Chain-of-Steps: New Mechanism for Reasoning in Video Diffusion Models
A new study published on arXiv (2603.16870) disputes the common belief that reasoning in diffusion-based video models happens in a sequential manner across frames, known as Chain-of-Frames (CoF). The authors reveal that reasoning predominantly takes place during the diffusion denoising steps, which they refer to as Chain-of-Steps (CoS). Their qualitative analysis and focused probing experiments indicate that models investigate several potential solutions in the initial denoising stages before gradually arriving at a conclusive answer. Additionally, the research highlights important emergent reasoning behaviors that enhance model performance, such as working memory for tasks needing consistent references (like object permanence) and self-correction. This paper marks a significant advancement in comprehending the reasoning abilities of video generation models.
Key facts
- Paper arXiv:2603.16870 challenges Chain-of-Frames (CoF) mechanism.
- Proposes Chain-of-Steps (CoS) as the primary reasoning mechanism.
- Reasoning emerges along diffusion denoising steps.
- Models explore multiple candidate solutions in early denoising steps.
- Models progressively converge to a final answer.
- Identifies working memory as an emergent reasoning behavior.
- Identifies self-correction as an emergent reasoning behavior.
- Based on qualitative analysis and targeted probing experiments.
Entities
Institutions
- arXiv