Researchers identify the "seriality gap" in video diffusion models, where performance degrades as causal chains lengthen during multi-ball hard-sphere dynamics simulations. Controlled experiments show that this degradation is caused by dependent-event structures rather than video length, as it disappears in single-ball controls without interactions.
- Standard bidirectional video diffusion fails to scale with longer causal chains despite increased denoising steps.
- Methods increasing effective serial computation, such as autoregressive generation and architectural depth, disproportionately improve performance.
- The study proves that deterministic video prediction denoising steps do not add serial computation beyond the backbone.
This finding indicates a structural obstacle for using video diffusion models on tasks requiring serial reasoning and simulation.