The Seriality Gap in Video Diffusion Models

2026-07-14

AI Summary

This work addresses the significant performance degradation of video diffusion models on tasks involving long causal chains—such as multi-ball collision sequences—and identifies a fundamental limitation in their sequential reasoning capability. The study introduces the concept of the “sequentiality gap,” revealing that the denoising process in standard diffusion models lacks a scalable sequential computational structure. Through a combination of bidirectional diffusion architectures, autoregressive and block-wise generation strategies, deep intervention analyses, controlled experiments, and theoretical reasoning, the authors demonstrate that enhancing effective sequential computation substantially improves long-range causal reasoning. Empirical results show that while conventional models suffer sharp performance drops as causal chain length increases, methods incorporating explicit sequential mechanisms yield disproportionately large gains, thereby validating the existence and critical role of the sequentiality gap.

Abstract

When one ball strikes another, then another, video models should predict the consequences of each bounce. In controlled experiments on multi-ball hard-sphere dynamics, we find that the performance of standard bidirectional video diffusion degrades as the causal chain lengthens, even when provided more denoising steps. In a length-matched single-ball control, where ball-ball interactions are absent, the degradation largely disappears, isolating dependent-event structure rather than video length as the cause. Across intervention studies, methods that increase effective serial computation improve performance disproportionately, including autoregressive/blockwise generation and architectural depth. We identify this pattern as the seriality gap: a mismatch between tasks requiring growing serial computation and video diffusion models whose denoising loop does not provide scalable serial compute. We then prove that, for deterministic video prediction, denoising steps do not add serial computation beyond the backbone, indicating a structural obstacle for video diffusion on serial reasoning and simulation tasks.

Problem

• seriality gap
• video diffusion models
• causal chain
• serial computation
• dependent-event structure

Innovation

• seriality gap
• video diffusion models
• serial computation
• causal chain
• denoising loop