A new AI framework called VISD is proposed to improve video reasoning in large language models. Current methods struggle with assigning credit for complex tasks, but VISD introduces a structured self-distillation approach that breaks down reasoning into multiple dimensions, including correctness and temporal grounding. This allows for more fine-grained supervision and better performance on various benchmarks. The framework also incorporates techniques like curriculum scheduling and teacher stabilization to improve optimization efficiency. --- Why it matters: This matters because it addresses the challenge of assigning credit in complex video tasks, which is crucial for developing reliable and efficient AI models. VISD's structured approach can lead to improved performance and sample efficiency, making it a significant advancement in video reasoning research. Source: https://arxiv.org/abs/2605.06094