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
VLA-Rolling reuses discarded action tails via confidence-guided refinement to accelerate VLA inference without retraining.
How can we reduce the computational cost of flow-matching VLA models without sacrificing performance by reusing previously predicted action sequences?
Vision-language-action (VLA) models generate future actions in chunks, but they discard the unexecuted "tail" of each chunk whenever they replan. This discards valuable, partially denoised motion information that could have served as a head start for the next inference round. VLA-Rolling treats these tails as reusable intermediate states, using a lightweight confidence estimator to decide how much of each tail to keep and how many denoising steps to allocate for refinement. This allows the model to adaptively balance speed and accuracy based on how much the environment has changed since the last plan. Across 16 model-benchmark pairs, VLA-Rolling achieves up to a 1.48× speedup and improves success rates in 14 pairs by up to 15.08 percentage points.
Paper Primer
The core mechanism is a three-stage inference loop: inheritance, confidence estimation, and adaptive refinement. The estimator acts as a gatekeeper: it predicts the reliability of the inherited tail, then uses that confidence to inject noise (renoising) into unreliable segments and allocate a custom denoising budget for the remaining steps.
VLA-Rolling provides consistent Pareto improvements in both inference speed and task success.
Across 16 model-benchmark pairs (including $\pi\$0.5, SmolVLA, X-VLA, and MolmoAct2), the framework reduced inference cost by 5.00%–32.52% while maintaining or improving success rates. Up to 1.48× speedup and 15.08 percentage point improvement in success rate.
The method hinges on the observation that action tails are not just stale predictions, but partially denoised trajectories. By treating the tail as an intermediate state in the flow-matching process rather than a fresh start, the model preserves temporal consistency while correcting for environmental drift.
Why is a confidence estimator necessary instead of just reusing the tail directly?
Directly reusing the tail without evaluation can degrade performance, as stale actions from a previous state may be inaccurate in a new environment. The estimator prevents "blind" reuse by identifying when an action is unreliable, triggering renoising and additional refinement steps to correct the trajectory.
Does this approach require retraining the base VLA model?
No. VLA-Rolling is a lightweight inference-time plugin (approx. 30M parameters) that keeps the underlying VLA backbone frozen, making it compatible with existing pre-trained models.
Introduction and Motivation
Identifies wasted compute in VLA inference and proposes reusing discarded action tails.
Flow‑matching VLA models transform Gaussian noise into a future‑action chunk, but under the receding‑horizon paradigm they execute only a short prefix before replanning, discarding the unexecuted remainder—the action tail.
A VLA model takes visual observations and language instructions, then iteratively denoises Gaussian noise until it becomes a coherent sequence of robot actions.
The planner predicts a full action chunk, executes only its leading segment, then observes the environment and replans a new chunk, discarding the remainder.
Discarding the action tail wastes the denoising work already performed; the tail already contains partially refined actions that overlap with the next planning window. However, because the tail was generated from an earlier observation, its usefulness depends on how much the environment has changed.
**Figure 1.** VLA-Rolling reuses rather than discards action tails. Confidence controls both how strongly inherited actions are renoised and how many denoising steps refine the next chunk.
The inefficiency of discarding action tails in current VLA inference is the key obstacle this work addresses.
Feasibility of Action Tail Reuse
Shows that reusing the unused action tail cuts denoising steps.
Flow‑matching VLAs predict a full $H$‑step action chunk, but only the first $K$ steps are executed before replanning. The remaining $L=H\!-\!K$ steps—called the action tail—are normally discarded, wasting compute.
The action tail is the block of predicted actions that lies beyond the current replanning horizon; keeping it gives the robot a warm‑start for the next planning round.
Execute $a_1$ and $a_2$; the remaining tail is $[a_3,a_4,a_5]$.
Initialize the next planning round with the tail $[a_3,a_4,a_5]$ and add Gaussian noise.
Run $N=4$ denoising steps on this initialization; the error drops below the 10‑step Gaussian baseline.
Compare to a fresh random initialization that would need $N=10$ steps to reach the same error.
Reusing the tail provides a warm‑start that cuts the required denoising steps roughly in half for many samples.
How does reusing the action tail differ from simply discarding the unused actions?
Discarding throws away a partially computed plan, forcing the model to start from scratch. Reusing the tail keeps the already‑computed suffix, so the subsequent denoising only refines what is already close to the target, saving inference cycles.
**Figure 2** Action tails recover a 10-step Gaussian prediction in fewer steps. Tail initialization crosses the Gaussian baseline in four steps for $\pi_{0.5}$ and six for SmolVLA on both benchmarks. Subtle shaded bands mark these earliest shared operating points.
The VLA-Rolling Paradigm
VLA-Rolling reuses the previous action tail, guided by confidence, to adapt denoising.
Discarding the unexecuted “action tail” wastes compute because the model already generated those actions but never uses them during replanning.
Instead of starting each planning round from scratch, the model treats the previously predicted but unexecuted actions as a warm‑start prior and refines it only as much as needed.
Compute $L = H - K = 3$; the tail consists of the last three tokens $[(1,1), (0,0), (1,0)]$.
Current observation $o_t$ and previous observation $o_{t-2}$ differ by $\Delta o_t = [0.1, -0.05]$ (small change).
Confidence estimator maps $\Delta o_t$ to a high confidence score $c=0.85$.
Noise scale is set to $(1-c)=0.15$; Gaussian noise of magnitude $0.15$ is added to each tail token.
Adaptive denoising runs $N_{\text{adapt}} = \lceil (1-c)\times 4\rceil = 1$ refinement step, producing the final prefix sent to the robot.
When the environment changes little, the tail is reused almost unchanged, saving the cost of a full denoising pass.
Action‑Tail Inheritance: retrieve the tail $A_{\text{tail},t-K}$ from the previous round and align it with the current time step.
Action‑Tail Estimation: compute $\Delta o_t$, $\Delta q_t$ and feed them to the Action Confidence Estimator to obtain confidence $c$.
Confidence‑guided Refinement: add Gaussian noise scaled by $(1-c)$ to the tail and run $N_{\text{adapt}}$ denoising steps, then execute the resulting prefix.
**Figure 4. Overview of VLA-Rolling.** VLA-Rolling reuses the unexecuted action tail from the previous inference round by estimating its confidence, which jointly determines tail preservation and the adaptive denoising budget while keeping the original VLA frozen. (a) VLA-Rolling inference. The Action Confidence Estimator predicts a confidence which controls its interpolation with Gaussian noise. The resulting initialization is refined using $N_{adapt}$ denoising steps. The executable prefix is sent to the robot, while the remaining tail rolls into the next round. (b) VLA-Rolling training. Confidence labels are constructed from the quality of inherited actions and normalized with EMA-based quantile statistics.
Action Confidence Estimation
Estimate tail reliability to adapt denoising budget per action.
The rolling VLA paradigm reuses the previously generated action tail, but without a measure of its reliability the system cannot know how much to trust it. We therefore introduce a lightweight confidence estimator that predicts how well the inherited tail will survive under the current environment.
It predicts a per‑position confidence score for the inherited action tail, using recent observation and query changes, so the system can inject just enough noise and allocate the right number of denoising steps.
Compute the raw error $E_t$ for each position as the squared difference to the target action (e.g., $E_{t,1}=0.04$, $E_{t,2}=0.09$, $E_{t,3}=0.01$, $E_{t,4}=0.16$).
Apply the log transform: $r_{t,i}=\log(E_{t,i})$ giving $r_{t}=[-3.22,-2.41,-4.61,-1.83]$.
Assume the batch 10 % and 90 % quantiles of $r$ are $q_{0.1}=-4.0$ and $q_{0.9}=-1.5$. Normalize each $r_{t,i}$: $\frac{r_{t,i}-q_{0.1}}{q_{0.9}-q_{0.1}}$ yields $[0.22,0.59,0.00,0.83]$.
Clip to $[0,1]$ (no change) and compute $c^{*}_{t,i}=1-$ that value, resulting in confidences $[0.78,0.41,1.00,0.17]$.
The estimator $f_{\phi}$ learns to predict these targets from the input deltas; after training it might output $c_{\phi}=[0.80,0.38,0.95,0.20]$.
Low confidence at position 4 signals a large observation change, so the system will inject stronger renoising there, while high confidence at position 3 lets the original action survive with little modification.
**Figure 3** Action tails require sample-dependent refinement. Each bar shows the fraction of samples with a given optimal denoising step $N^*$. Green denotes $N^*$ at or below the model-specific median (8 for $\pi_{0.5}$, 9 for SmolVLA), while red denotes $N^*$ above it. Dashed lines mark the median boundaries; brackets report the fraction of samples in each region. Means and standard deviations appear inside the panels.
How does this confidence estimator differ from simply regressing the raw error $E_t$?
Regressing $E_t$ would produce values that span many orders of magnitude, making it hard for a single network head to learn a stable mapping. By log‑transforming $E_t$ to $r_t$ and then quantile‑clipping to $c^{*}_t$, the estimator works on a bounded, roughly uniform scale and can focus on the relative ordering of errors rather than their absolute magnitude. Moreover, the asymmetric weighting $w_{+}>w_{-}$ penalizes over‑confidence more strongly, which a plain regression loss would not enforce.
Performance and Efficiency Results
VLA‑Rolling raises success rates and cuts inference cost across benchmarks.
VLA‑Rolling raises success rate by an average of +2.0 pp across the four benchmarks.
Table 1 shows per‑benchmark SR gains ranging from +0.20 pp to +5.93 pp, which average to +2.0 pp.
**Table.** Performance comparison of various VLA models with and without VLA-Rolling across four benchmarks (LIBERO, MetaWorld, LIBERO-Plus, and RoboTwin), reporting Success Rate (SR) and inference Cost.
Experimental Setup
VLA‑Rolling delivers up to 1.48× throughput boost with stable success rates.
VLA‑Rolling boosts action chunk throughput by up to 1.48× while keeping task success rates stable.
Table 1 shows the throughput increase and comparable success rates across LIBERO, MetaWorld, and LIBERO‑Plus.
We evaluated on LIBERO, MetaWorld, LIBERO‑Plus, and RoboTwin 2.0, running 100 rollouts per task to measure success rate (SR) and inference latency.
The table presents data on the Execution Horizon $K$ and its impact on various metrics including $\Delta o_t$, $\Delta q_t$, $c_{min}$, $\rho(\Delta o_t, c_{min})$, and $\rho(\Delta q_t, c_{min})$ for values of $K$ ranging from 10 to 40.
**Figure 5.** Left: Average denoising steps across task difficulty levels on MetaWorld. Successful trajectories consistently require more denoising steps than failed trajectories. Right: Distribution of episodes across different difficulty levels.
Observation Shift Analysis
Ablations quantify how each component affects success rate and cost.
The ablation study isolates the contribution of each design choice by removing it and measuring the impact on success rate and computational cost.
**Table 3.** Observation shift and confidence statistics under different execution horizon $K$. Correlation values $\rho$ are Pearson correlations computed within each $K$ group. $c_{min}$ is the average minimum of predicted confidence.
Removing the inherited action tail reduces success rate by 4.89 %.
Table 2 reports a 4.89 % drop when the action tail is used without renoising.
Reducing the number of denoising steps cuts success rate by 4.29 %.
Table 2 shows a 4.29 % decrease when the denoising‑step count is lowered.
Analysis of Denoising Strategies
Key ablations show how confidence estimation and VLA‑Rolling affect robustness and efficiency.
This section isolates the contribution of each component by removing it and measuring the impact on confidence, denoising budget, and overall efficiency.
Increasing the execution horizon $K$ beyond the default $K\!=\!10$ causes the confidence estimator to lose reliability, confirming that the estimator is only robust to modest environmental shifts.
Table 3 (LIBERO) shows a clear negative trend: as $K$ grows, predicted confidence drops sharply once the environment deviates beyond the training distribution.
Successful trajectories consistently require more denoising steps than failed ones, with a maximum gap of 2.33 steps across difficulty levels.
MetaWorld experiments report denoising‑step counts: Easy (350 vs 70), Medium (87 vs 78), Hard (48 vs 42), Very hard (24 vs 51).
Adding the VLA‑Rolling plugin improves success rate by +5 % and cuts inference cost by −50.78 ms, confirming the practical benefit of reusing the action tail.
Real‑world tests on the AgileX PiPER arm show $\pi$0.5 achieving 70 % SR at 227.28 ms, while +VLA‑Rolling reaches 75 % SR at 176.50 ms.
**Figure 6.** Real-world experiment of context mutations revealed a significant decrease in confidence at the mutation site, leading to an increase in the number of denoising steps.
Extended Benchmark Results
Calibrated $\alpha$=1.5 gives the best success rate while keeping inference fast.
$\alpha$ = 1.5 yields the highest success rate among all calibration settings.
Table 5 reports 98.50 % SR for $\alpha$ = 1.5 on the LIBERO benchmark.
**Table 6.** Offline trajectory-level action error under different denoising step convert strategies.
**Table 7.** Performance and inference cost of VLA-Rolling models on LIBERO. Success rate statistics are computed across 40 tasks, 10 episodes per task
Performance Metrics Summary
VLA‑Rolling achieves peak performance at 30 M parameters with minimal inference overhead.
VLA‑Rolling with 30 M parameters attains the highest success rate (98.0 %) while keeping inference cost low.
Table 7 shows 30 M models achieve 98.0 % ± 5.10 % success with 1.579 ± 0.335 ms inference cost, outperforming 10 M and 100 M variants.
**Table 8.** N* Locally Optimal Test.
**Table 9.** Observation and proprioceptive-state similarity across a 10-step execution horizon. We compare the source context at time $t$ with the updated context at time $t + 10$. **Image Cosine** denotes the cosine similarity between the two visual representations; **Image (1 - MAE)** denotes one minus the mean absolute error between the normalized visual representations; **State Cosine** measures the cosine similarity between the corresponding proprioceptive states.
The ablation and similarity analyses together support the design choice of reusing action tails rather than discarding them, enabling VLA‑Rolling to achieve strong performance with modest compute.
Real-World Robotic Experiments
Real‑world tests show VLA‑Rolling improves success and cuts latency.
VLA‑Rolling raises task success from 70 % to 75 % on the real‑world placement experiment.
Table 4 shows 14/20 successes for the baseline $\pi_{0.5}$ model versus 15/20 for the VLA‑Rolling configuration.
The experiment uses a PiPER 6‑DoF arm to grasp a potato and place it into a plate whose color is perturbed mid‑execution, forcing the controller to continue with the same linguistic command “place the potato into the plate”. Execution horizons $K_{\text{min}}$ and $K_{\text{max}}$ are derived from the estimated latency and control frequency, bounding the number of timesteps the planner can safely reuse.
**Table 4.** Model performance on real-world experiments.
Implementation Details
Implementation specifics of the VLA‑Rolling components and training pipeline.
The predictor estimates how trustworthy the inherited action tail is, enabling VLA‑Rolling to decide whether to keep or discard it during re‑planning.
How does this confidence predictor differ from a typical classifier that outputs a probability?
A standard classifier is trained to separate predefined classes, whereas the predictor is trained to regress a scalar indicating the quality of a future action sequence; its loss is asymmetric L1 rather than cross‑entropy, and its output directly modulates the denoising budget instead of selecting a discrete label.
Why not simply omit the Gaussian padding instead of masking it?
Omitting the padding would change the input length, breaking the transformer’s positional embeddings; padding preserves the expected sequence length while the mask guarantees that the noise does not affect the confidence computation.
Instead of restarting denoising from a fresh Gaussian state, the inherited action tail is treated as an intermediate point on the flow‑matching trajectory and refined for the remaining steps.
How does this continuation differ from simply re‑initialising denoising with a fresh Gaussian sample?
Re‑initialising would set $\tau=0$ and discard all information already present in the tail, effectively wasting the computation that produced the tail; the adaptive step starts at $\tau_{0}>0$, preserving the partially refined state and only spending budget on the remaining trajectory.
Training updates the confidence predictor and the EMA risk statistics while applying an asymmetric loss that penalises under‑confidence more than over‑confidence.
During evaluation the same VLA model and the VLA‑Rolling plugin are applied unchanged across multiple robotic benchmarks to assess robustness.
Inference rule that uses confidence to decide noise budget and whether to preserve the action tail.