PCD Separates Multimodal Failures into Perception and Reasoning Signals, Raising the 32B-to-8B Distillation Average to 61.22

When a multimodal reasoning model gives an incorrect answer, the final reward usually cannot reveal whether the model misread the image first or perceived it correctly but made an error during deduction. Conventional on-policy distillation applies the teacher signal across the student’s entire sampled trajectory. This may repeatedly correct perception that is already adequate while allowing the longer reasoning tokens to dominate the gradient. The newly proposed Perception-Correction Distillation (PCD) explicitly divides the output into a perception segment, `<aware>`, and a reasoning segment, `<cot>`, and trains them using distillation and verifiable rewards, respectively.

For each image-question pair, the method first samples multiple perception descriptions and then generates several reasoning trajectories from each fixed description. The study estimates downstream performance using the average success rate under the same perception, called PSR. However, it notes that a low PSR may result from insufficient perception, a difficult question, or a weak reasoner, so PSR alone cannot identify the cause. PCD therefore also calculates the KL divergence between the teacher’s and student’s perception tokens and uses `(1−PSR)×KL` as a soft-AND gate: the distillation weight for a perception is increased only when the outcome is poor and the teacher disagrees with what the student perceived. Within each batch, the weights are renormalized to have a mean of one, preventing the selection of which examples to train from being conflated with an increase in overall distillation strength.

Across eight multimodal benchmarks for Qwen3-VL-8B→2B and 32B→8B, PCD achieves macro averages of 47.28 and 61.22, respectively, compared with 44.50 and 56.94 for uniform OPD. In the 2B ablations, removing the PCD weighting or the decoupled rollout reduces the held-out average by 2.22 and 0.88 points, respectively. The method requires neither perception labels nor an additional discriminator, but it increases rollout computation for each perception branch.

The main limitation is that KL divergence represents only a “teacher-correctable difference,” not a calibrated probability that the perception is true or false. If the teacher and student share the same blind spot, or if the teacher performs worse on the relevant visual evidence, the multiplicative gate may miss the error or even amplify correction in the wrong direction. The paper also provides no public code or training checkpoints. The next step is to reproduce the results across different model families, non-mathematical visual tasks, and fixed GPU budgets to verify that the gains do not come from additional branch sampling or a specific output-tag format.