VLA-Rolling: adaptive action reuse for more efficient VLA models
In brief
- VLA models generate actions via progressive denoising but only execute a short prefix.
- VLA-Rolling reuses discarded action tails as intermediate states to reduce denoising steps.
- Results: up to 1.48x faster inference and +15.08 points success rate across 16 tested setups.
Vision-language-action (VLA) models based on flow-matching generate action sequences by progressively refining Gaussian noise. According to the announcement, these models typically only keep the beginning of each sequence, discarding the rest as unused.
The researchers show that this discarded part, called the "action tail," retains usable denoising progress. VLA-Rolling, a lightweight inference framework, reuses these action tails in three steps: inheritance, per-action confidence estimation, and confidence-guided refinement. Tested on four models and four benchmarks, it reduces inference costs in all cases (up to 1.48x faster) and improves success rates in 14 of 16 setups (up to +15.08 percentage points).
What it changes for you
- Industry & ManufacturingFactories using VLA-based robots could cut compute costs and improve precision in robotic tasks.
This summary gives you the essentials. The full article, its figures and quotes are on Hugging Face Daily Papers.
Read the original articleAI-generated summary of the source, checked by rules. How we write · Report an error