论文库

原文摘要

Video reasoning hinges on a small number of visually decisive tokens embedded in otherwise largely language-driven chains of thought. Yet prevailing post-training methods such as GRPO overlook this structure, assigning coarse trajectory-level rewards while requiring multiple rollouts and often costly external judges. We introduce Visually-Aware On-Policy Self-Distillation (VIOS), a post-training framework that identifies where a model under-utilizes visual evidence and concentrates learning at those positions. Our key insight is that a model can audit its own visual reliance: the divergence between its next-token distributions with and without the video reveals how strongly each prediction depends on the visual input. Measuring this reliance for both a student and an answer-privileged teacher exposes a per-token reliance gap, identifying positions where the teacher exploits visual evidence that the student overlooks. VIOS translates this signal into dense, targeted supervision through three complementary components: an exponential moving average (EMA) teacher that co-evolves with the student, a contrastive objective that encourages predictions to depend on the video, and a reliance-gap gate that focuses learning on visually under-utilized tokens. Crucially, VIOS requires only a single on-policy rollout from a single model, without external rewards, judges, or larger teachers. Across 12 benchmarks spanning general, temporal, spatial, knowledge-intensive, and long-video understanding, VIOS consistently improves three model backbones and outperforms contemporary video-reasoning systems, while training in over 9x fewer steps, 60x fewer FLOPs, and 111x fewer generated tokens. Project page: https://vision.cs.utexas.edu/projects/VIOS/