Back to AI Research

AI Research

Correcting What You Cannot See: Credit Assignment f... | AI Research

Key Takeaways

  • Perception-Correction Distillation (PCD) is a method designed to improve how multimodal large language models (MLLMs) learn to interpret images.
  • On-policy distillation provides dense supervision for multimodal reasoners, but its trajectory-level reward cannot determine whether a failed answer arose from perception or subsequent reasoning.
  • Perception Success Rate (PSR), estimated from multiple reasonings sharing one perception, remains ambiguous because low success conflates perceptual insufficiency with reasoning difficulty.
  • We introduce \textbf{Perception-Correction Distillation (PCD)}, a label-free method that identifies correctable perception failures using downstream failure and teacher--student disagreement as complementary witnesses.
  • Their product, , forms a soft AND gate that strengthens distillation only when both witnesses are present.
Paper AbstractExpand

On-policy distillation provides dense supervision for multimodal reasoners, but its trajectory-level reward cannot determine whether a failed answer arose from perception or subsequent reasoning. Perception Success Rate (PSR), estimated from multiple reasonings sharing one perception, remains ambiguous because low success conflates perceptual insufficiency with reasoning difficulty. We introduce \textbf{Perception-Correction Distillation (PCD)}, a label-free method that identifies correctable perception failures using downstream failure and teacher--student disagreement as complementary witnesses. Their product, , forms a soft AND gate that strengthens distillation only when both witnesses are present. We motivate this rule through Bayesian evidence combination and show that multiplication is the unique normalized bilinear gate that vanishes when either witness is absent. PCD uses separated perception--reasoning rollouts and mean-preserving weights, leaving the reasoning objective unchanged. Across eight benchmarks, PCD improves the 8B 2B macro average from 44.50 with OPD to 47.28 and the 32B 8B result from 56.94 to 61.22. In matched 2B ablations, removing PCD and separated rollout reduces held-out average by 2.22 and 0.88 points, respectively. Effective multimodal distillation therefore depends not only on what the teacher predicts, but also on identifying when perception is the appropriate target of correction.

Perception-Correction Distillation (PCD) is a method designed to improve how multimodal large language models (MLLMs) learn to interpret images. When these models fail to answer a question correctly, it is often unclear whether the error occurred because the model misread the image (perception) or because it failed to reason correctly from the information it gathered. PCD solves this "credit assignment" problem by identifying which failed perceptions are actually correctable by a teacher model, allowing the student to focus its learning on those specific visual mistakes.

The Problem of Ambiguous Failure

In standard on-policy distillation, models are trained using rewards based on the final answer. However, a zero reward is ambiguous: it does not distinguish between a bad initial observation and a flawed reasoning process. While researchers often use the Perception Success Rate (PSR)—the average reward of multiple reasoning attempts sharing the same perception—this metric is confounded. A low PSR could mean the perception was bad, or it could simply mean the reasoning task was too difficult. Because of this, uniform distillation often wastes training effort by correcting perceptions that were already sufficient.

How PCD Works

PCD introduces a "soft AND gate" to determine when to apply corrective supervision. It uses two "witnesses" to identify a correctable perception failure: 1. Downstream Failure: A low Perception Success Rate (PSR), indicating the model is struggling to reach the correct answer from a given observation. 2. Teacher-Student Disagreement: A high KL divergence between the student and the teacher on the perception span, indicating the teacher would have interpreted the image differently.
The method multiplies these two factors to create a deficiency score. This ensures that the model only receives extra supervision when both conditions are met: the student failed to solve the problem, and the teacher disagrees with the student’s visual interpretation. If the student and teacher agree, or if the student is already succeeding, the model does not receive additional, potentially unnecessary, correction.

Experimental Results

The authors tested PCD across eight benchmarks using two transfer settings: Qwen3-VL-8B to 2B and 32B to 8B.

  • For the 8B to 2B transfer, PCD improved the macro average from 44.50 (using standard on-policy distillation) to 47.28.

  • For the 32B to 8B transfer, the result improved from 56.94 to 61.22.
    Ablation studies on the 2B model showed that removing the PCD weighting reduced the held-out average by 2.22 points, while removing the separated perception-reasoning rollout reduced it by 0.88 points.

Key Considerations

PCD is designed to be a label-free method, meaning it does not require human-annotated data to identify perception errors. It relies on the teacher model's existing knowledge to guide the student. Because the method uses a mean-preserving weight, it reallocates the existing teacher-supervision budget rather than increasing the total amount of supervision. The authors note that this approach is complementary to other techniques, such as token-level visual focusing, and can be used alongside them to further refine model performance.

Comments (0)

No comments yet

Be the first to share your thoughts!