Demystifing Video Reasoning

Recent advances in video generation models have significantly transformed the landscape of the movie, gaming, and entertainment industries. However, most research has primarily focused on their ability to produce high-fidelity, realistic, and visually appealing videos. Recent studies have revealed an unexpected phenomenon: diffusion-based video models exhibit non-trivial reasoning capabilities in spatiotemporally consistent visual environments. Prior work attributes this behavior to a Chain-of-Frames (CoF) mechanism, suggesting that reasoning unfolds sequentially across video frames. Despite this intriguing discovery, the underlying mechanisms of video reasoning remain largely unexplored. With the recent release of large-scale video reasoning datasets and open-source foundation models, we now have the opportunity to systematically investigate this capability. Leveraging these resources, we conduct the first comprehensive dissection of video reasoning and uncover a fundamentally different mechanism: reasoning in diffusion-based video models primarily emerges along the denoising process rather than across frames.

Our key discovery challenges the prevailing Chain-of-Frames (CoF) hypothesis, which assumes that video reasoning unfolds sequentially across frames. Instead, we find that reasoning does not primarily operate along the temporal dimension. Rather, it emerges along the diffusion denoising steps, progressing throughout generation. We term this mechanism Chain-of-Steps (CoS). This finding suggests a fundamentally different view of how diffusion-based video models reason. Due to bidirectional attention over the entire sequence, reasoning is performed across all frames simultaneously at each denoising step, with intermediate hypotheses progressively refined as the process unfolds. Qualitative analysis reveals intriguing dynamics. In early denoising steps, the model often entertains multiple possibilities (populating alternative trajectories or superimposing candidate outcomes) before gradually converging to a final solution in later steps. Moreover, noise perturbation analysis shows that disruptions at specific denoising steps significantly degrade performance, whereas frame-wise perturbations have a much weaker impact. Further information propagation analysis identifies that the conclusion primarily solidifies during the middle diffusion steps.

Furthermore, we uncover several surprising emergent behaviors in video reasoning models that are strikingly similar to those observed in early studies of Large Language Models (LLMs). First, these models exhibit a form of working memory, enabling persistent reference. Second, we observe that video models can self-correct and enhance throughout generation. Third, video models exhibit a perception before action strategy, where early steps establish semantic grounding and later steps perform structured manipulation.

We further conduct a fine-grained analysis of the Diffusion Transformer by examining token representations within a single diffusion step. This reveals the self-evolved, diverse, task-agnostic functional layers throughout the network. Within a diffusion step, early layers focus on dense perceptual understanding (e.g., separating foreground from background and identifying basic geometric structures), while a set of critical middle layers performs the bulk of the reasoning. The final layers then consolidate the latent representation to produce the video state for the next step.

Motivated by these insights, we present a simple training-free strategy as a proof-of-concept, demonstrating how reasoning can be improved by ensembling latent trajectories from identical models with different random seeds. This approach encourages the model to retain a richer set of candidate reasoning trajectories during generation. As a result, the model explores more diverse reasoning paths and is more likely to converge to the correct solution, illustrating a way to utilize our findings to design more effective video reasoning systems.