Search papers, labs, and topics across Lattice.
This paper introduces Perception-Correction Distillation (PCD), a novel label-free method that enhances on-policy distillation for multimodal reasoners by identifying correctable perception failures through a combination of downstream failure and teacher-student disagreement. By employing a soft AND gate mechanism, PCD strengthens the distillation process only when both indicators signal a perception issue, effectively isolating perceptual insufficiencies from reasoning difficulties. The method demonstrates significant improvements across eight benchmarks, with notable increases in performance metrics for both 8B and 32B models, highlighting the critical role of perception correction in multimodal reasoning tasks.
Identifying when perception, rather than reasoning, is the failure point can boost multimodal reasoning performance by over 2.5 points on key benchmarks.
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.
