Researchers led by Abhinav Gupta have compared two ways to speed up a flow-matching robot policy, reusing computation from earlier planning calls or simply taking fewer denoising steps, according to a paper on arXiv. They tested Fast-WAM, a world action model with 6 billion parameters, on the LIBERO-Goal benchmark.
Two uncached steps matched the success of ten within two percentage points on all 500 starting states and cut a planning call on an RTX A6000 graphics card from 202 to 92 milliseconds. Any cache that keeps two steps fresh can't beat that, and caching the first of two steps would save at most 11 to 14% of a call, the authors calculate.
Cached policies were also fragile: with the C3ache method's eight cached steps, the policy succeeded on 63% of states against 95% without caching. The paper received an honourable mention at a CoRL 2026 workshop.



