acceptodds
Under review as a conference paper at ICLR 2027

PaceVLA: Adaptive Temporal Action Quantization in Vision-Language-Action Models

Abstract

Fast and reliable manipulation is crucial for practical robot deployment. Although Vision-Language-Action (VLA) models have recently shown remarkable progress as generalist robot policies, their execution often remains temporally inefficient. A key source of this inefficiency is that policies learn fine-grained temporal patterns from teleoperated demonstrations, requiring many low-level action steps. Temporal action quantization offers a simple way to accelerate execution by aggregating consecutive low-level actions into temporally coarser actions. However, applying action quantization without considering when fine-grained control is required introduces a speed–accuracy trade-off: accelerated actions may efficiently reduce redundant steps in some phases, while discarding critical motion details in precision-sensitive interactions. To address this, we propose PACEVLA, a unified framework for adaptive temporal execution in VLAs that integrates two coupled components: (1) mode-specific action decoders built on a shared action expert, supporting multiple temporal granularities and prediction horizons, (2) a quantizability-guided mode selection scheme which decides whether coarse execution is appropriate, and a context-conditioned router that selects an execution mode at each inference step. Experiments across three multi-task simulation benchmarks and six real-world tasks show that PACEVLA improves execution efficiency while maintaining task success rate across multiple embodiments, including long-horizon and dexterous manipulation. The is available.

Primary area:
applications to robotics, autonomy, planning
Keywords:
Vision-language-action models, Robot manipulation

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_z39SOpenReview ↗
Rating
2
Confidence
4(+1)
Informativeness
2
Presentation
2(−1)
Soundness
1
Read review
Paper summary

This paper proposes PaceVLA, a framework for accelerating VLA execution through adaptive temporal action quantization. The key idea is to dynamically adjust action granularity and prediction horizon according to the precision requirements of different manipulation phases. Specifically, the method introduces mode-specific action decoders built upon a shared action expert, together with a quantizability judge and a context-conditioned router for selecting execution modes. The quantizability judge is supervised using LVLM-generated annotations, while the router is trained based on mode-specific action prediction errors. Experiments across three simulation benchmarks and six real-world tasks demonstrate improved execution efficiency while largely preserving task success rates.

Critical strengths
  1. Practical and well-motivated problem. The paper addresses an important limitation of imitation-learned robot policies: inefficient execution inherited from conservative human demonstrations. Adaptive temporal granularity is a reasonable approach to balancing execution speed and control precision.

  2. Extensive experimental evaluation. The experiments cover multiple simulation benchmarks, two VLA backbones, and six real-world tasks across three robot platforms, including dexterous and humanoid manipulation. The real-world results are particularly valuable and demonstrate meaningful improvements in execution efficiency.

  3. Relatively lightweight integration. The shared action expert and mode-specific decoders provide a practical way to support multiple temporal resolutions without duplicating the entire policy. The ablation studies demonstrate the contributions of the main components, and the additional inference overhead is small.

Critical weaknesses
  1. Fundamentally questionable routing objective. The router is supervised by a target distribution constructed from mode-specific Flow Matching prediction errors, i.e., rk∝exp⁡(−ℓk/β)r_k \propto \exp(-\ell_k/\beta). However, these errors are not directly comparable across different temporal granularities and prediction horizons. Shorter prediction horizons are inherently less affected by long-term uncertainty, while temporally coarsened actions often discard fine-grained motion details and can be easier to predict. Consequently, a mode may receive a lower prediction loss simply because its prediction target is easier, rather than because it leads to better execution. Averaging losses over valid action elements does not resolve this fundamental issue. More importantly, prediction accuracy does not necessarily correlate with execution success or efficiency. The proposed routing objective therefore lacks a convincing justification, and the experiments do not establish whether it learns meaningful mode preferences or merely exploits differences in prediction difficulty.

  2. Insufficient experimental disentanglement. The method jointly changes action granularity and prediction horizon, but their individual contributions are not isolated. Moreover, the SAIL baseline is evaluated with action scheduling disabled, and the accelerated policies benefit from controller gain adjustments while the base VLA retains its default gains in simulation. These choices complicate the interpretation of the reported improvements.

  3. Limited methodological novelty. The framework mainly combines existing ideas, including temporal action quantization, multi-horizon action prediction, and expert routing. Although the integration is practically useful, the technical contributions appear relatively incremental. In particular, the routing formulation does not provide a principled objective for optimizing the speed–accuracy trade-off.

Questions for authors
  1. How are the Flow Matching losses calibrated across different modes? Shorter and coarser action predictions can naturally have lower errors. Have the authors analyzed whether the routing targets are systematically biased toward certain modes? How strongly do these losses correlate with actual execution success and efficiency?

  2. Can the authors compare the learned router with simple alternatives, such as always selecting the shortest horizon within the group chosen by the quantization judge, or using uniform/random routing? An oracle router based on rollout outcomes would also be informative.

4 comments
AuthorsOpenReview ↗

[Response to Reviewer z39S (1/2)]

Dear Reviewer z39S,

We sincerely appreciate your insightful comments and efforts in reviewing our manuscript. We address your comments and questions one-by-one below. In the revised draft, we mark our major revisions as “purple”.


[Q1, W1] Whether the per-mode losses are comparable across horizons and granularities, whether the router target is biased toward easier targets, and how the losses relate to execution success and efficiency.

[A1] The routing targets are not biased toward shorter or coarser modes, and following them improves both success and efficiency. Losses are compared only between the two horizons of the same quantization level, so coarser predictions are not compared with finer ones, and the judge decides quantization (Section 3.4).

The target is a softmax of the negative losses, the soft assignment analyzed in annealed multiple choice learning [1], and the balance term is the load-balancing loss of sparse mixture-of-experts routing [2], which keeps any single mode from being selected everywhere. Within each group, the shorter horizon has the lower loss in 53 percent of the chunks among fine-grained modes and 55 percent among quantized modes (9,747 demonstration chunks, RoboCasa-Kitchen; Appendix E.2).

A lower per-step prediction error keeps the executed actions closer to the demonstration, and imitation learning theory bounds the closed-loop performance gap by this error [3, 4, 5]. In closed loop, routing by the losses gives higher success and ES than random or fixed-horizon routing (66.9 and 1.14 vs 62.0 to 64.2 and 0.96 to 1.05, Q2).

[1] Perera et al., Annealed Multiple Choice Learning: Overcoming limitations of Winner-takes-all with annealing, NeurIPS 2024

[2] Fedus et al., Switch Transformers: Scaling to Trillion Parameter Models with Simple and Efficient Sparsity, JMLR 2022

[3] Xu et al., Error Bounds of Imitating Policies and Environments, NeurIPS 2020

[4] Block et al., Provable Guarantees for Generative Behavior Cloning: Bridging Low-Level Stability and High-Level Behavior, NeurIPS 2023

[5] Foster et al., Is Behavior Cloning All You Need? Understanding Horizon in Imitation Learning, NeurIPS 2024


[Q2] Learned router versus simple alternatives such as always-shortest-in-group, random routing, or a rollout-based oracle.

[A2] The learned router gives higher success and ES than every simple alternative, so the gain does not come from the judge alone. We replaced only the horizon choice within the judge's group on the same checkpoint (RoboCasa-Kitchen; Table 6a in Section 4.3):

\hline \text{Horizon choice within the group} & \text{SR} \uparrow & \text{TPR} \uparrow & \text{ES} \uparrow \newline \hline \colorbox{#EDE5F7}{Learned router (ours)} & \colorbox{#EDE5F7}{\textbf{66.9}} & \colorbox{#EDE5F7}{\textbf{2.75}} & \colorbox{#EDE5F7}{\textbf{1.14}} \newline \text{Uniform random} & 64.2 & 2.64 & 1.05 \newline \text{Always the shortest horizon} & 63.9 & 2.57 & 1.03 \newline \text{Always the longest horizon} & 62.0 & 2.33 & 0.96 \newline \hline \end{array}$$ Always taking the longer horizon keeps executing the end of a chunk even when it is poorly predicted, so errors accumulate until the next replan. Always taking the shorter one replans in the middle of motions that were predicted well. The learned router keeps the long plan when its predicted error is low and replans early when it is not. A rollout-based oracle would need to re-run every chunk under every mode from the same state, which is not possible in closed loop.
AuthorsOpenReview ↗

[Response to Reviewer z39S (2/2)]

[W2] Disentangling temporal quantization from horizon selection, the disabled action scheduling of SAIL, and the controller gains.

[A3] The speedup comes from temporal quantization, and neither the disabled action scheduling of SAIL nor the controller gains work in PaceVLA's favor.

Restricting, at evaluation time, the modes the router may select on the same checkpoint separates the two factors. Without the quantized modes the speedup disappears, and without horizon selection the policy executes poorly predicted chunk ends and success drops (RoboCasa-Kitchen; Tables 6a and 6b in Section 4.3):

\hline \text{Variant} & \text{SR} \uparrow & \text{TPR} \uparrow & \text{ES} \uparrow \newline \hline \colorbox{#EDE5F7}{PaceVLA (ours)} & \colorbox{#EDE5F7}{\textbf{66.9}} & \colorbox{#EDE5F7}{\textbf{2.75}} & \colorbox{#EDE5F7}{\textbf{1.14}} \newline \text{No horizon selection} & 62.0 & 2.33 & 0.96 \newline \text{No quantization} & 64.1 & 2.05 & 0.99 \newline \hline \end{array}$$ Action scheduling only hides inference latency when inference and execution run asynchronously. The simulation benchmarks and the Franka and OpenArm tasks run synchronously, so it has nothing to hide there, and on the asynchronous ALLEX task every method, SAIL included, used the same real-time chunking (Appendix C.6). A quantized action moves further per control step, so the higher arm gain is part of executing quantized actions. The base model with the same gains loses success, because its chunks are not produced for those gains (Table 14 in Appendix E.3): $$\begin{array}{lccc} \hline \text{RoboCasa-Kitchen} & \text{SR} \uparrow & \text{TPR} \uparrow & \text{ES} \uparrow \newline \hline \text{Base model, default gains} & 65.7 & 1.82 & 1.00 \newline \text{Base model, PaceVLA's gains} & 63.3 & 1.96 & 0.96 \newline \colorbox{#EDE5F7}{PaceVLA (ours)} & \colorbox{#EDE5F7}{\textbf{66.9}} & \colorbox{#EDE5F7}{\textbf{2.75}} & \colorbox{#EDE5F7}{\textbf{1.14}} \newline \hline \end{array}$$ --- **[W3] Limited methodological novelty, and no principled objective for the speed-accuracy trade-off.** **[A4]** **PaceVLA introduces a capability that prior acceleration methods lack:** one shared action DiT is trained on the same motion at two temporal granularities, and learning both makes the policy more accurate than the base model it was fine-tuned from, not just faster. On RoboCasa-Kitchen the full model reaches 66.9 against 65.7 for the base model at an ES of 1.14, while removing the quantized decoders and quantizing the base prediction instead drops it to 63.3 (Table 5a). Temporal quantization [1] fixes one granularity for the whole trajectory, and MoH [2] fuses heads of several horizons into a single granularity. PaceVLA's mode-specific decoders, quantizability-supervised judge and context-conditioned router are designed together to switch between granularities within a task, adding only about 2 ms of inference latency (Table 8; Section 2 and Appendix E.5). This design is what lets acceleration keep success. On the real robots, fixed quantization loses 21.9 to 28.2 points on Air Fryer and Stack Cup, while PaceVLA shortens execution by 33 to 45 percent with success close to the base policy (Table 2). The speed-accuracy trade-off has a dedicated objective for each side. The judge learns when quantized execution is safe, and its threshold sets how often it is used (Table 6c). The router learns which mode predicts most accurately within that group, and its objective has no term for speed, so a router trained without the judge does not accelerate execution (ES 1.01, Table 5a). [1] Tang et al., VLASH: Real-Time VLAs via Future-State-Aware Asynchronous Inference, arXiv 2025 [2] Jing et al., Mixture of Horizons in Action Chunking, arXiv 2025

Thank you for the detailed rebuttal and additional experiments. However, my main concern regarding the theoretical justification of the routing objective remains unresolved.

  1. The cited imitation learning theory does not justify comparing Flow Matching losses across different action horizons.

The authors argue that lower prediction error leads to better closed-loop performance, citing Xu et al. (2020), Block et al. (2023), and Foster et al. (2024). However, these references establish substantially different results:

  • Xu et al. (2020) derive an upper bound on the policy value gap based on the KL divergence between the learned and expert policies.
  • Block et al. (2023) establish bounds on imitation trajectory discrepancies using distributional errors under specific low-level stability and policy continuity assumptions.
  • Foster et al. (2024) analyze the relationship between behavior cloning with log loss, sample complexity, and task horizon. Notably, their Figure 6 also shows that MSE validation loss can be poorly correlated with rollout reward, even within a fixed task setting.

None of these works establishes that Flow Matching regression losses are directly comparable across different action horizons, let alone that their relative ordering predicts closed-loop performance.

The fundamental distinction is between bounding the performance gap using an imitation error under a given formulation and using errors from different prediction targets and execution horizons to rank competing policies. Even if lower imitation error yields a tighter theoretical upper bound, this does not imply better actual performance. Furthermore, different action horizons change the prediction targets, replanning frequency, and closed-loop error accumulation.

The authors' routing objective relies precisely on this unproven relationship. Therefore, I find the theoretical argument in the rebuttal insufficient and potentially misleading. The authors should either provide a theoretical justification specifically addressing cross-horizon loss comparability or acknowledge that this is an empirical heuristic rather than a theoretically supported routing criterion.

  1. The rebuttal claims that losses are compared only within the same quantization group, but Eq. (2) still defines the KL supervision over all K modes; restricting selection to a group at inference does not eliminate cross-granularity effects during training.

Overall, while I appreciate the additional experiments, I remain unconvinced by the central methodological justification.

AuthorsOpenReview ↗

Reply Rebuttal Comment by Authors

We appreciate the reviewer's response and the careful reading of the cited works. We acknowledge that these works do not directly justify comparing flow matching losses across different action horizons in robotic manipulation tasks. Our intent in citing these works was only to support a general point that a lower imitation error during training reduces the possibility to failure at test phase. We therefore address the reviewer's concerns directly below, with additional analysis.

(1) Comparing Flow Matching losses across action horizons. We note that the two losses within a group have the same scale, because each is an average of per-step errors and does not grow with the horizon. One might expect the longer horizon to have the larger loss in most chunks, since its later steps are more affected by long-term uncertainty, as the reviewer noted in the initial review. However, as reported in our initial response, the shorter horizon has the lower loss in only 53% of the fine-grained chunks and 55% of the quantized chunks, so which mode wins depends on the chunk, not on the horizon length.

We therefore train the router to predict this comparison for each chunk, and test it in closed loop. On the same checkpoint (RoboCasa-Kitchen, 24 tasks x 50 episodes), we replaced the learned router with simpler rules (random selection, short-horizon only, long-horizon only, and the router's less preferred horizon). As a result, every alternative lowers SR, TPR and ES (see the Table below). This indicates that if the policy always selects the longer action horizons, error accumulates even on the next replanning step. Likewise, if the policy always selects the shorter action horizons, it loses the chance to select the longer plans that were predicted well on the given phase. This makes the policy replan more often, the inconsistency between consecutive chunks grows, and more error accumulates over the full trajectory, which lowers the success rate.

\hline \text{Horizon choice within the group} & \text{SR} \uparrow & \text{TPR} \uparrow & \text{ES} \uparrow \newline \hline \text{Learned router (ours)} & 66.9 & 2.75 & 1.14 \newline \text{Uniform random} & 64.2 & 2.64 & 1.05 \newline \text{Always the shortest horizon} & 63.9 & 2.57 & 1.03 \newline \text{Always the longest horizon} & 62.0 & 2.33 & 0.96 \newline \text{Inverted router (less preferred horizon)} & 63.8 & 2.55 & 1.02 \newline \hline \end{array}$$ **(2) KL supervision over all K modes.** We note that training over all $K$ modes does not change which horizon the router learns to prefer within a group, and it becomes clear when training and inference are separated. In our earlier response, we wrote that losses are compared only within a group. What we meant there was how the router is used at inference, and we should have said so more clearly. During training, the reviewer notes that the KL supervision of Eq. (2) normalizes its target over all $K$ modes, so the losses of the other quantization group also affect the target values. However, they give more weights on the targets of the two modes in the same group by the same factor, so the preference of the two modes is decided by their own two losses within the group. The losses of the other group only change the target probability of their group, not changing the other group. At inference, the quantization judge decides the group, and the router only compares the two horizons inside it. The router's preference between the groups is not used, and its choice within the selected group follows the same preference as a target computed only within that group.
Reviewer_puynOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper argues that existing action quantization methods overlook when fine-grained control is needed. To address this issue, it proposes PaceVLA to accelerate VLA policies. PaceVLA combines mode-specific decoders built on a shared action expert with a quantizability-guided mode selection scheme that determines when coarse execution is appropriate. Experiments on LIBERO, RoboCasa, DexJoco and real world tasks demonstrate the effective of the method.

Critical strengths
  1. The motivation is clear: robots should adapt their execution speed to different task phases, which prior work does not fully account for. The proposed method directly addresses this issue and has a sound underlying intuition.

  2. The experiments cover several standard benchmarks, including RoboCasa-Kitchen and LIBERO, and the real-world results are promising. The ablation studies provide further evidence supporting the effectiveness of the proposed components.

Critical weaknesses
  1. The argument that free-space motion can tolerate coarser execution while grasping requires finer control is reasonable, but the paper provides limited direct empirical evidence to support it. In particular, the evaluation does not clearly establish how different levels of quantization affect execution success across task phases, or whether the judge predicted scores reliably capture these differences. Moreover, the predefined temporal modes and the judge threshold are not sufficiently justified, either theoretically or empirically. The paper provides limited explanation of how these configurations were selected and how sensitive the reported gains are to alternative choices. This raises concerns about the amount of task-specific tuning required and the robustness of the method across tasks and embodiments.

  2. The technical novelty appears limited, as the framework primarily combines established ideas, including mode-specific decoders, learned routing, and VLM-assisted supervision. Its main contribution lies in integrating these components for adaptive temporal execution. In particular, it remains unclear what adaptations are required to address VLA-specific challenges, why this combination is particularly suitable for VLA policies, and how the resulting mechanism differs from existing routing and temporal adaptation approaches. A more explicit comparison with the closest prior methods would help establish the contribution.

Questions for authors
  1. How were the temporal mode configurations and the judge threshold selected?How sensitive are the reported gains to these settings?

  2. As for the router,is there any evidence supporting using lower error as a critertion for selection the excution mode?

  3. What methodological advances does PaceVLA offer over existing approaches, and why is its design particularly suited to VLA policies?

  4. What are the individual contributions of adaptive temporal quantization and adaptive horizon selection? Can the author conduct more experiments on this?

  5. Beyond the simulation in the main paper, Could the authors additionally evaluate the method on LIBERO-PLUS to provide stronger evidence of its generalization capabilities?

4 comments
AuthorsOpenReview ↗

[Response to Reviewer puyn (1/2)]

Dear Reviewer puyn,

We sincerely appreciate your insightful comments and efforts in reviewing our manuscript. We address your comments and questions one-by-one below. In the revised draft, we mark our major revisions as “purple”.


[Q1, W1-1] Selection and Sensitivity of the temporal modes and judge threshold

[A1] Both the temporal modes and the judge threshold follow fixed rules across all tasks, rather than being tuned for each task. We describe each rule and its sensitivity below.

For the temporal modes, with H the action chunk length, the fine-grained modes predict H and H/2 steps and the quantized modes predict the same horizons at quantization ratio q ≈ 2. This rule is used on RoboCasa-Kitchen, LIBERO, DexJoCo, and the three real robots. A larger ratio lowers the success rate while barely shortening the successful episodes (RoboCasa-Kitchen; Table 13 in Appendix E.1):

\hline \text{Quantization ratio} & \text{SR} \uparrow & \text{TPR} \uparrow & \text{ES} \uparrow \newline \hline \colorbox{#EDE5F7}{\textbf{2.0}} & \colorbox{#EDE5F7}{\textbf{66.9}} & \colorbox{#EDE5F7}{\textbf{2.75}} & \colorbox{#EDE5F7}{\textbf{1.14}} \newline 2.5 & 64.0 & 2.64 & 1.04 \newline 4.0 & 58.0 & 2.01 & 0.83 \newline \hline \end{array}$$ For the judge threshold, we cut the quantizability score, which is a relative score within a dataset, and we set it by the same rule everywhere: the most aggressive value that keeps the success rate of the base model. Lowering it quantizes more chunks and costs some success, raising it quantizes fewer and costs some speed, and ES stays above 1 and TPR above the base model on both sides (Table 6c in Section 4.3). $$\begin{array}{cccc} \hline \text{Threshold} & \text{SR} \uparrow & \text{TPR} \uparrow & \text{ES} \uparrow \newline \hline 0.45 & 63.3 & \textbf{3.04} & 1.04 \newline \colorbox{#EDE5F7}{0.55} & \colorbox{#EDE5F7}{\textbf{66.9}} & \colorbox{#EDE5F7}{2.75} & \colorbox{#EDE5F7}{\textbf{1.14}} \newline 0.65 & 65.8 & 2.53 & 1.08 \newline 0.75 & 64.8 & 2.13 & 1.02 \newline \hline \end{array}$$ --- **[Q2] Evidence that the router should favor the mode with the lower prediction error.** **[A2]** The most direct evidence comes from closed-loop rollouts. On the same checkpoint, we replaced the learned router with simpler rules and kept everything else fixed. **Every alternative lowers success.** Uniform random selection loses 2.7 points, always choosing the shortest horizon loses 3.0, and always choosing the longest loses 4.9 (RoboCasa-Kitchen, 24 tasks x 50 episodes; Table 6a in Section 4.3). $$\begin{array}{lccc} \hline \text{Router criterion} & \text{SR} \uparrow & \text{TPR} \uparrow & \text{ES} \uparrow \newline \hline \colorbox{#EDE5F7}{\textbf{Learned router (ours)}} & \colorbox{#EDE5F7}{\textbf{66.9}} & \colorbox{#EDE5F7}{\textbf{2.75}} & \colorbox{#EDE5F7}{\textbf{1.14}} \newline \text{Uniform random} & 64.2 & 2.64 & 1.05 \newline \text{Always the shortest horizon} & 63.9 & 2.57 & 1.03 \newline \text{Always the longest horizon} & 62.0 & 2.33 & 0.96 \newline \hline \end{array}$$ The criterion works because the router only chooses between two horizons at the same temporal resolution; the judge has already decided whether to quantize. Between those two, lower prediction error means the executed actions stay closer to the demonstration, which is the quantity that bounds the closed-loop gap in imitation learning [1, 2, 3] (Section 3.4). The failures in the table follow from this. Always taking the long horizon keeps executing the second half of a plan even when it was poorly predicted, so errors build up until the next replan. Always taking the short horizon replans in the middle of motions that were predicted well. The learned router keeps the long plan when its expected error is low and replans early when it is high. [1] Xu et al., Error Bounds of Imitating Policies and Environments, NeurIPS 2020 [2] Block et al., Provable Guarantees for Generative Behavior Cloning: Bridging Low-Level Stability and High-Level Behavior, NeurIPS 2023 [3] Foster et al., Is Behavior Cloning All You Need? Understanding Horizon in Imitation Learning, NeurIPS 2024
AuthorsOpenReview ↗

[Response to Reviewer puyn (2/2)]

[Q3, W2] What is new in PaceVLA beyond temporal action quantization, and why it fits VLAs.

[A3] Where prior methods speed up a trained policy from outside, PaceVLA learns quantized actions as part of the policy itself. Its shared action DiT is trained on each motion at two temporal granularities, and this makes the policy more accurate than the base model it was fine-tuned from as well as faster: 66.9 against 65.7 on RoboCasa-Kitchen at an ES of 1.14, while quantizing the base prediction without the quantized decoders gives 63.3 (Table 5a).

This is also where PaceVLA differs from DemoSpeedUp and SAIL (Section 2 and Appendix E.5). DemoSpeedUp resamples the demonstrations, so the policy is trained on altered data. SAIL raises the robot's execution speed, so the learned actions no longer match how the robot actually responds to them. PaceVLA leaves both the dataset and the control frequency unchanged, and chooses the granularity within a task at about 2 ms of added inference latency (Table 8).


[Q4, W1-2] Whether different phases need different quantization, how reliably the judge captures this, and the separate contributions of quantization and horizon selection.

[A4] The judge's selection is reliable: quantizing the chunks it selects keeps success above the base model, while quantizing 58 percent of chunks at random, or the chunks it rejects, lowers success and loses the speedup. Quantization and horizon selection also contribute separately, with quantization giving the speedup and horizon selection keeping the success (Table 6b in Section 4.3).

\hline \text{Which chunks are quantized} & \text{SR} \uparrow & \text{TPR} \uparrow & \text{ES} \uparrow \newline \hline \colorbox{#EDE5F7}{PaceVLA (chunks the judge selects)} & \colorbox{#EDE5F7}{\textbf{66.9}} & \colorbox{#EDE5F7}{\textbf{2.75}} & \colorbox{#EDE5F7}{\textbf{1.14}} \newline \text{Random (same probability as judge selects)} & 62.1 & 2.49 & 0.97 \newline \text{Chunks the judge rejects} & 61.9 & 2.36 & 0.96 \newline \hline \end{array}$$ The two are separate decisions because they depend on different things. Quantizing depends on whether consecutive actions can be aggregated, while following a prediction longer depends on how reliable its later part is. A chunk can be safe to quantize but uncertain toward its end, or slow but well predicted, so the judge decides the first and the router decides the second (RoboCasa-Kitchen, same checkpoint; Tables 6a and 6b in Section 4.3): $$\begin{array}{lccc} \hline \text{Variant} & \text{SR} \uparrow & \text{TPR} \uparrow & \text{ES} \uparrow \newline \hline \colorbox{#EDE5F7}{PaceVLA (ours)} & \colorbox{#EDE5F7}{\textbf{66.9}} & \colorbox{#EDE5F7}{\textbf{2.75}} & \colorbox{#EDE5F7}{\textbf{1.14}} \newline \text{No horizon selection} & 62.0 & 2.33 & 0.96 \newline \text{No quantization} & 64.1 & 2.05 & 0.99 \newline \hline \end{array}$$ --- **[Q5] Generalization under the perturbations of LIBERO-Plus.** **[A5]** Thank you for the suggestion. **PaceVLA keeps its speedup under the perturbations of LIBERO-Plus,** with successful episodes about 24 percent shorter than the base model and an overall ES of 1.13. We evaluated the LIBERO checkpoints of Table 1 zero-shot on all 10,030 LIBERO-Plus variants (7 perturbation types, the benchmark's official setting), with the threshold and the per-task gains of our LIBERO setup unchanged. The results per perturbation type are below (Table 7 in Section 4.3 and Appendix E.4). $$\begin{array}{lcccccccc} \hline & \text{Background} & \text{Camera} & \text{Language} & \text{Light} & \text{Layout} & \text{Robot Init} & \text{Noise} & \text{All} \newline \hline \text{Base (SR / TPR)} & 95.4 / 6.83 & 53.9 / 2.31 & 90.4 / 6.02 & 95.4 / 6.71 & 74.4 / 4.36 & 45.6 / 1.16 & 81.3 / 5.35 & 74.9 / 4.47 \newline \colorbox{#EDE5F7}{\textbf{PaceVLA} (SR / TPR)} & \colorbox{#EDE5F7}{96.0 / 8.63} & \colorbox{#EDE5F7}{54.7 / 3.66} & \colorbox{#EDE5F7}{88.8 / 7.71} & \colorbox{#EDE5F7}{93.5 / 8.29} & \colorbox{#EDE5F7}{75.6 / 6.07} & \colorbox{#EDE5F7}{42.4 / 1.40} & \colorbox{#EDE5F7}{82.9 / 7.44} & \colorbox{#EDE5F7}{74.5 / 5.93} \newline \colorbox{#EDE5F7}{ES} & \colorbox{#EDE5F7}{1.24} & \colorbox{#EDE5F7}{1.10} & \colorbox{#EDE5F7}{1.15} & \colorbox{#EDE5F7}{1.16} & \colorbox{#EDE5F7}{1.18} & \colorbox{#EDE5F7}{0.96} & \colorbox{#EDE5F7}{1.22} & \colorbox{#EDE5F7}{1.13} \newline \hline \end{array}$$
Reviewer_puynOpenReview ↗

Response to rebuttal

Thank you for the detailed rebuttal, which addresses some of my concerns. However, I am still not fully convinced that the paper’s methodological contribution is substantial enough to warrant acceptance at ICLR. Moreover, I share Reviewer z39S’s concerns regarding the theoretical flaws. I believe the paper requires major revision to address these issues. I therefore maintain my score of 2.

AuthorsOpenReview ↗

Reply Rebuttal Comment by Authors

We sincerely thank you for your time and careful consideration of our rebuttal. We completely understand that you may share some of Reviewer z39S's concerns. We have posted a new comment addressing that reviewer's additional points, and we would be very grateful if you could take a look and see whether it also addresses your concerns. Please do not hesitate to let us know if anything remains unclear.

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

The authors investigate how vision-language-action models transform visual and language information into robot actions. They use activation patching to study the internal representations of three VLA models, including π0.5, MolmoAct2, and MolmoBot. The authors identify a narrow group of intermediate to late layers through which the vision-language backbone provides important conditioning information to the action expert. They further analyze how visual, semantic, and geometric information is represented across different tokens and layers. Based on these observations, they propose simplifying the connections between the backbone and action expert. Experiments on MolmoBot demonstrate a 33% reduction in inference FLOPs with comparable task success, while real robot experiments on π0.5 provide additional validation.

Critical strengths

The paper presents an interesting perspective on improving VLA efficiency through mechanistic interpretability. I appreciate that the authors go beyond conventional pruning heuristics and attempt to understand which internal representations actually influence robot actions. The activation patching experiments across three different VLA architectures are particularly interesting, and the distinction between visual, semantic, and geometric information provides useful insights into the internal computation of these models. Another strength is that the authors connect their analysis to practical efficiency improvements rather than stopping at interpretability observations. The MolmoBot experiments demonstrate meaningful computational savings while maintaining comparable task success, and the real robot evaluation further supports the feasibility of simplifying the action-conditioning interface.

Critical weaknesses

My main concern is whether the causal analysis provides a sufficiently strong advantage over existing layer selection and model compression methods. Several previous works have investigated VLA representations and the importance of different computational components, including Not All Features Are Created Equal, Observing and Controlling Features in Vision-Language-Action Models, and Revisiting Parameter Redundancy in Vision-Language-Action Models. Other methods such as DeeR-VLA have also demonstrated that skipping redundant backbone layers can substantially reduce inference computation. The authors propose selecting layers based on causal influence, but the paper does not provide a systematic comparison with simpler alternatives, such as selecting late layers based on validation performance, layer-wise sensitivity, or conventional early exit. The comparison with layer 12 is useful, but I do not think a single poorly performing layer is sufficient to establish that causal analysis is necessary for identifying an efficient interface. Another concern is that the causal analysis is conducted primarily on simplified object selection scenarios with only two object locations, and the interventions are performed at the beginning of an episode. It is unclear whether the same important layers remain sufficient during more complex manipulation stages involving contact, grasping, and fine-grained control. Finally, the efficiency improvements appear strongly dependent on the model architecture. While MolmoBot achieves a 33% reduction in computation, the reduction for π0.5 is only about 5%, and the authors do not provide measured inference latency for this model. The reported reduction in real robot task completion time is interesting, but the authors acknowledge that it does not directly result from faster inference. I would therefore like to see stronger evidence that the proposed analysis leads to broadly applicable efficiency improvements. Several recent studies have already explored computational redundancy and adaptive inference in VLA models, including DeeR-VLA, MoLe-VLA, EfficientVLA, VLA-ADP, A1, and Revisiting Parameter Redundancy in Vision-Language-Action Models. While these methods adopt different strategies, they collectively demonstrate that substantial computation can be removed from VLA architectures while preserving task performance. The authors propose using causal analysis to identify important conditioning layers, but it remains unclear whether this strategy provides a meaningful advantage over existing importance estimation, dynamic layer selection, or structured pruning approaches. I would appreciate a more systematic discussion and comparison with these related methods.

Questions for authors

Could the authors compare their causally selected layers against simpler layer selection strategies under matched computational budgets? In particular, how much improvement comes from identifying the action-conditioning interface rather than simply truncating the VLM backbone? It would also be helpful to separate the computational savings from removing cross-attention connections and from early termination of backbone computation. Second, do the important conditioning layers remain consistent across different stages of manipulation, especially during contact-rich or fine-grained actions? The current causal analysis mainly considers the beginning of episodes. Third, can the authors clarify why the computational savings differ so substantially between MolmoBot and π0.5, and whether similar efficiency improvements can be achieved on MolmoAct2? Finally, the sparse architectures are evaluated after training or adaptation. Could the authors discuss whether the same layer selection remains effective without additional training and how the cost of the causal analysis compares with simpler compression procedures?

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.