A Mechanistic View on Video Generation as World Models: State and Dynamics

TL;DR

This paper proposes a new framework for video generation models that emphasizes state construction and dynamics modeling, aiming to develop more robust world models capable of physical reasoning and causal understanding.

Contribution

It introduces a novel taxonomy for classifying video models based on state and dynamics, bridging the gap between current architectures and classical world model theories.

Findings

Categorizes video models into implicit and explicit state construction paradigms.

Analyzes dynamics modeling through knowledge integration and architectural reformulation.

Recommends shifting evaluation from visual quality to functional benchmarks like causality.

Abstract

Large-scale video generation models have demonstrated emergent physical coherence, positioning them as potential world models. However, a gap remains between contemporary "stateless" video architectures and classic state-centric world model theories. This work bridges this gap by proposing a novel taxonomy centered on two pillars: State Construction and Dynamics Modeling. We categorize state construction into implicit paradigms (context management) and explicit paradigms (latent compression), while dynamics modeling is analyzed through knowledge integration and architectural reformulation. Furthermore, we advocate for a transition in evaluation from visual fidelity to functional benchmarks, testing physical persistence and causal reasoning. We conclude by identifying two critical frontiers: enhancing persistence via data-driven memory and compressed fidelity, and advancing causality through latent factor decoupling and reasoning-prior integration. By addressing these challenges, the field can evolve from generating visually plausible videos to building robust, general-purpose world simulators.