Megadose The AI curve, compressed.

Cache or Denoise Less? Residual Reuse in a Flow-Matching World Action Model

· ArXiv · AI/CL/LG ·
Two denoising steps preserved ten-step success while cutting Fast-WAM’s replanning time from 202 to 92 milliseconds.

On LIBERO-Goal’s 500 initial states, two uncached steps were non-inferior within two percentage points, tested with one checkpoint and noise seed. The timing result used an RTX A6000 and PyTorch’s default compile mode. Reusing cached computation proved fragile: under Fast-WAM’s current time shift, C³ache’s eight cached steps covered 80% of the flow path, with success falling to 63% against 95% uncached. The authors call for fewer-step baselines and testing caches under the sampler and noise settings used in deployment.

ArXiv · AI/CL/LG's note

score 5

Categories: Research