C$^3$ache: Accelerating World Action Models with Cross Inference Chunk Cache
Training-free speedup for world-action models: reuses denoising residuals across consecutive inference chunks at the same step, since they correlate during smooth motion. Up to 2.5x with a Fast-WAM backbone.无需训练的世界动作模型加速方法：平滑运动时相邻推理块在同一去噪步的残差高度相关，因此跨块缓存复用。以 Fast-WAM 为骨干最高提速 2.5 倍。
BibTeX
@misc{zhao2026c,
title = {{C$^3$ache: Accelerating World Action Models with Cross Inference Chunk Cache}},
author = {Weisen Zhao and Lam Nguyen and Zhicong Lu and Yuzhang Shang},
year = {2026},
eprint = {2606.08962},
archivePrefix = {arXiv},
primaryClass = {cs.LG},
url = {https://arxiv.org/abs/2606.08962},
}