Correcting What You Cannot See: Credit Assignment for Perception Distillation in Multimodal Reasoners

Analysis

Why This Paper Matters

Multimodal reasoners, which combine perception and reasoning, are typically trained via distillation from larger models. However, standard on-policy distillation provides only trajectory-level rewards, making it impossible to attribute a failed answer to either perception or reasoning errors. This paper addresses a critical gap: how to correct perception errors without explicit labels. The proposed method, Perception-Correction Distillation (PCD), offers a principled, label-free approach to identify when perception is the culprit, enabling targeted supervision.

The significance lies in its potential to improve the efficiency of training smaller multimodal models. By focusing distillation on perception failures, PCD avoids over-correcting reasoning errors and instead provides dense, targeted supervision. This is particularly important as multimodal models become more complex and data-hungry; effective distillation can reduce the need for large-scale labeled data.

Technical Contributions

• Complementary Witnesses: PCD uses two signals: downstream failure (the model's final answer is wrong) and teacher-student disagreement (the student's perception differs from the teacher's). The product of these signals forms a soft AND gate, ensuring that distillation is strengthened only when both witnesses are present.
• Bayesian Justification: The authors derive the multiplication rule from Bayesian evidence combination, proving that it is the unique normalized bilinear gate that vanishes when either witness is absent. This provides a theoretical foundation for the heuristic.
• Separated Rollouts: PCD uses separated perception and reasoning rollouts, which allows the method to isolate perception errors from reasoning errors. This is a key innovation that enables the credit assignment.
• Mean-Preserving Weights: The method uses mean-preserving weights, ensuring that the reasoning objective remains unchanged, so PCD only affects the perception-related supervision.

Results

Across eight benchmarks, PCD improves the macro average from 44.50 with OPD to 47.28 for the 8B→2B distillation, and from 56.94 to 61.22 for the 32B→8B setting. These are substantial gains, indicating that PCD effectively boosts the performance of smaller models. In matched 2B ablations, removing PCD reduces held-out average by 2.22 points, and removing separated rollout reduces it by 0.88 points, confirming the importance of both components.

Significance

This work advances the field of multimodal distillation by shifting the focus from what the teacher predicts to when perception is the appropriate target of correction. It provides a principled, label-free method that can be integrated into existing distillation pipelines. The theoretical grounding and empirical results suggest that PCD could become a standard tool for training efficient multimodal reasoners, potentially reducing the need for large-scale labeled data and enabling deployment on resource-constrained devices. Future work could explore extending PCD to other modalities or tasks, and investigating its interaction with different teacher architectures.