- Rating
- 2
- Confidence
- 2
- Informativeness
- 2
- Presentation
- 2
- Soundness
- 1
Read review
Paper summary
The paper targets the cost of multi-step generation in the action head of flow-based VLAs and proposes ActionCache, a training-free method. It uses VLM output embeddings to retrieve actions cached from successful episodes, then either reuses them directly or runs a small amount of refinement. When similarity is too low, it falls back to full-step generation. No model weights are changed.
Actions can be reused across timesteps and across episodes. The paper evaluates SR and latency on VLABench, LIBERO, and real-robot tasks, and also looks at what affects action reuse. After cache warmup, π₀.₅ on VLABench gets a 10.44× action head speedup over the default ten-step base model with similar SR. On the real robot, the overall speedup is 1.26×–1.88×.
Critical strengths
S1. The method is simple, and the key design choice is backed by experiments. ActionCache needs no extra training and doesn't touch the model weights. Using VLM output embeddings as the cache key source is supported by an ablation.
S2. The speedup on VLABench is clear. After cache warmup, the method keeps SR close to full-step generation while cutting action head latency. The real-robot experiments also show a speedup over the default base model.
Critical weaknesses
W1. The advantage over just reducing NFE only holds up on VLABench. On LIBERO, one-step inference is faster than ActionCache, with similar or higher SR. The real-robot experiments don't include a low-NFE base model for comparison, so we can't yet tell whether the cache adds anything beyond simple few-step inference.
W2. With the default threshold, fallback doesn't keep base model performance when the cache is small. In the small-cache experiments on VLABench, hit rate stays high but SR is clearly below the base model, and fallback rarely kicks in. This doesn't match the paper's claim that fallback preserves the original policy's robustness when cache coverage is insufficient.
W3. The comparison with EfficientVLA and Falcon isn't well specified. The baseline configs, tuning ranges, and how these methods were adapted to flow-based VLAs aren't described in enough detail, and the repeated-evaluation setup isn't consistent either. This makes the "outperforms existing acceleration methods" claim harder to trust. The experimental conditions need to be spelled out more clearly, with more thorough repeated evaluation.
Questions for authors
Q1. What exactly does "first" mean here, and how is this different in substance from retrieval-based priors and warm-start work? Please clarify whether the "first" claim refers to caching in general or specifically to action retrieval across episodes, and explain what's new compared to the most directly related work.
Q2. How much prior experience does real-world deployment need? The current warmup summary isn't enough to judge the setup cost for each real-robot task. Please report, per task, the number of successful and failed episodes needed for cache warmup, the interaction time, and the cost of manual success verification.
5 comments
Rebuttal (2/3)
Clarification of evaluation settings (W3)
W3.1: The comparison with EfficientVLA and Falcon isn't well specified. The baseline configs, tuning ranges, and how these methods were adapted to flow-based VLAs aren't described in enough detail
W3.2: The experimental conditions need to be spelled out more clearly
We revised Section 4.2, Appendix C.2, and C.6 to clarify the baseline configurations, evaluation settings, and implementation details for EfficientVLA and Falcon.
W3.3: (The experimental conditions need to be spelled out more clearly,) with more thorough repeated evaluation.
W3.4: repeated-evaluation setup isn't consistent either.
We originally evaluated the base model and ActionCache over 800 episodes because success rates on VLABench exhibited variations of several percentage points with only 200 episodes. We therefore prioritized additional evaluations for these two methods to obtain more reliable comparisons. For EfficientVLA and Falcon, the initial 200-episode evaluations already showed substantially less favorable trade-offs. Given the time constraints, we therefore decided to stop their evaluations at 200 episodes, considering the results sufficient to capture the overall trend. We will extend their evaluations to 800 episodes and update the results in the camera-ready version if accepted.
Clarification of “first” claim and novelty of ActionCache (Q1)
Q1.1: Please clarify whether the "first" claim refers to caching in general or specifically to action retrieval across episodes
The “first” claim in Section 2 refers to introducing action-level caching as a general, training-free acceleration paradigm for pretrained flow-based VLAs, rather than caching in general. We revised Section 2 of the main manuscript as follows:
“To the best of our knowledge, ActionCache is the first method that introduces action-level caching as a general, training-free acceleration paradigm for pretrained flow-based VLAs and optimizes both inference speed and accuracy.”
Q1.2: Please (…) explain what's new compared to the most directly related work.
As discussed in Section 2, ActionCache differs from the most closely related approaches in two key aspects. First, unlike diffusion caching methods developed for image generation (e.g., [1]), ActionCache enables retrieval of intermediate actions based on multimodal VLA representations. Second, unlike existing diffusion-policy warm-starting methods (e.g., STEP[2] and SDP[3]), which primarily rely on temporal continuity or learned modules, ActionCache retrieves and reuses actions across temporally distant contexts and episodes without additional training. We will add this clarification in Section 2.
Rebuttal (3/3)
Real-World Deployment Overhead (Q2)
Q2.1: How much prior experience does real-world deployment need? The current warmup summary isn't enough to judge the setup cost for each real-robot task. Please report, per task, the number of successful and failed episodes needed for cache warmup, (…).
We report the numbers of successful and failed warmup episodes for each real-robot task, along with estimated interaction times, in the table below.
| Task | Cache size | Number of episodes | Success rate | ||
|---|---|---|---|---|---|
| Success | Failure | Total | |||
| pick_sausage (vegetable) | 1000 | 111 | 17 | 128 | 86.72% |
| pick_and_close | 1000 | 86 | 12 | 98 | 87.76% |
| push_button | 300 | 101 | 2 | 103 | 98.06% |
Cache warmup required 86–111 successful episodes per task in our real-robot experiments, fewer than the 150 episodes used for fine-tuning each task. We also corrected the description in Appendix E, where the previously reported range of 11.6–26.7 successful episodes per task applied only to simulation experiments.
Q2.2: Please report (…) the interaction time, and the cost of manual success verification.
We did not quantitatively measure the human interaction time and effort required for manual success verification in our experiments. However, the verification procedure was limited to a single binary success/failure judgment at the end of each episode, based solely on whether the predefined final task state had been reached, and thus required minimal human effort. We also note that, as discussed in Appendix C.4, ActionCache requires only an episode-level success signal for cache admission, so this signal could in principle be provided by an automated task-completion detector.
We would like to thank you again for your careful assessment and valuable comments. We hope our answers have addressed your concerns. Please let us know if you have any questions or concerns for improving your score.
References
[1] Shubham Agarwal, Subrata Mitra, Sarthak Chakraborty, Srikrishna Karanam, Koyel Mukherjee, and Shiv Kumar Saini. “Approximate Caching for Efficiently Serving Text-to-Image Diffusion Models,” NSDI, 2024. (https://www.usenix.org/conference/nsdi24/presentation/agarwal-shubham)
[2] Jinhao Li, Yuxuan Cong, Yingqiao Wang, Hao Xia, Shan Huang, Yijia Zhang, Ningyi Xu, and Guohao Dai. “STEP: Warm-Started Visuomotor Policies with Spatiotemporal Consistency Prediction,” ICML, 2026. (https://icml.cc/virtual/2026/poster/61717)
[3] Sigmund H Høeg, Yilun Du, and Olav Egeland. “Fast Policy Synthesis with Variable Noise Diffusion Models,” ICRA, 2025. (https://ieeexplore.ieee.org/document/11127858)
Rebuttal (1/3)
We thank the reviewer for the careful review and valuable feedback on our experiments. We address the concerns and questions below:
Advantage of ActionCache over simple NFE reduction (W1)
While the base model with NFE = 1 achieves a better success-rate–latency trade-off on LIBERO, this does not imply that our approach offers limited benefits in practical settings. In real-world deployments, it is difficult to determine a priori which tasks or task phases can be reliably handled with a single denoising step. In such settings, ActionCache provides a generalizable acceleration strategy that does not require prior knowledge of task-specific NFE sensitivity, as discussed in Appendix G. Specifically, with a single refinement step (i.e. Nhit=1), ActionCache maintains success rates close to those of the NFE = 1 baseline on NFE-insensitive tasks such as LIBERO, while achieving substantial latency reductions with minimal accuracy degradation on NFE-sensitive tasks such as VLABench.
Even if the NFE = 1 base model were to outperform ActionCache in our real-robot experiments, this would not necessarily negate ActionCache's benefits beyond simple few-step inference, as discussed for LIBERO in W1.1.
With this distinction in mind, our real-robot experiments aim to validate whether ActionCache can accelerate inference in real-world environments, as described in Section 4.5. Thus, we compared ActionCache at Nhit=1 with the base model at NFE=10. Beyond this primary objective, a comparison with the NFE = 1 base model would further clarify the latency–success-rate trade-off in real-world settings. Given the limited rebuttal period, we plan to include the additional comparison in the camera-ready version if accepted.
Fallback Behavior with Small Caches (W2)
We would like to clarify the role of the hit threshold in the fallback mechanism.
The results in small-cache experiments (Figs. 9 and 10) do not indicate a failure of fallback, but an inappropriate hit threshold for small caches. In Figs. 9 and 10, the threshold remains fixed at 0.85 as the cache size decreases. As a result, less relevant cached actions are still accepted, leading to a high hit rate but reduced success rate.
As shown in Fig. 4 (a), higher retrieval similarity leads to higher success rates. Figs. 4(b) and (c) further show that increasing the threshold improves retrieval quality while increasing the fallback. Thus, raising the threshold allows ActionCache to reject less reliable cached actions and use the original policy more frequently, as intended by the fallback mechanism.
We would like to clarify the intended scope of our statement in Section 3.1 (“This fallback strategy ensures that … populated.”). Our claim is that, in unseen or complex situations, ActionCache can preserve the original policy's behavior by falling back to full-step generation when no sufficiently similar cached action is found. The results in Figs. 9 and 10 are consistent with this explanation, since the low threshold allows less relevant cached actions to be accepted instead of triggering fallback.
We revised the statement in Section 3.1 as follows: “This fallback strategy ensures that the VLA model retains its original robustness and generalization capabilities in unseen or complex situations where no sufficiently similar cached action is found.”