We sincerely thank you for your insightful comments and constructive feedback. Below, we address your comments and questions in detail.

Q1. Which visual tokens are used to compute attention?

R1: Thanks for the question. We compute task relevance scores using the decoder self-attention weights from the previous timestep (t−1). We do not access or recompute visual tokens from multiple encoder layers, nor do we require raw image tokens as in OpenVLA. Instead, we reuse existing attention weights that were already computed during action prediction at the previous timestep.

Q2. Experiment: Reusing only task-relevant tokens (no static filtering)

R2: Our method is designed to reuse static tokens, those unlikely to influence final predictions, to reduce redundant computation without harming performance. Following your suggestion, we conducted an experiment on the LIBERO Spatial benchmark where only task-relevant tokens were reused, and all static tokens were recomputed. The results are shown below:

As expected, reusing all task-relevant tokens leads to complete performance failure (SR = 0%), confirming that these tokens are essential for accurate action prediction and should not be reused. In contrast, VLA-Cache reuses only static tokens and preserves task-relevant computations, thus maintaining high success rates while improving efficiency.

Q3. Token reuse for wrist camera input

R3: Thank you for the thoughtful question. The wrist camera input is processed identically to the third-person camera in our system. While it is true that wrist-view observations contain more motion (due to hand-object interaction and egocentric perspective shifts), not all tokens are equally dynamic. Our top-$k$ similarity-based selection captures local temporal consistency. For example, background areas, parts of the robot arm, or regions with limited motion—still exhibit high similarity across frames. These static regions enable effective token reuse.

As shown in Q2-R2, the token reuse ratios are surprisingly similar between third-person view (31.62%) and wrist view (35.47%). This indicates that, despite the higher motion, the wrist camera input still contains substantial redundancy that VLA-Cache can exploit for acceleration without hurting performance.

Q4. Control Frequency Improvement

R4: Thanks for the nice question. Across OpenVLA, OpenVLA-OFT, and CogACT, we apply the same configuration. Token reuse ratios are consistently around 35% as reported above (Q2-R2).

VLA-Cache primarily accelerates the language decoder, which accounts for over 90% of total compute in VLA models (based on FLOPs profiling). As such, control frequency improvements scale with how often the decoder is called. In low-frequency models like OpenVLA (≈4Hz), which predict one action per timestep, the performance gain is naturally limited, even a large reduction in decoder latency translates to only ~0.2–0.4Hz gain, due to other system bottlenecks like policy update frequency or observation preprocessing. In contrast, high-frequency models like OpenVLA-OFT (≈65Hz), which use action chunking to predict multiple actions per step, benefit much more. For example, we observe a +13.88Hz gain in Table 2.

We sincerely appreciate your valuable suggestions and hope our rebuttal can address your concerns. We are happy to answer any remaining questions during the discussion phase.

We sincerely appreciate your constructive and detailed feedback. Below, we address each concern point-by-point.

Q1&W1. Lack of Real-World Robot Experiments

R1: Thank you for the suggestion. We would like to clarify that VLA-Cache has already been deployed and evaluated on a real robotic platform, the Kinova Jaco2 arm. The real-world results are thoroughly reported in Section 5.2, Section 5.4, and Appendix E.3, with quantitative performance summarized in Table 5, and qualitative outcomes illustrated in Figure 3 (right) and Figure 4b. Across four real-world manipulation tasks, VLA-Cache improved success rate by 2.4% on average, reduced FLOPs by 28.2%, and lowered latency by 19.2% compared to the OpenVLA baseline.

In addition, we conducted experiments under dynamic real-world disturbances such as human motion and background clutter to further assess robustness, as detailed in Section 5.4 and Table 7. The method maintained a high success rate while significantly reducing FLOPs and latency.

Furthermore, demonstration videos of the real-robot execution has been provided via anonymous link in the abstract, showcasing the method’s practical applicability in real-world settings.

Q2&W2. Limited Backbone Generalization

R2: Thanks for the nice question. We would like to clarify that besides OpenVLA, we have already evaluated VLA-Cache across more diverse and representative VLA architectures (i.e., CogACT and OpenVLA-OFT). Notably, CogACT is a diffusion-based VLA model, which differs structurally from standard autoregressive LLM decoders, and our results on it are reported in Table 3. On the Visual Aggregation task, VLA-Cache improved success rate by +1.0%, reduced FLOPs by 17.4%, latency by 27.0%, and increased control frequency by +2.12 Hz.

Moreover, most recent state-of-the-art VLA systems (e.g., OpenVLA, CogACT, OpenVLA-OFT, UniVLA) are all built upon LLaMA2 as the language decoder. Since our current evaluation already covers this dominant backbone family across multiple variants (autoregressive, diffusion-based, and OFT-based), we believe the generalization ability of VLA-Cache is well demonstrated.

Q3&W3. Limitations of Using Attention scores as Task-Relevance Proxies

R3: Thank you for the constructive suggestion. We acknowledge that attention scores may not perfectly align with human-understood semantics. To assess the reliability of attention scores as a proxy for semantic relevance, we conducted a direct comparison between our attention-based token filtering and an object-mask-based approach using Efficient Track Anything [1,2] on the LIBERO Spatial benchmark. The results are shown in the following table.

We observe that attention-based filtering not only improves accuracy (more than 10%) but also reduces latency and computational cost, compared to the object-mask variant. While object masks provide spatial localization, they often miss fine-grained or contextual signals essential for manipulation tasks, especially when background clutter or small, relevant parts might be involved. In contrast, attention scores are dynamically generated and tightly coupled with the model’s internal reasoning, providing a lightweight, task-adaptive proxy for relevance.

Due to this year’s rebuttal policy, we cannot include visualizations, but we will incorporate illustrative examples in the final version to better explain this behavior.

Reference:

[1] Xiong, Y., et al. (2024). Efficient Track Anything. arXiv:2411.18933
[2] Xiong, Y., et al. (2024). EfficientSAM: Leveraged Masked Image Pretraining for Efficient Segment Anything. CVPR 2024

Q4. Compatibility with Token Merging and Pruning Strategies

R4: Thanks for the insightful comment. To validate this compatibility, we implemented and evaluated VLA-Cache in combination with token merging (SparseVLM [3]) and token pruning (FastV [4]) methods. The results are summarized below:

As VLA models are highly sensitive to subtle visual cues, both token merging and pruning tend to disrupt spatial consistency, an essential requirement for robotic manipulation. Our results in Table 4 of the main paper further confirm that applying these methods leads to a noticeable drop in success rate and inconsistent efficiency improvements.

In contrast, our reuse-based caching strategy is explicitly designed to preserve spatial integrity by avoiding token alteration and instead reusing computation only for static regions. While technically compatible, combining VLA-Cache with merging or pruning introduces instability and harms performance. Therefore, we recommend using our reuse strategy independently to maintain both accuracy and efficiency.

Reference:

[3] Zhang, Y., et al. (2024). Sparsevlm: Visual token sparsification for efficient vision-language model inference. ICML 2025
[4] Chen, L., et al. (2024). An image is worth 1/2 tokens after layer 2: Plug-and-play inference acceleration for large vision-language models. ECCV 2024

W4. Contribution is local rather than fundamental

R5: We respectfully clarify that our contribution extends beyond engineering optimization. Our work addresses a critical and underexplored challenge in the field: the inference inefficiency of billion-scale VLA models in real-time robotics, which currently limits their practical deployment.

Rather than implementing a software-level tweak, VLA-Cache is a modular, architecture-agnostic algorithm carefully designed to exploit intrinsic structural properties of VLA models. It integrates three steps including static token detection, task-relevance filtering, and layer-adaptive reuse. Importantly, our method is plug-and-play: it requires no retraining or fine-tuning, and generalizes across diverse VLA architectures.

Beyond practical acceleration, VLA-Cache also offers insight into the learning and information flow of VLA models. It reveals that token-level redundancy and temporal locality are prominent, exploitable features. Selective reuse of stable representations preserves performance while substantially reducing computational cost. These findings open new directions for building efficient and deployable VLA systems.

We sincerely appreciate your valuable suggestions and hope our rebuttal can address your concerns. We are happy to answer any remaining questions during the discussion phase.

We sincerely thank you for your efforts in reviewing our paper. We have provided corresponding responses and results, which we believe have covered your concerns. We hope to further discuss with you whether your concerns have been addressed or not. Please let us know if you still have any unclear parts of our work.

Best Regards,

Authors of Submission 5408

We sincerely thank you for your effort in reviewing our paper. Your insightful and helpful comments helped us for refining our paper better. Below, we address your comments and questions in detail.

Q1. Code release

R1: Thank you for raising this important point. To this end, we have included the full implementation of VLA-Cache in the supplementary material for verification.

We are fully committed to open-sourcing the complete codebase upon paper acceptance, including scripts for reproducing all benchmarks and real robot deployment. We hope this will support further research and practical deployment.

Q2.1. Applicability to Point Cloud or Non-2D Inputs

R2.1: Thanks for this insightful question. Our current work focuses on vision-language-action (VLA) models that process 2D image inputs, which remain the dominant form across recent models (e.g., OpenVLA, OpenVLA-OFT, $\pi_0$) and large-scale datasets like Open X-Embodiment.

We acknowledge that the proposed VLA-Cache is designed and evaluated in this 2D visual context. However, the core idea of selective token reuse based on temporal visual redundancy is not inherently tied to 2D images. For multi-modal VLA models (e.g., PointVLA [1]) that integrate point cloud data, VLA-Cache can still operate on the image branch. Moreover, similar principles such as identifying static or low-saliency subsets of unordered point cloud tokens, can also be extended to non-2D domains with appropriate adaptation. While evaluating this on point cloud inputs is beyond the current scope, we believe our method offers general insights into computation-efficient inference and is compatible with broader extensions in future work.

Reference:

[1] Li, C., et al. (2025). Pointvla: Injecting the 3d world into vision-language-action models. arXiv preprint arXiv:2503.07511.

Q2.2. Applicability to Robot-Mounted Cameras (e.g., Wrist Cameras)

R2.2: We validated VLA-Cache on OpenVLA-OFT, which includes inputs from both third-person and wrist-mounted (egocentric) cameras. As shown in Figure 4c/d and Table 2, the method maintains robustness under dynamic viewpoints and is able to reuse a significant portion of tokens even with egocentric inputs.

Q3. Applicability to Other VLA Architectures (e.g., DP or RDT)

R3: Thank you for this valuable suggestion. VLA-Cache is designed to accelerate the language decoder stage, and is therefore not directly applicable to VLA models without a VLM backbone, such as standalone DP-based architectures. However, for hybrid architectures, where a VLM encoder/decoder is combined with a diffusion policy head (e.g., CogACT [2]), VLA-Cache remains fully applicable. In fact, we have already validated our method on CogACT, achieving both efficiency and performance gains (please see Table 3).

Beyond this, transformer-based models such as RDT or DiT also exhibit inter-frame redundancy, and our token-reuse strategy can be adapted to these settings as well. We consider this an exciting direction for future generalization.

Reference:

[2] Li, Q., et al. (2024). Cogact: A foundational vision-language-action model for synergizing cognition and action in robotic manipulation. arXiv:2411.19650

We sincerely appreciate your valuable suggestions and hope our rebuttal can address your concerns. We are happy to answer any remaining questions during the discussion phase.

We sincerely thank you for your efforts in reviewing our paper. We have provided corresponding responses and results, which we believe have covered your concerns. We hope to further discuss with you whether your concerns have been addressed or not. Please let us know if you still have any unclear parts of our work.

Best Regards,

Authors of Submission 5408

Thank you for your efforts of reviewing! Please remember to read the rebuttal as soon as possible, and discuss with the authors for questions not resolved. According to the guideline this year, reviewers need to participate in discussions with authors before submitting “Mandatory Acknowledgement”.

The authors sincerely than you for your insightful comments and constructive feedback. Below, we address your comments and questions in detail.

Q1. Sensitivity of Thresholds (k) and ($\tau$)

R1: The parameters $k$ (for static token selection) and $\tau$ (for task relevance thresholding) can influence the trade-off between accuracy and efficiency. To further examine the sensitivity of these parameters, we conducted additional experiments on the LIBERO Spatial benchmark using the OpenVLA-OFT model. The results are summarized below.

Table 1. Varying $k$ values (with $\tau$ = 0.5):

Table 2. Varying $\tau$ values (with k = 100):

These results show that VLA-Cache is generally robust to changes in $k$ and $\tau$. While efficiency (FLOPs and latency) improves with higher values, task success rate remains consistently high across settings. Our default configuration (k = 100, τ = 0.5) offers a strong trade-off between performance and efficiency.

Q2. Whether Task Relevance Computation Uses KV-Cache

R2: Thanks for the question. We do not recompute $Q$, $K$ and $V$ values during inference for task-relevant token selection. Instead, we reuse the attention scores from the decoder at the previous timestep (t−1). These attention scores were already computed for action prediction and reflect semantically meaningful task-driven importance.

To validate this choice, we conducted an ablation comparing:

• Current Implementation: Using full decoder attention from the previous timestep $t−1$;

• A baseline variant that recomputes $Q$, $K$ and $V$ for the first two decoder layers at timestep $t$, and uses their attention scores.

The results show that our method achieves both better efficiency and higher task success. Recomputing shallow attention at the current timestep leads to worse SR and latency, likely due to insufficient task-level semantics in early decoder layers. This supports our design choice of reusing full attention from the prior timestep.

Q3. Applicability to Diffusion-Based VLA Models

R3: We would like to clarify that our paper already includes experiments on CogACT [1], a representative diffusion-based VLA model. Specifically, Table 3 in our main paper reports the performance of VLA-Cache applied to CogACT in the SIMPLER environment. These results demonstrate that our method is compatible with diffusion-based VLA models and remains effective in this setting.

VLA-Cache operates at the VLM decoder level by identifying and reusing static visual tokens across time. This design is agnostic to the decoder architecture and naturally integrates with diffusion-based policies like CogACT. Since token caching occurs before action generation, it does not interfere with the diffusion process.

As shown in our results, applying VLA-Cache to CogACT improves inference efficiency (e.g., lower FLOPs and latency) without compromising task success, confirming both compatibility and benefit.

Table3: Comparison of VLA-Cache with CogACT in SIMPLER Environment

Reference:

[1] Li, Q., et al. (2024). Cogact: A foundational vision-language-action model for synergizing cognition and action in robotic manipulation. arXiv:2411.19650

Q4. Performance Drop in PickPot and Limited Control Frequency Improvement

R4: Thanks for the nice concern. We acknowledge that the PickPot task shows a 5% drop in real-robot success rate. However, this change is within the expected statistical variance, given only 20 trials per task. Other real-world tasks such as PlaceCube show significant improvements (e.g., +7.3%), indicating no systematic degradation from applying VLA-Cache. Importantly, VLA-Cache is designed to improve computational efficiency while maintaining performance. Across all evaluated domains (LIBERO, SIMPLER, real-robot), task success rates remain stable or improve, confirming that caching does not harm control quality.

Regarding the improvement in control frequency, we note that our method accelerates the language decoder, which accounts for most of the VLA model’s computation. The end-to-end gain scales with decoder usage: high-frequency models like OpenVLA-OFT see substantial improvements (+13.88 Hz), while low-frequency pipelines (e.g., OpenVLA real-robot) are also constrained by external system bottlenecks such as sensor and actuator delays. Thus, the modest frequency gain in PickPot reflects platform-level limits rather than algorithmic shortcomings.

We sincerely appreciate your valuable suggestions and hope our rebuttal can address your concerns. We are happy to answer any remaining questions during the discussion phase.

We sincerely thank you for your efforts in reviewing our paper. We have provided corresponding responses and results, which we believe have covered your concerns. We hope to further discuss with you whether your concerns have been addressed or not. Please let us know if you still have any unclear parts of our work.

Best Regards,

Authors of Submission 5408