Paper 2603.16870
Demystifying Video Reasoning
- Published
- Mar 2026
- Research lab
- Independent
- Citations
- 6
- GitHub
- 47 stars
01 In brief
Summary
This paper challenges the Chain-of-Frames (CoF) hypothesis for reasoning in diffusion-based video models, proposing instead that reasoning primarily emerges along the diffusion denoising steps, termed Chain-of-Steps (CoS).
Through qualitative analysis and noise perturbation experiments, the authors show that models explore multiple candidate solutions in early denoising steps and progressively converge to a final answer.
They identify emergent behaviors: working memory, self-correction, and perception before action.
Layer-wise analysis of Diffusion Transformers reveals that middle layers (around 20-29) are critical for reasoning.
Based on these insights, they propose a Training-Free Ensemble (TFE) method that averages latents from three identical models with different seeds at mid-layers, improving performance on benchmarks like VBVR-Bench (0.685 to 0.716) and V-Reason Bench (8.94 to 12.12).
The study provides a systematic dissection of video reasoning mechanisms, positioning video models as a new substrate for intelligence.
02 From the paper
Abstract
Recent advances in video generation have revealed an unexpected phenomenon: diffusion-based video models exhibit non-trivial reasoning capabilities. Prior work attributes this to a Chain-of-Frames (CoF) mechanism, where reasoning is assumed to unfold sequentially across video frames. In this work, we challenge this assumption and uncover a fundamentally different mechanism. We show that reasoning in video models instead primarily emerges along the diffusion denoising steps. Through qualitative analysis and targeted probing experiments, we find that models explore multiple candidate solutions in early denoising steps and progressively converge to a final answer, a process we term Chain-of-Steps (CoS). Beyond this core mechanism, we identify several emergent reasoning behaviors critical to model performance: (1) working memory that supports tasks requiring consistent reference, such as object permanence; (2) self-correction and enhancement, allowing recovery from incorrect intermediate solutions; and (3) perception before action, where early steps establish semantic grounding and later steps perform structured manipulation. Moreover, analysis of Diffusion Transformer layers shows that middle layers conduct key reasoning procedures. Motivated by these insights, we present a simple Training-Free Ensemble (TFE) as a proof-of-concept, demonstrating how reasoning can be improved by ensembling latent trajectories from identical models with different random seeds. Overall, our work provides the first systematic dissection of the mechanisms underlying video reasoning, offering a foundation to guide future research in better exploiting the inherent reasoning dynamics of video models as a new substrate for intelligence.