DriveCache: Action-Aware Caching Boosts Driving World Model Inference

A recent study published on arXiv (ID 2608.16354) presents DriveCache, a caching controller that operates without training and is sensitive to actions, aimed at enhancing the inference process of driving world models. These models are essential for the advancement of autonomous driving, facilitating simulation, planning assessment, and offline data creation by generating controllable future driving scenarios. However, the computational demands of diffusion-based driving generators arise from the need for multiple evaluations of extensive backbones during denoising steps, which hampers throughput. While current methods for diffusion acceleration address this issue, they fail to utilize driving-specific signals like ego speed and intended trajectories prior to generation. DriveCache optimizes feature reuse across scenes based on planned motion and employs dynamic programming to manage reuse during denoising within a calibrated response budget. A causal drift check updates features and adjusts the remaining schedule if generation strays from expectations. Experiments demonstrate that cache tolerance is influenced by ego translation and rotation, denoising progress, and the length of consecutive reuse, highlighting the approach's effectiveness.