arXiv：cs.LG· Kai Ding, Yang He, Ruijie Quan, Yi Yang·· 4 小时前
WAM-Cache：为世界动作模型提供陈旧度受限的 KV 复用
WAM-Cache: Staleness-Bounded KV Reuse for Efficient World Action Models
AI 导读
WAM-Cache 是一个免训练框架，通过跨 chunk 保留 World Action Models 的逐层 KV 表示、仅重算稀疏刷新 token 集，将 video DiT 预填充 FLOPs 降低 32-42%。
正文
Abstract:World Action Models (WAMs) enable generalist robot manipulation by conditioning an action expert on representations from a pretrained video Diffusion Transformer (DiT). In closed-loop control, the video DiT runs at every chunk to encode the current observation into layerwise key-value (KV) pairs that the action expert queries. This prefill dominates the per-chunk computational cost, yet existing training-free accelerations leave it fully dense. We present WAM-Cache, a training-free framework that retains layerwise key-value representations across chunks and recomputes only a sparse refresh set of tokens. Crucially, we find that the intuitive heuristic of refreshing visually drifted tokens plateaus far below the dense baseline, even with an oracle predicting ground-truth KV drift. Downstream action accuracy is instead governed by where the action expert attends, not by what moved. WAM-Cache therefore selects the refresh set by uniting the action expert's cross-attention with visual latent surprise, complemented by a strict age bound that suppresses compounding error. On Fast-WAM, WAM-Cache cuts video DiT prefill FLOPs by 32-42% across RoboTwin 2.0, LIBERO, and real-world experiments, while staying within 0.7-1.8 percentage points of the dense policy in simulation and 2.5 points on a real robot.
| Comments: | 19 pages, 5 figures, 7 tables. Project page: this https URL |
| Subjects: | Robotics (cs.RO); Computer Vision and Pattern Recognition (cs.CV); Machine Learning (cs.LG) |
| Cite as: | arXiv:2610.11401 [cs.RO] |
| (or arXiv:2610.11401v1 [cs.RO] for this version) | |
| https://doi.org/10.48550/arXiv.2610.11401 arXiv-issued DOI via DataCite (pending registration) |
Submission history
From: Kai Ding [view email]
[v1]
Thu, 8 Oct 2026 07:33:59 UTC (543 KB)
来源：arXiv：cs.LG · arxiv.org