acceptodds
Under review as a conference paper at ICLR 2027

ActionCache: Training-Free Acceleration for Vision-Language-Action Models with Action Caching and Refinement

10 views

Abstract

Vision-Language-Action (VLA) models is a promising approach for generalizable robotic manipulation. Recently, VLA models that employ flow-matching-based action heads have shown remarkable success, as they can generate precise and smooth action sequences while capturing multimodal action distributions. However, these action heads require iterative denoising, which acts as a major computational bottleneck in inference, posing a critical challenge for real-time VLA deployment. To address this challenge, we propose ActionCache, a plug-and-play external cache that opportunistically reuses past intermediate actions. Specifically, ActionCache stores intermediate actions with compact multimodal keys and retrieves them from similar past contexts to warm-start generation from the vicinity of target actions. This mechanism reduces the number of required denoising steps, leading to substantially lower inference latency. Experimental results in simulation and real-world environments demonstrate that ActionCache maintains high task success rates in a low-latency regime, achieving action head inference acceleration of up to and for representative flow-matching-based VLA, and GR00T-N1.6, respectively.

Primary area:
applications to robotics, autonomy, planning
Keywords:
Vision-Language-Action Models, Robotics, Inference Acceleration, Efficient Inference

Reviews

Among:
Shape:
1234this paper, Bottom 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.25 over 4 reviews, above 49% of the others.
Reviewer_GGBzOpenReview ↗
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
AuthorsOpenReview ↗

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)

W1.1: 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.

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=1N_\mathrm{hit}=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.


W1.2: 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.

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=1N_\mathrm{hit}=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)

W2.1: 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.

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.


W2.2: This doesn't match the paper's claim that fallback preserves the original policy's robustness when cache coverage is insufficient.

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.”

AuthorsOpenReview ↗

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.

AuthorsOpenReview ↗

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.

TaskCache sizeNumber of episodesSuccess rate
SuccessFailureTotal
pick_sausage (vegetable)10001111712886.72%
pick_and_close100086129887.76%
push_button300101210398.06%

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)

Reviewer_5MPeOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

ActionCache is a training-free framework that accelerates flow-based VLA inference by retrieving cached intermediate action chunks from similar past contexts. Retrieved actions are executed directly or refined with a few denoising steps, while low-similarity queries fall back to full generation. Experiments cover two VLA models, simulation benchmarks, and real-robot tasks.

Critical strengths
  1. The method is simple and requires no additional training or changes to pretrained model parameters.
  2. VLABench results demonstrate substantial benefits over direct step reduction in the low-NFE regime.
  3. The evaluation includes real-world deployment, cache-management ablations, warmup analysis, and some end-to-end latency measurements.
Critical weaknesses

W1. Fallback does not guarantee robustness. Low-similarity queries trigger fallback, but high-similarity retrievals may still produce inappropriate actions. The paper lacks false-hit analysis and perturbation tests showing whether the threshold reliably rejects unsuitable reuse.

W2. Cache construction and feedback costs are incompletely evaluated. The main results assume a populated cache containing successful trajectories, with human-verified outcomes used for real-world updates. Despite the warmup analysis, cumulative end-to-end savings from an empty cache and sensitivity to imperfect success detection remain unclear.

W3. Generalization beyond the cache-population distribution is insufficiently tested. Different random seeds do not establish robustness to distribution shifts. Although model weights remain unchanged, behavior may become overly dependent on cached experience. Evaluation on LIBERO-Plus or comparable perturbation benchmarks, using caches populated only under nominal conditions, would clarify the method’s generalization boundaries.

W4. Head-level speedups do not fully characterize deployment benefits. The headline 10.44× and 40.17× gains concern the action head, and the latter accompanies a success-rate decrease from 34.0% to 30.8%. Although some end-to-end results are provided, the main comparison should also report total latency and average NFE including full-generation cache misses.

Questions for authors

Q1. How frequently do high-similarity cache hits produce unsuitable actions, and how reliably does the threshold detect such cases under controlled perturbations?

Q2. How many episodes are required before cumulative latency savings offset cache-population overhead? How sensitive is performance to errors in success verification?

Q3. How does a cache populated under nominal conditions perform on unseen perturbations, such as those in LIBERO-Plus? Please report success rate, hit rate, and end-to-end speedup.

Q4. Can the main comparison include end-to-end latency and actual average NFE accounting for cache misses, together with the corresponding success-rate trade-offs?

6 comments
AuthorsOpenReview ↗

Rebuttal (1/3)

We appreciate the reviewer's constructive comments and feedback. We address the concerns and questions below:


Fallback reliability (W1, Q1)

W1: Low-similarity queries trigger fallback, but high-similarity retrievals may still produce inappropriate actions. The paper lacks false-hit analysis and perturbation tests showing whether the threshold reliably rejects unsuitable reuse.

We would like to clarify that a "false hit" cannot be uniquely defined for an individual cache hit in VLA control. A VLA policy executes tasks through a closed-loop process: it observes the current state, generates and executes an action chunk, and then generates subsequent actions based on the updated observation. When generating each action chunk, ActionCache searches the cache and retrieves a cached chunk on a hit. Therefore, it is difficult to determine whether an individual retrieved action is appropriate in isolation.

For example, consider a robot reaching toward a target object using two consecutive action chunks. Even if the first cache hit produces a suboptimal movement, the policy may correct its trajectory when generating the second chunk. Moreover, the robot may approach the target from either the left or right, so there is no unique correct action. Thus, an individual cache hit does not necessarily determine task success or failure. Consequently, a per-hit "false-hit rate" is not well-defined, and we cannot conduct a meaningful false-hit evaluation.


Q1: How frequently do high-similarity cache hits produce unsuitable actions, and how reliably does the threshold detect such cases under controlled perturbations?

As explained in W1.1 above, since the suitability of individual cache hits cannot be uniquely defined, neither the frequency of unsuitable hits nor the threshold's ability to detect them can be meaningfully evaluated, even under controlled perturbations.


Cache construction duration and success verification (W2, Q2)

W2.1: Despite the warmup analysis, cumulative end-to-end savings from an empty cache remain unclear.

Q2.1: How many episodes are required before cumulative latency savings offset cache-population overhead?

We emphasize that no additional inference latency overhead is incurred during cache population. As described in Appendix E, cache population during warmup occurs in parallel with normal VLA inference. Specifically, during warmup, our method performs the same full-step inference as the base model while progressively caching successful action sequences in parallel with inference, resulting in the same inference latency as the base model. Once the cache is sufficiently populated, our method seamlessly transitions to cache-accelerated inference, achieving lower latency than the base model. Consequently, when starting with an empty cache, cumulative latency savings remain zero during warmup and increase after warmup, with no additional cache-population latency overhead to offset.

We briefly clarified this in the revised main manuscript (Section 4.4).


W2.2: sensitivity to imperfect success detection remain unclear.

Q2.2: How sensitive is performance to errors in success verification?

To evaluate sensitivity to errors in success verification, we conducted an experiment on VLABench with π0.5\pi_{0.5} by deliberately including cache entries from failed episodes. Specifically, we constructed caches containing 0-50% entries from failed episodes. We then froze each cache and evaluated success rates across all 10 VLABench tasks, using 80 episodes per task (800 episodes per configuration).

We show task success rates under different failure-entry ratios below.

Failure entry ratio
NhitN_\mathrm{hit}ThitT_\mathrm{hit}0%5%10%25%50%
00.8544.42%44.33%44.07%42.95%34.48%
00.9042.95%42.71%43.65%42.91%35.97%
10.8541.86%38.51%38.66%42.98%36.70%
10.9040.89%39.87%38.28%42.77%36.32%

Success rates remained relatively stable with up to 25% failure entries, while a clear degradation was observed at 50%. This trend was consistent across different NhitN_\mathrm{hit} and ThitT_\mathrm{hit} settings. Because errors in success verification can cause entries from failed episodes to be mistakenly admitted to the cache, the observed robustness to failure entries suggests that ActionCache is also likely to be robust to moderate errors in success verification.

We will conduct comprehensive experiments and report results and discussions in a camera-ready version if accepted.

AuthorsOpenReview ↗

Rebuttal (2/3)

Generalizability beyond the cache-population distribution (W3, Q3)

W3: Evaluation on LIBERO-Plus or comparable perturbation benchmarks, using caches populated only under nominal conditions, would clarify the method’s generalization boundaries.

Q3: How does a cache populated under nominal conditions perform on unseen perturbations, such as those in LIBERO-Plus? Please report success rate, hit rate, and end-to-end speedup.

We evaluated ActionCache under unseen perturbations on LIBERO-Plus using a cache populated under nominal LIBERO conditions. We used π0.5\pi_{0.5} and evaluated 10 episodes per task (400 episodes per configuration), with balanced coverage of all seven LIBERO-Plus perturbation types: object layouts, camera viewpoints, robot initial states, language instructions, lighting conditions, background textures, and sensor noise.

MethodNhitN_\mathrm{hit}ThitT_\mathrm{hit}Success rateAction head latency (ms)Cache hit rateE2E latency (ms)E2E speedup
Base model (NFE=10)––81.25%53.04–99.451.00×
Base model (NFE=1)––81.25%5.67–52.051.91×
ActionCache00.80061.50%10.5380.08%57.131.74×
ActionCache00.82572.00%24.0554.92%70.711.41×
ActionCache00.85077.00%34.7633.94%81.341.22×
ActionCache00.87580.50%43.6217.74%90.331.10×
ActionCache10.80070.00%14.9480.25%61.571.62×
ActionCache10.82578.25%25.8557.29%72.481.37×
ActionCache10.85080.75%35.7436.31%82.321.21×

At the original LIBERO threshold (Thit=0.85T_\mathrm{hit}=0.85, a setting used in Appendix G), ActionCache showed a success-rate degradation at Nhit=0N_\mathrm{hit}=0. However, increasing ThitT_\mathrm{hit} to 0.875 restored near-baseline success rates while retaining a 17.74% hit rate and 1.10× end-to-end speedup. This demonstrates that the fallback mechanism can preserve success rates under unseen perturbations through appropriate thresholding. Meanwhile, cache hit rates decreased compared to nominal LIBERO, likely because perturbations reduce the similarity between queries and cached nominal contexts, causing more fallback to full-step generation.

Overall, these results demonstrate that appropriate threshold calibration enables ActionCache to maintain near-baseline success rates while still providing end-to-end acceleration under unseen perturbations. We will add more detailed results and analyses in a camera-ready version.

AuthorsOpenReview ↗

Rebuttal (3/3)

End-to-end latency, speedup, and cache hit rate (W4, Q4)

W4.1: Head-level speedups do not fully characterize deployment benefits. The headline 10.44× and 40.17× gains concern the action head, and the latter accompanies a success-rate decrease from 34.0% to 30.8%.

W4.2: Although some end-to-end results are provided, the main comparison should also report total latency and average NFE including full-generation cache misses.

Q4. Can the main comparison include end-to-end latency and actual average NFE accounting for cache misses, together with the corresponding success-rate trade-offs?

As shown in Fig. 1, the flow-based action head is a major inference bottleneck in current VLAs. Since ActionCache specifically targets this bottleneck, we report head-level speedups (up to 10.44× and 40.17×) as headline results. To clarify the end-to-end benefits, we have expanded Tab. 1 to include E2E latency and speedup, cache hit rate, and average NFE accounting for full-generation cache misses.

ModelMethodCfg. NFEAvg. NFESR (%)HR (%)Head latency (ms)Head speedupE2E latency (ms)E2E speedup
π₀.₅Base model101038.8 ± 3.0–18.81.00×56.31.00×
π₀.₅Base model116.8 ± 2.9–2.57.52×40.01.41×
π₀.₅EfficientVLA (Head only)2217.9–5.23.62×42.71.32×
π₀.₅EfficientVLA (Head only)117.3–3.65.22×41.11.37×
π₀.₅Falcon–6.0613.1–13.01.45×50.61.11×
π₀.₅Falcon–2.937.6–6.82.76×44.41.27×
π₀.₅ActionCache (Ours)22.0139.4 ± 2.899.95.53.42×42.91.31×
π₀.₅ActionCache (Ours)11.0141.0 ± 1.299.93.65.22×41.01.37×
π₀.₅ActionCache (Ours)00.0440.9 ± 2.899.61.810.44×39.21.44×
GR00T-N1.6Base model4434.0 ± 1.2–24.11.00×43.41.00×
GR00T-N1.6Base model1124.6 ± 1.4–6.93.49×26.31.65×
GR00T-N1.6EfficientVLA (Head only)2219.7–15.31.58×34.61.25×
GR00T-N1.6EfficientVLA (Head only)118.2–9.22.62×28.61.52×
GR00T-N1.6Falcon–2.8926.1–18.81.28×38.91.12×
GR00T-N1.6Falcon–2.6119.3–15.61.54×35.71.22×
GR00T-N1.6ActionCache (Ours)22.0233.9 ± 0.999.213.61.77×34.01.28×
GR00T-N1.6ActionCache (Ours)11.0232.1 ± 3.299.37.53.21×27.81.56×
GR00T-N1.6ActionCache (Ours)00.0530.8 ± 1.898.80.640.17×20.92.08×

At Thit=0.85T_\mathrm{hit}=0.85, ActionCache achieves nearly 100% cache hit rates. For π0.5\pi_{0.5}, it achieves a 1.44× E2E speedup while maintaining the baseline success rate. For GR00T-N1.6, the maximum 2.08× E2E speedup comes with a success-rate decrease from 34.0% to 30.8%, as noted in W4.1. However, a more conservative setting (Nhit=2N_\mathrm{hit}=2) achieves a 1.28× E2E speedup while maintaining the baseline success rate (33.9% vs. 34.0%). Our central claim is that ActionCache substantially improves the latency–success-rate trade-off, not that the maximum-speedup configuration necessarily preserves the baseline success rate.


We would like to thank you again for your detailed review and constructive suggestions. We hope our answers and the new experimental results have addressed your concerns. Please let us know if you have any questions or concerns for improving your score.

Reviewer_guzBOpenReview ↗
Rating
3
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes ActionCache, a training-free, plug-and-play cache for accelerating the action head of flow-based VLAs. It uses VLM output embeddings as keys to store intermediate actions from past successful episodes. On a cache hit, generation starts from the retrieved action and requires only 0–1 denoising steps. The authors report action-head speedups of up to 10.44× on π0.5 and 40.17× on GR00T-N1.6.

Critical strengths
  1. Clear motivation and a fresh perspective. The paper shows that the action head accounts for 37–66% of end-to-end latency in flow-based VLAs, a bottleneck that prior acceleration methods targeting the VLM backbone largely leave unaddressed. Reframing warm-starting as cross-episode retrieval of past outputs is a simple and interesting idea.

  2. Training-free and plug-and-play. The method leaves model weights unchanged, caches actions directly, and builds keys from existing VLM output embeddings with negligible overhead. The fallback to full-step generation on cache misses is a reasonable safeguard.

  3. Clear analysis of retrieval quality versus success rate. The paper characterizes how Top-1 similarity, hit threshold, and cache size jointly determine success and hit rates, and compares different cache replacement policies. This provides tunable knobs for balancing accuracy, latency, and memory.

Critical weaknesses
  1. Limited generalization, with clear degradation on long-horizon tasks. Warmup and test episodes share the same tasks and scenes and differ only in random seeds, so the evaluation mainly measures in-distribution repetitiveness. On LIBERO, ActionCache at NFE = 0 reduces the average success rate from 97.1% to 92.1%, with a 12.0-point drop on LIBERO-Long. At the 40× speedup point, GR00T-N1.6 also drops from 34.0% to 30.8%. Please evaluate on unseen scenes and tasks, and analyze the failure cases on long-horizon tasks.

  2. A comparison with one-step distilled action heads is missing. On LIBERO, the base model with NFE = 1 (96.9%, 5.77 ms) outperforms every ActionCache setting in both success rate and latency. This suggests that ActionCache is beneficial only when the base model collapses at low NFE, a setting in which distilling the action head to a single step is a more natural alternative. Please compare against such methods under the same settings.

  3. The evaluation scope is limited. Although the method is claimed to be architecture-agnostic, experiments cover only two flow-based VLAs (π0.5 and GR00T-N1.6), and the main results are reported only on VLABench. Please add results with a standard Diffusion Policy (e.g., on Push-T and RoboMimic) and on additional long-horizon benchmarks such as CALVIN or RoboTwin.

  4. The real-world tasks are overly simple. The real-world evaluation uses only three short-horizon tasks on a single SO-101 arm (pick sausage, pick and close, push button), and only NFE = 1 is tested. The 94.4% hit rate on push button largely reflects the repetitive nature of the task. Even in this simple setting, the hit rate drops to 41.8% on pick sausage, which contains distractors. Please add multi-stage, long-horizon real-world tasks and evaluate generalization under variations in object position, appearance, and scene layout.

Questions for authors

Refer to the weakness part.

5 comments
AuthorsOpenReview ↗

Rebuttal (1/4)

We thank the reviewer for the positive review and valuable comments. We would like to answer the reviewer’s concerns below:


Generalization and success degradation in long-horizon tasks or high-speedup points (W1)

W1.1: Warmup and test episodes share the same tasks and scenes and differ only in random seeds, so the evaluation mainly measures in-distribution repetitiveness

Evaluating in-distribution performance directly reflects the intended use case of our method. This is because ActionCache is designed to accelerate inference by reusing cached actions from similar past contexts.


W1.2: On LIBERO, ActionCache at NFE = 0 reduces the average success rate from 97.1% to 92.1%, with a 12.0-point drop on LIBERO-Long. At the 40× speedup point, GR00T-N1.6 also drops from 34.0% to 30.8%.

The reported success-rate drops reflect aggressive Nhit=0N_{\mathrm{hit}}=0 and Thit=0.85T_{\mathrm{hit}}=0.85 settings that prioritize speedup, not an inherent accuracy limitation of ActionCache. At Nhit=2N_{\mathrm{hit}}=2, ActionCache achieves baseline-level performance on LIBERO-Long (96.0% vs. 95.4%) while still accelerating inference. This result shows that the success degradation can be mitigated even on long-horizon tasks. Similarly, at Nhit=2N_{\mathrm{hit}}=2, ActionCache achieves near-baseline success rates on GR00T-N1.6 on VLABench (33.9% vs. 34.0%).


W1.3: Please evaluate on unseen scenes and tasks

Regarding unseen scenes, we conducted evaluation of ActionCache under unseen perturbations on LIBERO-Plus using a cache populated under nominal LIBERO conditions. We used π0.5\pi_{0.5} and evaluated 10 episodes per task (400 episodes per configuration), with balanced coverage of all seven LIBERO-Plus perturbation types: object layouts, camera viewpoints, robot initial states, language instructions, lighting conditions, background textures, and sensor noise.

MethodNhitN_{\mathrm{hit}}ThitT_{\mathrm{hit}}Success rateAction head latency (ms)Cache hit rateE2E latency (ms)E2E speedup
Base model (NFE=10)––81.25%53.04–99.451.00×
Base model (NFE=1)––81.25%5.67–52.051.91×
ActionCache00.80061.50%10.5380.08%57.131.74×
ActionCache00.82572.00%24.0554.92%70.711.41×
ActionCache00.85077.00%34.7633.94%81.341.22×
ActionCache00.87580.50%43.6217.74%90.331.10×
ActionCache10.80070.00%14.9480.25%61.571.62×
ActionCache10.82578.25%25.8557.29%72.481.37×
ActionCache10.85080.75%35.7436.31%82.321.21×

As shown in this table, increasing ThitT_{\mathrm{hit}} to 0.875 at Nhit=0N_{\mathrm{hit}}=0 restored near-baseline success rates (80.50% vs. 81.25%) while retaining 1.10× E2E speedup. These results demonstrate that appropriate thresholding enables ActionCache to maintain near-baseline success rates even under unseen perturbations.

In Appendix B, we provide cross-task evaluation, where a cache is evaluated on unseen tasks. With a conservative threshold of Thit=0.925T_{\mathrm{hit}}=0.925, ActionCache achieves progress scores comparable to the base model (20.0% vs. 21.0% and 51.7% vs. 50.5%, respectively).


W1.4: Please (…) analyze the failure cases on long-horizon tasks.

As discussed in Appendix G, we observed that small deviations in retrieved actions sometimes prevented accurate grasping or placement of target objects. Refinement mitigates these errors: on LIBERO-Long, increasing NhitN_{\mathrm{hit}} from 0 to 2 improves success rate from 83.4% to 96.0%, comparable to the 95.4% baseline.

AuthorsOpenReview ↗

Rebuttal (2/4)

Comparison with distillation method and effective range of ActionCache (W2)

W2.1: A comparison with one-step distilled action heads is missing.

W2.2: Please compare against such methods under the same settings.

A direct comparison with one-step distillation would not be appropriate, as ActionCache aims to accelerate pretrained flow-based VLAs without additional training, whereas one-step distillation requires substantial training.


W2.3: On LIBERO, the base model with NFE = 1 (96.9%, 5.77 ms) outperforms every ActionCache setting in both success rate and latency. This suggests that ActionCache is beneficial only when the base model collapses at low NFE

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=1N_\mathrm{hit}=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.

AuthorsOpenReview ↗

Rebuttal (3/4)

Clarification of evaluation scope (W3)

W3.1: the main results are reported only on VLABench.

W3.2: (Please add results on) additional long-horizon benchmarks such as CALVIN or RoboTwin.

We emphasize that our VLABench evaluation includes ten diverse manipulation tasks involving different scenes, object configurations, and task objectives, providing evidence across a broad range of task settings. We also report additional results on LIBERO in Appendix G.

As the reviewer suggested, evaluations on additional long-horizon benchmarks such as CALVIN and RoboTwin would further strengthen our experimental evidence. Due to the limited time available during the rebuttal period, we commit to conducting these evaluations and presenting the results in the camera-ready version if accepted.


W3.3: Although the method is claimed to be architecture-agnostic, experiments cover only two flow-based VLAs (π0.5 and GR00T-N1.6)

Our evaluation of π0.5\pi_{0.5} and GR00T-N1.6 supports our claim that ActionCache is architecture-agnostic, as the two models adopt distinct architectural designs. In π0.5\pi_{0.5}, the action head accesses the VLM's KV cache through cross-attention, whereas GR00T-N1.6 directly feeds VLM output tokens into the action head and processes robot-state information separately from the VLM. These architectural differences provide meaningful evidence that ActionCache is applicable across different flow-based VLA designs.

Nevertheless, we agree that broader evaluation would further strengthen this claim. As additional experiments are difficult to complete within the limited rebuttal period, we will extend our evaluation in the camera-ready version if accepted.


W3.4: Please add results with a standard Diffusion Policy (e.g., on Push-T and RoboMimic)

ActionCache targets flow-based VLAs, and cannot be naively applied to standard Diffusion Policy architectures due to differences in model architecture and input modalities. However, extending ActionCache to broader architectures is an interesting direction for future work. We clarify this in Section 5 of the revised paper.

Reviewer_RBuaOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper presents ActionCache, a training-free, plug-and-play framework for accelerating flow-matching-based vision-language-action models by reusing past action generations. It caches intermediate action chunks from successful episodes using compact keys constructed from multimodal features through sparse random projection. At inference time, actions retrieved from sufficiently similar contexts are either executed directly or refined with a few flow steps, otherwise, the method falls back to standard full-step generation. This design enables action reuse across timesteps and episodes without modifying the pretrained policy. The approach is evaluated with π0.5\pi_{0.5} and GR00T-N1.6 on VLABench, alongside additional π0.5\pi_{0.5} experiments on LIBERO and three real-world manipulation tasks. The results show improved trade off between latency and success rate on VLABench, with maximum action head speedups of 10.44× and 40.17× for the two models, respectively, after cache warmup. Real-world experiments demonstrate reduced inference latency with comparable task success rates. Further analyses examine retrieval quality, cache capacity and replacement policies, hyperparameter sensitivity, and performance during warmup and sustained operation.

Critical strengths
  • Simple and effective acceleration. The paper’s main strength is that a straightforward action caching and refinement mechanism substantially reduces inference latency without requiring additional training or architectural changes to the pretrained policy. The experimental results substantiate the practical value of this design. After cache warmup, π0.5\pi_{0.5} achieves a 10.44× action-head speedup on VLABench with comparable task success (40.9% versus 38.8%; Table 1), while the three real-robot tasks demonstrate 1.26×–1.88× end-to-end inference speedups with comparable success rates (Table 2). This combination of methodological simplicity and demonstrated efficiency gains provides a practical contribution to VLA deployment.
Critical weaknesses
  • Limited methodological novelty. The core idea of retrieving cached intermediate samples to skip denoising steps is already established in approximate diffusion caching [1], which the paper cites. RT-Cache [2] also explores training-free retrieval and replay of stored action sequences for robotic control, but is not discussed. ActionCache differs by integrating VLM-derived multimodal keys and a cache populated from successful policy rollouts with intermediate flow-state refinement and fallback to the original policy. However, these changes primarily adapt and combine established ideas within the VLA setting, and the paper does not sufficiently substantiate a substantial methodological advance beyond these precedents. Although the resulting approach is practically useful, its incremental novelty remains limited.

  • Insufficient comparison with closely related acceleration methods. The main comparison evaluates EfficientVLA and Falcon on VLABench, with VLA-Cache considered separately in the end-to-end study. However, relevant training-free alternatives, including BAC [3] for intermediate-feature caching and ProbeFlow [4] for adaptive flow-step allocation, are neither evaluated nor discussed. Since their published implementations use different backbones, comparisons on the same pretrained policies or a clear explanation of adaptation barriers would be appropriate. Furthermore, EfficientVLA and Falcon are not included in the LIBERO or real-robot evaluations, leaving the comparative evidence against these methods concentrated on one benchmark. This limited coverage makes it difficult to assess the scope of ActionCache’s advantage over other relevant approaches to reducing action-generation latency.

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. Paper

[2] Owen Kwon, Abraham George, Alison Bartsch, and Amir Barati Farimani. RT-Cache — Training-Free Retrieval for Real-Time Manipulation. Humanoids, 2025. Paper

[3] Kangye Ji, Yuan Meng, Hanyun Cui, Ye Li, Jianbo Zhou, Shengjia Hua, Lei Chen, and Zhi Wang. Block-wise Adaptive Caching for Accelerating Diffusion Policy. ICLR, 2026. Paper

[4] Zhou Fang, Jiaqi Wang, Yi Zhou, and Qiongfeng Shi. ProbeFlow — Training-Free Adaptive Flow Matching for Vision-Language-Action Models. arXiv 2603.17850v1, 2026. Paper

Questions for authors
  1. Clarification of methodological novelty. Could you clarify the specific methodological contributions of ActionCache relative to approximate diffusion caching?

  2. Clarification of baseline selection and comparative evidence. Do you have additional comparisons that would clarify ActionCache’s advantage over other relevant acceleration methods?

6 comments
AuthorsOpenReview ↗

Rebuttal (1/3)

We thank the reviewer for the insightful comments on novelty and related work. We address the concerns and questions below:


Clarification of methodological novelty (W1, Q1)

W1.1: RT-Cache [2] also explores training-free retrieval and replay of stored action sequences for robotic control, but is not discussed.

We added a discussion of RT-Cache in Section 2. The additional discussion further clarifies the distinction between RT-Cache and ActionCache. While RT-Cache relies solely on image-based retrieval and directly replays (or averages) retrieved actions without refinement, ActionCache incorporates multimodal retrieval and action refinement.


W1.2: ActionCache differs by integrating VLM-derived multimodal keys and a cache populated from successful policy rollouts with intermediate flow-state refinement and fallback to the original policy. However, these changes primarily adapt and combine established ideas within the VLA setting, and the paper does not sufficiently substantiate a substantial methodological advance beyond these precedents. Although the resulting approach is practically useful, its incremental novelty remains limited.

We emphasize that a central methodological novelty of our work lies in introducing a unified, training-free, plug-and-play action caching framework for pretrained flow-based VLAs, a widely adopted class of VLA architectures.

Whereas existing training-free VLA caching methods primarily reuse intermediate computations across adjacent observations or denoising steps, our framework enables cross-episode reuse by retrieving intermediate flow states from prior policy rollouts to warm-start action generation. Furthermore, beyond this framework-level contribution, our work also introduces novelty at the component level. In particular, constructing action-relevant multimodal retrieval keys from VLM output representations and using them to retrieve cached intermediate flow states is not an established approach in prior work. Thus, we believe that these framework- and component-level contributions constitute a meaningful methodological advance in training-free acceleration of flow-based VLAs.


Q1: Could you clarify the specific methodological contributions of ActionCache relative to approximate diffusion caching?

Relative to approximate diffusion caching, ActionCache's specific methodological contributions are (1) VLM-derived multimodal retrieval keys and (2) success-conditioned cache population.

For (1), the contribution of ActionCache lies in enabling action retrieval using multimodal keys derived from high-dimensional inputs (vision, language, and robot states), with compact, action-relevant key design. This differs from approximate diffusion caching, which relies solely on text-based retrieval. We conducted an additional experiment on LIBERO-Spatial with π0.5\pi_{0.5} and found that using text-based CLIP embeddings as retrieval keys drastically reduced the task success rate to 1.0%, compared to 89.4% with ActionCache. This result suggests that text-based retrieval alone is insufficient for reliable action retrieval in flow-based VLAs, and shows the effectiveness of our multimodal retrieval mechanism. We will extend this additional evaluation and add results in the camera-ready version if accepted.

For (2), while diffusion caching stores intermediate states from cache-miss generations, ActionCache admits intermediate actions only from successful policy rollouts. This ensures that cache population is conditioned on reliable task success rather than solely on the occurrence of cache misses.

AuthorsOpenReview ↗

Rebuttal (2/3)

Clarification of baseline selection and comparative evidence (W2, Q2)

W2.1: However, relevant training-free alternatives, including BAC [3] for intermediate-feature caching and ProbeFlow [4] for adaptive flow-step allocation, are neither evaluated nor discussed. Since their published implementations use different backbones, comparisons on the same pretrained policies or a clear explanation of adaptation barriers would be appropriate.

Regarding BAC: BAC reuses intermediate features across denoising steps, a mechanism also employed by EfficientVLA. Thus, Tab. 1 has already provided an indirect comparison with this type of caching. As our method outperforms EfficientVLA, we expect it to offer advantages over BAC as well. Unfortunately, a direct comparison with BAC is difficult within the limited rebuttal period, but we commit to conducting a comprehensive evaluation for the camera-ready version if accepted.

Regarding ProbeFlow: We conducted an additional direct comparison with ProbeFlow using π0.5 on VLABench. As shown in the following table, ActionCache achieves a better NFE–success-rate tradeoff than ProbeFlow. Even with a higher average NFE, ProbeFlow fails to reach the baseline success rate, whereas ActionCache achieves baseline-level success even at an average NFE of 0.

MethodεOSCAvg. NFESuccess rate (%)
Base model––10.0038.9%
Base model––5.0037.3%
Base model––1.006.8%
ProbeFlow0.001Off8.5833.5%
ProbeFlow0.001On––
ProbeFlow0.002Off7.8028.5%
ProbeFlow0.002On7.8329.1%
ProbeFlow0.004Off6.5924.0%
ProbeFlow0.004On6.9523.8%
ProbeFlow0.008 (ref.)Off4.9117.4%
ProbeFlow0.008 (ref.)On––
ActionCache (Ours)––5.0040.3%
ActionCache (Ours)––4.0038.7%
ActionCache (Ours)––3.0138.6%
ActionCache (Ours)––2.0039.4%
ActionCache (Ours)––1.0141.1%
ActionCache (Ours)––0.0140.9%

We used the following hyperparameters for ProbeFlow: Δtprobe=0.5\Delta t_{\mathrm{probe}}=0.5, Nmin⁡=2N_{\min}=2, Nmax⁡=10N_{\max}=10, ΔN=2\Delta N=2, rref=0.02r_{\mathrm{ref}}=0.02, seasy=1.50s_{\mathrm{easy}}=1.50, shard=0.75s_{\mathrm{hard}}=0.75, smin⁡=0.60s_{\min}=0.60, smax⁡=1.60s_{\max}=1.60, and δ=10−6\delta=10^{-6}. The OSC-on results for ϵ=0.001\epsilon=0.001 and 0.0080.008 were unavailable due to execution errors; however, since OSC had little impact on success rates in other settings, similar performance to the corresponding OSC-off results are expected.

We will include a more comprehensive evaluation in the camera-ready version.

AuthorsOpenReview ↗

Rebuttal (3/3)

W2.2: EfficientVLA and Falcon are not included in the LIBERO or real-robot evaluations, leaving the comparative evidence against these methods concentrated on one benchmark. This limited coverage makes it difficult to assess the scope of ActionCache’s advantage over other relevant approaches to reducing action-generation latency.

To address this concern, we extended our comparison with EfficientVLA and Falcon to LIBERO benchmark, using π0.5\pi_{0.5}.

Experimental setup. We evaluated both methods on all four LIBERO suites (Spatial, Object, Goal, and Long), comprising 40 tasks in total, and measured success rates and action-head latency.

  • EfficientVLA. We evaluated NFE = 1, 2, and 5, using 50 episodes per task (2,000 episodes per NFE setting). All other configurations follow those used in Tab. 1.
  • Falcon. We evaluated eight hyperparameter configurations using 10 episodes per task (400 episodes per configuration). Due to the limited rebuttal period, we could not conduct 50 episodes per task for all eight configurations; nevertheless, 400 episodes per configuration provide a useful indication of the latency–success-rate trade-off. Since the exploration rate had little impact on success rates, we fixed δ=0.05\delta = 0.05 across all configurations. In the results below, ∣B∣|B| denotes the buffer size, and kmin⁡k_{\min} denotes the minimum number of remaining denoising steps after warm-start candidate selection.

Results.

MethodConfigurationSuccess rate (%)Action head latency (ms)
SpatialObjectGoal10Overall
Base modelNFE=1097.499.096.695.497.1054.07
Base modelNFE=197.698.097.494.696.905.77
EfficientVLANFE=597.298.896.496.897.3026.76
EfficientVLANFE=297.099.097.295.497.1513.12
EfficientVLANFE=193.898.496.089.494.408.53
Falconε=0.01, |B|=20, kmink_\mathrm{min}=126.00.014.00.010.0033.05
Falconε=0.01, |B|=20, kmink_\mathrm{min}=596.0100.096.093.096.2553.41
Falconε=0.01, |B|=5, kmink_\mathrm{min}=124.00.025.00.012.2523.85
Falconε=0.01, |B|=5, kmink_\mathrm{min}=596.099.098.094.096.7545.69
Falconε=0.05, |B|=20, kmink_\mathrm{min}=10.00.01.00.00.2529.14
Falconε=0.05, |B|=20, kmink_\mathrm{min}=597.098.085.093.093.2552.18
Falconε=0.05, |B|=5, kmink_\mathrm{min}=10.00.02.00.00.5019.15
Falconε=0.05, |B|=5, kmink_\mathrm{min}=596.099.087.091.093.2544.78
ActionCache (Ours)Nhit=2N_\mathrm{hit}=296.098.497.696.097.0014.25
ActionCache (Ours)Nhit=1N_\mathrm{hit}=196.499.096.293.696.3011.31
ActionCache (Ours)Nhit=0N_\mathrm{hit}=089.499.096.483.492.106.57
  • EfficientVLA. At NFE = 2 and 5, EfficientVLA achieved success rates comparable to the base model, while a slight degradation was observed at NFE = 1. Compared with ActionCache at Nhit=0N_\mathrm{hit}=0, EfficientVLA at NFE = 1 achieved a slightly higher success rate but incurred 1.3× higher latency. Overall, on LIBERO benchmark, EfficientVLA and ActionCache exhibit broadly comparable success-rate–latency trade-offs.
  • Falcon. With a conservative setting of kmin=5k_{\mathrm{min}} = 5, Falcon achieved success rates comparable to or slightly below the base model while providing approximately 1.2× speedup. However, more aggressive configurations led to substantial degradation in success rates. As discussed in Section 4.2, this degradation is likely due to the limited correctability of recent-trajectory warm-starting under deterministic flow integration.

Overall, the results of EfficientVLA and Falcon on LIBERO exhibit trends similar to those observed on VLABench in Tab. 1. These additional results broaden our comparative evaluation beyond VLABench and clarify the conditions under which ActionCache provides advantages over existing acceleration methods. We will add these results in the camera-ready version if accepted.


We would like to thank you again for your thoughtful review and valuable feedback. We hope our answers and the new experimental results have addressed your concerns. Please let us know if you have any questions or concerns for improving your score.

Comments on the paper

2 comments
AuthorsOpenReview ↗

Summary of Initial Reviews and Author Responses (For PCs, SACs, and ACs)

We sincerely thank all the reviewers and area chairs involved in the review process for our paper. We appreciate the reviewers' recognition of our reframing of warm-starting as cross-episode action retrieval (guzB), a simple, practical, training-free approach achieving substantial inference speedups (RBua, guzB, 5MPe, GGBz), and a systematic analysis of ActionCache (guzB).

We received several concerns in the original reviews, particularly regarding the novelty, evaluation, and practical applicability of ActionCache. To address these concerns, we conducted additional experiments, provided clarifications, and revised the manuscript, as summarized below.


Summary of Responses

  1. Methodological Novelty and Related Work (RBua, GGBz): We clarified how ActionCache reframes warm-starting as training-free, cross-episode retrieval of intermediate flow states. We also explained its differences from prior diffusion caching and RT-Cache in multimodal retrieval, success-conditioned cache population, and action refinement, and revised Section 2.
  2. Additional Baseline Comparisons and Evaluation Details (RBua, GGBz): We compared ActionCache with ProbeFlow on VLABench and extended the evaluation of EfficientVLA and Falcon to LIBERO. ActionCache achieves better or comparable accuracy–efficiency trade-offs across these comparisons. We also clarified the baseline implementations and evaluation settings in Section 4.2 and Appendices C.2 and C.6.
  3. Generalization and Retrieval Robustness (guzB, 5MPe, GGBz): We evaluated ActionCache under unseen perturbations on LIBERO-Plus. The results demonstrate that appropriate thresholding maintains baseline-level success rates while still accelerating inference.
  4. End-to-End Acceleration (5MPe): We expanded Table 1 to report end-to-end speedups, cache hit rates, and average NFE accounting for cache misses. The results confirm that ActionCache achieves end-to-end acceleration while maintaining baseline-level success rates.
  5. Advantages over Simple NFE Reduction (guzB, GGBz): We clarified that ActionCache can accelerate inference on both NFE-sensitive and NFE-insensitive tasks, without requiring prior knowledge of whether simple NFE reduction is effective.
  6. Success Verification and Real-world Deployment Requirements (5MPe, GGBz): To address concerns about imperfect success verification, we evaluated caches containing entries from failed episodes, finding that success rates remain stable under moderate verification errors. We also clarified that cache population occurs during normal inference and reported warm-up requirements for real-world deployment.
AuthorsOpenReview ↗

Final Summary for PCs, SACs, and ACs

We sincerely thank the reviewers and chairs for their time and consideration. As summarized earlier, we have provided responses to all weaknesses raised in the initial reviews through additional experiments, clarifications, and manuscript revisions. We also conducted further experiments in response to Reviewer GGBz's follow-up concerns regarding one-step inference in a real-robot task and fallback behavior with small caches.

Regarding Reviewer 5MPe, their follow-up response refers to “inference-time future generation,” “state probes,” and “latent interventions,” none of which are discussed in our paper. The response may have been intended for a different submission. We respectfully ask the chairs to consider this mismatch when evaluating Reviewer 5MPe's assessment.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.