The paper introduces Visual Attribution Distillation (VAD) for multimodal on-policy distillation. Instead of treating every teacher correction as visual supervision, VAD evaluates a fixed privileged-view teacher with relevant evidence present and removed. The change in centered log-probabilities becomes a signed proxy for the direction and strength of visual evidence. VAD projects the teacher correction onto this proxy, discards the proxy-unexplained component from primary supervision, and reconstructs a student-anchored target. Across six fine-grained visual benchmarks at 4B and 9B scales, the reported results outperform direct privileged-view distillation and visual-advantage weighting.
No heat snapshots are available in the last 24 hours.