Title: WorldCache: Accelerating World Models for Free via Heterogeneous Token Caching

Abstract:

Although diffusion-based world models demonstrate significant promise for comprehensive world simulation, the iterative denoising process involved remains prohibitively expensive for interactive applications and long-term rollouts. Although feature caching offers a method to speed up inference without requiring additional training, we identify that policies optimized for single-modal diffusion do not transfer effectively to world models. This limitation stems from two specific challenges inherent to world models: token heterogeneity, arising from multi-modal coupling and spatial variations, and non-uniform temporal dynamics, where a limited subset of difficult tokens causes error accumulation, rendering uniform skipping either unstable or excessively cautious.

To address these issues, we present WorldCache, a caching framework specifically designed for diffusion world models. Our approach introduces Curvature-guided Heterogeneous Token Prediction, which employs a physics-based curvature metric to assess token predictability. For chaotic tokens exhibiting sudden direction shifts, it utilizes a Hermite-guided damped predictor. Additionally, we developed Chaotic-prioritized Adaptive Skipping, a mechanism that tracks a dimensionless, curvature-normalized drift signal. This system triggers recomputation only when critical bottleneck tokens start to drift.

Evaluated on diffusion world models, WorldCache achieves end-to-end speedups of up to 3.7$\times$ while preserving 98\% of the rollout quality. These results highlight the substantial benefits and viability of WorldCache, particularly in environments with limited resources. The source code is available at https://github.com/FofGofx/WorldCache.

Source: arXiv Generated at: 2026-06-02 00:00:00 UTC