negative teacher. Their logit differences define a modality-balance trust region for selecting on-policy tokens. Tests cover **6 benchmarks**, **4 MLLM backbones**, and **5 post-training methods**.
Builders training vision agents should inspect whether generated reasoning actually depends on the image, not merely whether privileged inputs are available. Token selection based on modality balance may make self-distillation more targeted and reduce wasted training work.
OPD-V uses a zoomed image as a positive teacher and a masked image as a negative teacher. Their logit differences define a modality-balance trust region for selecting on-policy tokens. Tests cover **6 benchmarks**, **4 MLLM backbones**, and **5 post-training methods**. Builders training vision agents should inspect whether generated reasoning actually depends on the image, not merely whether privileged inputs are available. Token selection based on modality balance may make self-distillation more targeted and reduce wasted training work. The abstract reports consistent reasoning gains and lower training cost but gives no effect sizes, task-level results, or implementation details. It remains unclear how well the method transfers beyond the tested backbones and benchmarks.
OPD-V narrows multimodal self-distillation from general teacher–student disagreement to whether each generated token is genuinely supported by vision. It complements cross-view supervision and other on-policy distillation methods with a modality-balance criterion, but the missing effect sizes leave its relative value and deployment economics unresolved.