On-policy self-distillation (OPSD) has recently emerged as an effective post-training paradigm that provides dense token-level supervision from a privileged self-teacher, yet it remains underexplored for Video Large Language Models (Video-LLMs). Existing methods typically construct privileged teachers by augmenting their context with additional information while keeping the primary input unchanged for both teacher and student. In video reasoning, however, privileged information can also be found within the primary input itself: videos contain substantial temporal redundancy, and only a few frames carry the evidence needed to answer a question. Building on this observation, we present $\textbf{Video-OPSD}$, an OPSD framework that exploits privileged visual evidence for both self-teacher construction and knowledge transfer. First, our Evidence-Grounded Self-Teacher conditions the teacher exclusively on annotated evidence frames, while the student reasons over a broader set of frames that contains them. This focused visual context enables more targeted supervision without introducing visual content unavailable to the student. Second, our Evidence-Guided Token Optimization adaptively weights token-level distillation according to an estimate of each reasoning token's reliance on privileged visual evidence, thereby emphasizing perceptually grounded reasoning. Across five video benchmarks and three Qwen-VL backbones, $\textbf{Video-OPSD}$ consistently outperforms Standard OPSD by 1.8-2.3 percentage points on average, establishing an effective and efficient post-training approach for Video-LLMs.
翻译:暂无翻译