发表机构
Thinking Machines Lab; DeepSeek-AI（思维机器实验室; 深度求索公司）
机构由 AI 辅助整理，请以论文原文为准。AI 中文总结
该研究针对多模态推理器的感知蒸馏信用分配问题，提出感知修正蒸馏（PCD）方法，经8个基准测试，可提升不同规模模型的宏平均和结果，消融实验验证其有效性。
AI 中文摘要
在线策略蒸馏为多模态推理器提供密集监督，但其轨迹级奖励无法判断错误答案源于感知还是后续推理。感知成功率（PSR）由共享同一感知的多次推理估计，却因低成功率将感知不足与推理难度混淆而模糊不清。我们提出感知修正蒸馏（PCD），这是一种无标签方法，利用下游失败和师生分歧作为互补证据识别可修正的感知失败，二者的乘积形成软与门，仅当两种证据均存在时才强化蒸馏。我们通过贝叶斯证据组合论证该规则，证明乘法是唯一归一化双线性门，当任一证据缺失时会消失。PCD采用分离的感知-推理展开和均值保持权重，使推理目标保持不变。在8个基准测试中，PCD将8B 2B宏平均从使用在线策略蒸馏（OPD）的44.50提升至47.28，将32B 8B结果从56.94提升至61.22；在匹配的2B消融实验中，移除PCD和分离展开分别使保留平均降低2.22和0.88个百分点。因此，有效的多模态蒸馏不仅取决于教师的预测，还取决于识别何时应修正感知。
英文摘要
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.
