VLA-Rolling: Fewer Denoising Steps, Better Control via Adaptive Action Reuse
Sicheng He, Yi Pan, Qiuting Yu, Jiaming Huang, Qi Lu, Shuyan Wu, Hongyan Feng, Xuanyu Liu and 5 others
Zhejiang University
Published October 18, 2026 · on the feed October 21, 2026 · ▲ 64 · arXiv 2610.25043
The abstract
Flow-matching vision-language-action (VLA) models generate an action chunk by iteratively denoising Gaussian noise, but typically execute only a short prefix before replanning with updated observations. Under this receding-horizon paradigm, the unexecuted remainder---the action tail---is discarded, and the next chunk is regenerated from fresh noise. We show that this discarded tail retains transferable denoising progress: action-tail initialization can match or outperform a native 10-step Gaussian baseline using fewer denoising steps. However, the optimal denoising budget varies substantially across samples, indicating that tail reuse should be adapted to the updated environment. Based on this observation, we introduce VLA-Rolling, a lightweight inference framework that reuses action tails through three stages: action-tail inheritance, per-action confidence estimation, and confidence-guided refinement through renoising and denoising budget allocation, while the underlying VLA remains frozen. Across π_{0.5}, SmolVLA, X-VLA, and MolmoAct2 on LIBERO, MetaWorld, LIBERO-Plus, and RoboTwin 2.0, VLA-Rolling reduces action-chunk inference cost in all 16 model--benchmark pairs, achieving up to a 1.48times speedup, while improving success rate in 14 pairs by up to 15.08 percentage points. These results show that action tails are not merely discarded predictions, but reusable intermediate states that enable more efficient and accurate VLA inference. These results demonstrate that the action tail can support the generation of better actions with fewer denoising steps. Our project is available at https://chenbao123ww.github.io/VLA-Rolling/.
The Brief
2:32A short podcast summary of the paper. Written and voiced by AI.
Transcript
VLA-Rolling. Fewer Denoising Steps, Better Control via Adaptive Action Reuse. By Sicheng He and twelve others, from Zhejiang University. Published on the eighteenth of October. Its code repository is worked out: about forty source files, with a readme.
Flow-matching vision-language-action models, or VLA models, produce an action chunk by iteratively denoising Gaussian noise. They typically execute only a short prefix of that chunk, and then replan with updated observations. This is the receding-horizon paradigm. Under it, the part of the chunk that never got executed, which the authors call the action tail, is simply discarded. The next chunk is then regenerated from fresh noise, as though the previous one had never been computed. Every replan starts the denoising from the very beginning.
The authors show that the discarded tail retains denoising progress that transfers to the next chunk. Initializing from the action tail can match, or outperform, a native ten-step Gaussian baseline while using fewer denoising steps. In other words, the old tail is a better starting point than noise.
There is a complication. The best denoising budget varies substantially from one sample to the next. That variation indicates that tail reuse should not be fixed. It should be adapted to the updated environment that each new observation reveals, sample by sample, rather than applied with one budget for all.
VLA-Rolling is a lightweight inference framework built on those two observations. It reuses action tails in three stages. The first is action-tail inheritance, where the leftover tail is carried forward. The second is per-action confidence estimation. The third is confidence-guided refinement, which works through renoising and through how the denoising budget is allocated. Throughout, the underlying VLA remains frozen. Nothing about the model is retrained, and only the way inference is run changes.
The evaluation spans four models, pi zero point five, SmolVLA, X-VLA and MolmoAct2, on four benchmarks, LIBERO, MetaWorld, LIBERO-Plus and RoboTwin 2.0. VLA-Rolling reduces action-chunk inference cost in all sixteen model-benchmark pairs, reaching a speedup of up to one point four eight times. And in fourteen of those pairs, success rate improves as well, by as much as fifteen point zero eight percentage points.
Cheaper and better at once is the unusual combination here, and it comes from recasting what a tail is. Instead of a leftover prediction to be discarded, it becomes a reusable intermediate state, a head start that lets the next chunk arrive at a better action in fewer denoising steps.
Paper Quest · 7 chapters · about 41 min
Play this paper
Learn it by doing: make the calls the authors made, guess the result before you see it, and defend it in a viva graded on your own words. Chapter 1 is right here.
Chapter 1 🧩 · The problem
Every replan starts from scratch. Here is the work being thrown away.
Your robot policy turns noise into a chunk of future actions by denoising it step by step. To react to the world, the robot runs only the first few actions, then looks again and replans.
The road ahead
- 1 · The problem ~7 min A robot policy denoises a whole chunk of future actions, runs only the first few, then throws the rest away and starts again from noise.
- 2 · What came before ~3 min
- 3 · The key idea ~9 min Start each new chunk from the old chunk's unused actions, and trust each one only as far as a small confidence estimator says.
- 4 · Choose your own research ~3 min
- 5 · Build the method ~8 min Run a prefix, carry the leftover tail forward, score each inherited action's confidence, then renoise and budget denoising steps by that score.
- 6 · The experiment ~6 min You’ll guess: Across four VLA models and four benchmarks, VLA-Rolling cut action-chunk inference cost. What was its largest speedup?
- 7 · The viva ~5 min Three questions from the examiners, answered in your own words. Win it for the paper’s trophy.
Free, with Google or an email address. Your stars and progress are kept.