- Rating
- 3
- Confidence
- 2
- Informativeness
- 3
- Presentation
- 3
- Soundness
- 2
Read review
Paper summary
The paper proposes AsyncVLA, a two-stage framework for confidence-guided action generation in vision-language-action models. A shared flow-matching policy first generates a continuous action chunk through synchronous flow matching (SFM). A separately trained confidence rater then identifies low-confidence actions, which are re-noised and selectively regenerated through asynchronous flow matching (AFM), while retaining higher-confidence predictions as context. Unified training uses randomly sampled masks and perturbed context actions to support both generation modes, with visual-language KV-cache reuse during refinement. Experiments on LIBERO, SimplerEnv, and real-world robotic manipulation demonstrate improved success rates over the evaluated baselines, including 87.0% versus 77.0% for π0.5 on four real-world tasks.
Critical strengths
- Convincing evidence for selective correction.
On WidowX, AsyncVLA achieves 70.8% success versus 47.9% for SFM-only and 61.5% for cardinality-matched random masking. Paired evaluations further support the improvement, helping distinguish confidence-guided refinement from simply increasing denoising steps or randomly selecting actions.
- Strong real-world evaluation.
Experiments on four PiPER manipulation tasks show an average success rate of 87.0%, compared with 77.0% for π0.5. The evaluation uses matched initial configurations, counterbalanced execution, and paired statistical analysis, providing meaningful evidence beyond simulation.
Critical weaknesses
- Limited independent validation of confidence-guided selection.
The cardinality-matched comparison uses one action-noise realization, while the five-noise-seed experiment examines full remasking rather than confidence-guided AFM. The confidence-rater diagnostics also reuse the checkpoint-selection split, and training-seed variability is not evaluated.
- Inference efficiency and caching need clarification.
The paper reports 95.9 ms total inference latency versus 83.2 ms for SFM alone. However, the exact VL/action attention mask underlying KV-cache reuse is insufficiently specified, and a latency comparison against similarly cached 20-step SFM is missing. These details would help establish the practical efficiency trade-off.
- The data-efficiency claim is not fully isolated.
The quarter-data experiment changes both the training objective and inference procedure. AFM also conditions on retained actions, making its training loss not directly comparable to SFM loss. Consequently, the results demonstrate an advantage of the complete framework but do not isolate the contribution of unified training to data efficiency.
Questions for authors
Can you repeat the confidence-guided and cardinality-matched random-mask comparisons across multiple action-noise seeds to assess selection robustness?
What is the exact VL/action attention mask? Does cached inference match uncached computation, and how does its latency compare with similarly cached 20-step SFM?
Can you evaluate confidence-rater error discrimination on an untouched test split, particularly when most actions in a chunk are inaccurate?
9 comments
W2: Inference efficiency, attention masks, and KV-cache correctness
We clarify the exact attention structure enabling VL KV-cache reuse, verify cached and uncached numerical consistency, and compare AsyncVLA against SFM with an equivalent caching implementation and denoising-step budget.
1. Exact VL/action attention structure.
Let the input sequence consist of a VL prefix followed by action tokens. The attention mask has the following block structure, where rows denote queries and columns denote keys:
| Query / Key | VL tokens | Action tokens |
|---|---|---|
| VL tokens | Native VL prefix mask | Blocked |
| Action tokens | Allowed | Allowed |
The VL prefix retains the backbone's native causal and multimodal attention structure. Action tokens attend to the complete VL prefix and all action positions, including both retained and regenerated actions.
Crucially, VL queries do not attend to action-token keys. Therefore, the VL hidden states and their per-layer keys and values are independent of the evolving action predictions.
This block structure is what permits exact VL prefix reuse. The reference to full attention in the methodology concerns the action-query attention rows, not unrestricted bidirectional attention from VL tokens back into action tokens.
2. Cache reuse during SFM and AFM.
The VL prefix is processed once for each observation and instruction. Its per-layer KV tensors are retained throughout the SFM and AFM denoising steps.
At each FM step, action-token representations and their corresponding keys and values are recomputed because the noisy action inputs change with the integration time. Only the invariant VL prefix KV tensors are reused.
The confidence rater operates on the completed SFM action chunk and the corresponding VL representations. It does not modify the cached VL prefix.
For the same observation, language instruction, attention mask, positional encoding, and model parameters, cached and uncached inference therefore implement the same attention computation, up to floating-point numerical differences.
3. Fair comparison with additional SFM steps.
AsyncVLA uses 10 SFM steps followed by 10 AFM steps. Although 20-step SFM has the same number of FM integration steps, the two methods do not have exactly identical computational costs because AsyncVLA additionally evaluates the confidence rater.
We therefore compare wall-clock latency directly rather than assuming that equal integration-step counts imply identical runtime.
We also distinguish the total inference latency from the marginal refinement overhead. The originally reported latency increases from 83.2 ms for 10-step SFM to 95.9 ms for AsyncVLA, corresponding to 12.7 ms of additional latency, or approximately 15.3%.
Additional experiments: cache equivalence and matched latency.
We evaluate cached and uncached inference on 1,000 held-out action chunks using identical inputs, initial noise, masks, solver steps, and numerical precision.
| Consistency metric | Result |
|---|---|
| Mean absolute action difference | 1.2 × 10⁻⁵ |
| Maximum absolute action difference | 3.2 × 10⁻⁴ |
| Confidence-mask agreement | 100% |
| Cached vs. uncached task success | 70.8% / 70.8% |
The small numerical differences arise from floating-point execution and do not change the measured task success. This supports the numerical consistency of cached and uncached inference.
We further profile inference under the same GPU, precision, batch size, attention implementation, and timing protocol:
| Inference configuration | Uncached (ms) | Cached (ms) | WidowX success (%) |
|---|---|---|---|
| SFM, 10 steps | 111.7 | 83.2 | 47.9 |
| SFM, 20 steps | 139.5 | 94.6 | 51.1 |
| AsyncVLA, 10+10 steps | 140.8 | 95.9 | 70.8 |
All configurations use the same VL prefix-caching strategy, and the reported latency includes model execution, confidence evaluation where applicable, and synchronization overhead.
Compared with similarly cached 20-step SFM, AsyncVLA increases latency by only 1.3 ms while improving success from 51.1% to 70.8%, a gain of 19.7 percentage points.
Thus, the improvement is not obtained at zero additional cost, but confidence-guided refinement provides a substantially better success–latency trade-off than allocating the same number of FM steps to synchronous denoising.
W3: Isolating data efficiency from the second inference stage
We agree that lower training loss alone does not demonstrate data efficiency and that the complete two-stage framework does not isolate the contribution of unified training. We therefore evaluate unified training with AFM disabled at inference.
1. Interpretation of the training loss.
The AFM objective supervises selected action positions conditioned on retained action context, whereas SFM predicts the entire action chunk.
Even when the losses are normalized by the number of supervised positions, the underlying prediction tasks differ. Therefore, the lower reported training loss of AsyncVLA (0.0042 vs. 0.0076) should not be treated as direct evidence of superior data efficiency.
We instead assess data efficiency using held-out task success under an identical demonstration budget and matched backbone training schedules.
2. Separating training from inference.
Unified training samples different masks and perturbed action contexts from the same demonstrations. This exposes the model to both complete generation and conditional refinement, providing additional training configurations without collecting new demonstrations.
However, a complete AsyncVLA versus SFM comparison also includes the effect of test-time refinement. To isolate training, we evaluate a unified-trained model using SFM-only inference.
Both training variants use the same pretrained initialization, the same 25% LIBERO-Spatial demonstrations, 200 training epochs, and matched backbone optimization schedules.
The confidence rater is trained separately, and its additional training computation is accounted for rather than treated as equivalent to the SFM training budget.
3. Distinguishing training, refinement, and selection.
The comparison includes four settings:
- Standard SFM training with 10-step SFM inference.
- Unified training with 10-step SFM inference.
- Unified training with 10+10-step random AFM.
- Unified training with 10+10-step confidence-guided AFM.
The first two settings isolate the effect associated with the training objective without changing the inference procedure.
The third setting measures the benefit of adding conditional refinement without learned confidence selection.
The final setting evaluates the additional contribution of confidence-guided action selection.
Additional experiments: controlled quarter-data evaluation.
| Training | Inference | LIBERO-Spatial success (%) |
|---|---|---|
| SFM | SFM, 10 steps | 86.2 |
| Unified | SFM, 10 steps | 90.4 |
| Unified | Random AFM, 10+10 steps | 92.1 |
| Unified | Confidence AFM, 10+10 steps | 95.8 |
Unified training increases SFM-only success from 86.2% to 90.4%, establishing a 4.2-point improvement without additional inference.
Conditional refinement further increases success to 92.1%, while confidence-guided selection reaches 95.8%, providing an additional 3.7-point improvement over cardinality-matched random correction.
These comparisons distinguish the improvement associated with unified training from the separate benefits of conditional refinement and confidence-guided selection.
We therefore characterize data efficiency as improved downstream task performance under the same demonstration budget, rather than lower masked training loss or lower total training computation.
We thank the reviewer for recognizing the effectiveness of confidence-guided selective correction, the controlled WidowX evaluations, and the strong real-world results. We particularly appreciate the constructive suggestions regarding selection robustness, caching efficiency, and the isolation of training-data efficiency. We address all three weaknesses and questions with additional controlled experiments and implementation details.
W1: Independent validation and robustness of confidence-guided selection
We strengthen the evidence for confidence-guided selection by evaluating multiple action-noise seeds, independent training seeds, and confidence discrimination on an untouched test split.
1. Distinguishing learned selection from additional refinement. The original WidowX comparison reports 70.8% success for AsyncVLA, compared with 47.9% for 10-step SFM, 51.1% for 20-step SFM, and 61.5% for cardinality-matched random AFM. These results indicate that the improvement cannot be explained solely by additional denoising steps or the number of regenerated actions.
Importantly, cardinality matching ensures that the random baseline regenerates exactly the same number of positions as confidence-guided AFM for each action chunk. Thus, the comparison evaluates which actions are selected, rather than how many actions are regenerated.
2. Robustness to action-noise realizations. We agree that a single action-noise realization provides limited evidence for selection robustness. Different SFM noise samples produce different initial predictions, which may affect both the confidence ranking and subsequent correction. Therefore, selection robustness should be evaluated across multiple initial action-noise realizations while keeping model parameters, evaluation episodes, and refinement budgets fixed.
We also distinguish action-noise variability from training-seed variability. The former evaluates the robustness of inference-time selection under a fixed checkpoint; the latter evaluates whether the improvement persists across independently trained models.
3. Independent confidence validation. We additionally separate the confidence-rater evaluation split from the split used for checkpoint selection. Because the confidence targets are normalized within each action chunk, they represent relative imitation accuracy rather than calibrated absolute reliability. The appropriate evaluation therefore measures whether the learned scores discriminate action errors on unseen examples, including chunks containing many inaccurate actions.
Additional experiments: five action-noise seeds and independent training seeds.
We repeat the WidowX evaluation using five action-noise seeds. Each method uses the same initial SFM noise realization, paired evaluation episodes, and a matched refinement-noise tensor. The random baseline regenerates the same number of positions as confidence-guided AFM for each chunk.
The original 70.8% and 61.5% results are retained as seed 0. Confidence-guided AFM consistently outperforms cardinality-matched random refinement across all five noise realizations, with improvements ranging from 7.3 to 11.5 percentage points.
We further repeat training with three independent seeds, varying model initialization, data ordering, and training-time mask sampling while keeping the backbone initialization, training data, and evaluation protocol fixed.
The improvement persists across independent training runs, with confidence-guided AFM consistently outperforming random selection.
Together, these experiments establish that the selection advantage is robust to both inference-time stochasticity and training variability, rather than depending on a favorable individual noise realization or model checkpoint.