VLA·Rolling Code
RESEARCH PROJECT / VISION · LANGUAGE · ACTION

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 Jun Zhang Jie Zhang Wenqi Zhang Xuhong Zhang Jianwei Yin

OmniAI Group of ZJU ACES Lab Zhejiang University

First author: Sicheng He · chen123ww@outlook.com

Why restart from noise when useful actions are already there? VLA-Rolling carries forward the unexecuted action tail and adapts its refinement to the changing scene—without modifying the frozen VLA backbone.

Paper (PDF) Code Model arXiv
1.48×up to inference
throughput speedup
+15.08%up to success-rate
improvement
16/16model–benchmark pairs
with lower latency
14/16pairs with improved
task success
00 / VIDEO DEMO

See it in action.

A video introduction to VLA-Rolling: reusing action tails and adapting refinement to the changing scene.

VLA-Rolling demo. A video overview of the method. Open video ↗
01 / OVERVIEW

VLA-Rolling at a glance.

When the target moves, useful actions can still carry forward. This illustrated example shows how VLA-Rolling evaluates the inherited action tail and adapts its refinement to the updated scene.

Illustrated VLA-Rolling overview: when a plate moves, current observations and cached action tails inform per-action confidence. Useful actions receive less noise, uncertain actions receive more refinement, and the frozen action expert generates an updated plan whose remaining tail is reused next cycle.
Overview. A moving-plate example illustrates action-tail reuse, confidence estimation, and adaptive refinement. View full-size illustration ↗ Original method diagram (PDF) ↗
02 / INTRODUCTION

Make every action count.

Traditional receding-horizon control throws away the unexecuted tail of each action chunk. VLA-Rolling turns that discarded computation into the starting point for the next decision.

Figure 1: Standard receding-horizon VLA discards the action tail, whereas VLA-Rolling passes it to the next round for confidence-guided refinement.
Figure 1. Instead of restarting from Gaussian noise each round, VLA-Rolling inherits and selectively refines previously predicted actions. From the manuscript, p. 2. View original PDF ↗
CENTRAL INSIGHT

An action tail is not a finished plan and not fresh noise. It is a reusable intermediate state whose value depends on how the environment has changed.

03 / THE METHOD

Reuse. Evaluate. Refine.

A lightweight, three-stage inference framework learns when to preserve an inherited action and when to correct it. The original VLA stays frozen.

Figure 4: The VLA-Rolling inference pipeline and offline training of the action confidence estimator.
Figure 4. VLA-Rolling inference (left) and confidence estimator training (right). From the manuscript, p. 5. View original PDF ↗
01INHERIT

Action-tail inheritance

Execute a short prefix and align the unexecuted actions with the next planning window. Cache the prior context and the motion already predicted by the VLA.

Atail = AK:H · L = H − K
02ESTIMATE

Per-action confidence

A ~30M-parameter action-centric Transformer scores the reuse reliability of each tail action using visual, proprioceptive, and temporal-change information.

ci = fφ(Atail, o, Δo, q, Δq)i
03REFINE

Confidence-guided refinement

High-confidence actions receive less noise; uncertain actions are regenerated more aggressively. The minimum confidence sets the denoising budget for the next chunk.

Nadapt ∝ (1 − min ci)
✳

No VLA retraining. The original VLA backbone remains frozen; only the lightweight confidence estimator is trained offline on temporally adjacent demonstrations.

04 / WHY IT WORKS

The right amount of refinement.

Tail reuse reduces the amount of denoising needed, but no single denoising budget works for every sample. Confidence makes reuse adaptive rather than unconditional.

OBSERVATION A · FIGURE 2

Fewer steps can be enough.

Tail initialization reaches the 10-step Gaussian baseline with 4 steps for π₀.₅ and 6 for SmolVLA, across the tested diagnostic benchmarks.

Figure 2: Error ratios for tail initialization with 2, 4, 6 and 8 denoising steps against a 10-step Gaussian reference.
Action-tail initialization vs. Gaussian restart.
OBSERVATION B · FIGURE 3

One budget does not fit all.

Sample-wise optimal denoising steps span 1–10. The distribution varies considerably across samples and model backbones.

Figure 3: Distributions of optimal denoising steps for pi zero point five and SmolVLA.
Optimal denoising steps are sample-dependent.
05 / EXPERIMENTS

Efficiency, no need to trade-off.

Evaluated on four VLA backbones and four benchmarks. VLA-Rolling lowers chunk inference latency in all 16 model–benchmark combinations, while success rate rises in 14.

ACROSS ALL 16 PAIRS

Less compute.
Better control.

The largest latency reduction is 32.52% (MolmoAct2 on LIBERO); the largest success gain is +15.08% (SmolVLA on LIBERO-Plus).

−32.52%Largest chunk-latency reduction
+15.08%Largest success-rate increase

Benchmark comparison

Choose a benchmark to inspect all four VLA backbones.

Table 1 · Main paper
BackboneSuccess rate (%)baseline → oursΔ SuccessLatency (ms / chunk)baseline → oursΔ Latency
π₀.₅96.38 → 97.83+1.45%168.09 → 129.60−22.90%
X-VLA96.03 → 96.65+0.62%135.24 → 112.79−16.60%
SmolVLA73.85 → 79.78+5.93%127.80 → 89.57−29.91%
MolmoAct295.50 → 95.70+0.20%291.34 → 196.61−32.52%

LIBERO · Results reproduced from Table 1 of the manuscript. Success-rate changes are absolute differences (ours − baseline), shown with %; latency changes are relative percentages.

Latency measures model inference time per action chunk, not robot control frequency or total task duration. Two success-rate results decline slightly (X-VLA on LIBERO-Plus and RoboTwin), while their latency still improves.

COMPONENT ABLATION / TABLE 2

Each part has a purpose.

Which components turn action reuse into better control? With π₀.₅ fixed, Table 2 isolates tail inheritance, renoising, and adaptive denoising across three benchmarks.

What each component contributes

Baseline measurements and component ablations reproduced from Table 2.

Table 2 · Main paper
A Action-tail inheritanceB Confidence-guided renoisingC Adaptive-step denoisingBfixed Fixed 50% renoising
MethodLIBEROMetaWorldLIBERO-Plus
SR (%) ↑Cost (ms) ↓SR (%) ↑Cost (ms) ↓SR (%) ↑Cost (ms) ↓
π₀.₅ baseline96.38166.0467.22217.2884.73161.92
Variant vs. baselineΔ SR (%) ↑Δ Cost (%) ↓Δ SR (%) ↑Δ Cost (%) ↓Δ SR (%) ↑Δ Cost (%) ↓
+ A−2.13%+0.02%+0.63%+0.01%−4.89%+0.02%
+ A + Bfixed+0.20%+0.02%+0.61%+0.02%−0.16%+0.02%
+ A + B+0.37%+0.55%+0.40%+0.42%−0.22%+0.84%
+ A + C−1.48%−22.19%−0.62%−15.93%−4.29%−11.91%
Full model · A + B + C+1.45%−23.03%+0.66%−16.74%+0.69%−12.07%

SR = success rate; Cost = model inference time per action chunk. Δ SR is the absolute success-rate difference (variant − baseline), shown with %; Δ Cost is a relative percentage change. Table 2 uses its own baseline inference-cost measurements, which differ slightly from Table 1. Source: manuscript, p. 8.

Reuse needs correction.

Tail inheritance alone (+A) improves MetaWorld but reduces success on LIBERO and LIBERO-Plus. On LIBERO-Plus, the change is −4.89%; adding renoising brings it close to baseline (−0.16% with fixed noise; −0.22% with confidence-guided noise). Reusing a tail therefore needs a mechanism to correct actions that no longer fit the current observation.

The full combination delivers both gains.

Adaptive steps without renoising (+A+C) cut inference cost by 11.91–22.19%, but success remains below baseline on all three benchmarks. The full model is the only tested variant that improves success and lowers cost on all three: +1.45%, +0.66%, and +0.69% in success, alongside cost reductions of 23.03%, 16.74%, and 12.07%.

CONFIDENCE ANALYSIS / FIGURE 5

Different outcomes, different budgets.

On MetaWorld, successful and failed trajectories receive different denoising budgets within every task-difficulty group. Figure 5 compares their average steps and episode counts.

Figure 5 bar chart: Successful versus failed episodes average 7.23 versus 5.80 denoising steps on easy tasks, 8.34 versus 6.01 on medium, 6.20 versus 4.35 on hard, and 7.03 versus 5.96 on very hard.
Figure 5. Average denoising steps by task difficulty. From the manuscript, p. 9. View in paper ↗

A consistent gap within each group.

Successful episodes average 1.43, 2.33, 1.85, and 1.07 more steps than failed episodes on easy, medium, hard, and very hard tasks. The largest gap is on medium tasks: 8.34 versus 6.01. These averages show that confidence-guided budget allocation distinguishes trajectories with different outcomes.

More steps are not a guarantee.

The average budget does not rise monotonically with task difficulty: successful medium tasks use more steps than successful hard or very hard tasks. Episode counts also differ across groups (from 350 successes on easy tasks to 24 on very hard tasks). Figure 5 shows an association with outcomes; it does not establish that simply adding denoising steps causes success.

06 / REAL-WORLD VALIDATION

Ready when the world changes.

An AgileX PiPER 6-DoF robot is asked to put a potato into a plate. When the target plate is switched during execution, confidence reacts and denoising increases.

Real-world demo. Robot execution with adaptive denoising steps shown in the upper-left inset. Open video ↗
07 / CITE THIS WORK

Citation & resources.

Read the October 2026 manuscript from the OmniAI Group of ZJU ACES Lab, Zhejiang University, or use the BibTeX entry to cite this work.

Read the paper on arXiv Read the manuscript (PDF) View the code repository View the model on Hugging Face
BIBTEX ARXIV:2610.25043
@misc{he2026vlarollingfewerdenoisingsteps,
      title={VLA-Rolling: Fewer Denoising Steps, Better Control via Adaptive Action Reuse},
      author={Sicheng He and Yi Pan and Qiuting Yu and Jiaming Huang and Qi Lu and Shuyan Wu and Hongyan Feng and Xuanyu Liu and Jun Zhang and Jie Zhang and Wenqi Zhang and Xuhong Zhang and Jianwei Yin},
      year={2026},
      eprint={2610.25043},
      archivePrefix={arXiv},
      primaryClass={cs.AI},
      url={https://arxiv.org/abs/2610.25043},
}

BibTeX citation for arXiv:2610.25043.