Perception-Correction Distillation: A New Method for Multimodal Reasoners

A new research paper introduces Perception-Correction Distillation (PCD), a label-free method to improve multimodal reasoners by identifying and correcting perception failures. The paper, available on arXiv (2607.28336), addresses the challenge of on-policy distillation where trajectory-level rewards cannot distinguish whether a failed answer stems from perception or reasoning. The authors propose using downstream failure and teacher-student disagreement as complementary witnesses, forming a soft AND gate through their product. This gate strengthens distillation only when both witnesses are present, based on Bayesian evidence combination. The method uses separated perception-reasoning rollouts and mean-preserving variance reduction. The paper is categorized under AI and machine learning, focusing on improving the training of multimodal AI systems.