acceptodds
Under review as a conference paper at ICLR 2027

DART-VLA: Dynamic Action-Relevance Token Pruning for Efficient Vision-Language-Action Models

Abstract

Vision-Language-Action (VLA) models map multimodal observations and instructions to continuous control actions. However, the large number of visual tokens imposes substantial computational overhead on the language-model backbone, hindering real-time closed-loop inference. Existing VLA pruning methods primarily rely on semantic attention or static importance estimates, leaving the potential of action signals for guiding token-level pruning largely underexplored. Our key insight is that a visual token can be better valued by its influence on the predicted action, as this directly reflects its contribution to the decision process. Based on this insight, we propose DART-VLA, a dynamic action-relevance token pruning framework for efficient vision-language-action models. DART-VLA backpropagates gradients from the action outputs to estimate each visual token's influence on action prediction and prunes tokens accordingly. DART-VLA first adapts the overall retention budget to the current motion state. Considering the distinct functional roles of the main and wrist views, we further allocate the token budget between them based on end-effector motion and independently retain the tokens most relevant to action prediction within each view. DART-VLA achieves a 97.40% average task success rate while delivering a 1.26\(\times\) inference speedup. Moreover, at a comparable inference speedup, DART-VLA outperforms the strongest pruning baseline by 1.90 % in average task success.

Primary area:
applications to robotics, autonomy, planning
Keywords:
vision-language-action models, visual token pruning, efficient inference, robotic manipulation, action-guided token selection

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

This paper proposes DART-VLA, a training-free framework for visual token pruning in vision-language-action models. At periodic full-scoring queries, the method constructs an L1 objective from the predicted normalized action chunk and uses absolute gradient-times-attention scores from the final layer to rank visual tokens. The selected indices are reused in subsequent queries to select freshly computed visual embeddings. Action-Relative Dynamic Budgeting adjusts the retention budget according to planned motion relative to episode history, while Action-Motion-Conditioned View Allocation distributes that budget between the main and wrist cameras. The paper evaluates the method on OpenVLA-OFT and OpenPI’s π0.5 across four LIBERO suites, together with experiments on a single-arm ALOHA platform. It reports average LIBERO success rates of 97.40% and 97.35%, with inference speedups of 1.26× and 1.15×, respectively.

Critical strengths

The paper provides empirical evaluation across multiple settings. DART-VLA is evaluated on OpenVLA-OFT and π0.5 across four LIBERO suites, with additional real-robot experiments. The reported results suggest that the framework can reduce inference latency while largely preserving task success. In particular, on OpenVLA-OFT it achieves 97.40% success compared with 95.50% for SP-VLA at similar reported latencies of 54.06 and 53.48 ms. These results provide encouraging evidence of practical utility for efficient VLA inference.

Critical weaknesses
  1. Efficiency accounting needs clarification. The paper describes periodic full-scoring queries, but does not clearly explain how their scoring overhead is included in the reported latency and FLOPs. This makes the net efficiency gains difficult to assess.

  2. The benefit of dynamic budgeting needs stronger support. Fixed 50% pruning in Table 2(d) achieves nearly the same success rate as the dynamic method at comparable latency. Without uncertainty estimates, the reliability of the small reported improvement remains unclear.

  3. The methodological contribution appears incremental. VLA-Pruner [1] already incorporates action relevance into token selection, while SP-VLA [2] and ADP [3] adapt pruning to motion dynamics. DART-VLA introduces gradient-weighted scoring and specific budget and view-allocation rules, but these refinements represent a relatively limited conceptual advance over existing action-aware pruning methods.

[1] Ziyan Liu et al. Bridging the Semantic-Action Gap in Visual Token Pruning for Efficient VLA Inference. arXiv:2511.16449, 2025.

[2] Ye Li et al. SP-VLA: A Joint Model Scheduling and Token Pruning Approach for VLA Model Acceleration. ICLR, 2026.

[3] Xiaohuan Pei et al. Action-aware Dynamic Pruning for Efficient Vision-Language-Action Manipulation. ICLR, 2026.

Questions for authors
  1. Accounting for scoring overhead. Do the reported latency and FLOPs include the full cost of scoring queries, including gradient computation and token selection? Please provide a cost breakdown and explain how the reported speedup is obtained when all costs are averaged across full-scoring and pruned queries.

  2. Evidence for dynamic budgeting. Fixed 50% pruning achieves 97.20% success versus 97.40% for the dynamic method at essentially identical latency (Table 2). What evidence supports a reliable advantage from dynamic budgeting? Please quantify the uncertainty in this comparison.

  3. Advance over prior work. What is the main new insight provided by Grad×Attention and the budget/view-allocation rules relative to VLA-Pruner, SP-VLA, and ADP? Please explain how the existing scoring and component ablations substantiate that advance.

Reviewer_NpcQOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

DART-VLA prunes visual tokens in VLA models without retraining. At periodic full-scoring queries, it backpropagates the L1 norm of the predicted action chunk to the last-layer action-to-vision attention and keeps the tokens with the highest |gradient × attention|. AR-DB sets the retention ratio from planned motion, and AMC-VA splits the budget between the main and wrist cameras. Earlier VLA pruners mainly use action signals to set budgets; this method ranks tokens by action sensitivity, using a gradient-weighted attention rule close to Chefer et al. (ICCV 2021), which is not cited. Experiments cover LIBERO with OpenVLA-OFT and π0.5 against VLA-Cache, ADP and SP-VLA, plus four ALOHA tasks.

Critical strengths

The scoring rule is shown to matter. In Table 2(b), on π0.5 with the same AR-DB budget rule (70–75% retention) for every criterion, Grad×Attention reaches 97.35% against 95.05% for attention-only and 42.7% for random selection. The gap over attention-only is well above sampling noise. The margin over gradient-only (96.85%) is within noise, and the test uses π0.5 only. On the same backbone, DART-VLA beats SP-VLA by 2.05 points at the same 1.15× speedup, also outside noise. Together these support the central claim that tokens should be ranked by their effect on the action. This is the paper's main strength, helped by the method working on two backbones without retraining.

Critical weaknesses

The "highest success" gains fall within noise, and on OpenVLA-OFT the latency-matched baseline row does not fit the protocol. DART-VLA beats the unpruned model and VLA-Cache by 0.40–0.55 points. With 2,000 rollouts the standard error of such a difference is about 0.5 points, and no seeds are reported. The abstract calls SP-VLA the "strongest" baseline, but on OFT it is the lowest-scoring one; it is used because it is latency-matched (53.48 vs 54.06 ms). SP-VLA's OFT values of 96.30% and 93.70% cannot arise from 500 rollouts per suite, even though §4.1 says all methods share rollout counts. On π0.5 the equal-speed gain over SP-VLA (2.05 points at 1.15×) is outside noise, so the claim rests mainly on that backbone; the OFT comparison needs the SP-VLA row explained.

The ablations do not show that the adaptive components are needed. At matched latency, fixed 50% pruning gives 97.20% at 54.05 ms against DART-VLA's 97.40% at 54.06 ms (Table 2(d)), and fixed 75% gives 97.10% at a faster 1.36×. Both gaps are within noise. Fixed 50% retention gives 96.75% at 1.18× in Table 2(a) but 97.20% at 1.25× in Table 2(d), and the +AMC-VA row (97.15%, 1.20×) matches neither, so some undescribed setting differs. AR-DB barely changes FLOPs (3.238 vs 3.240 T), and the AR-DB and AMC-VA gains (+0.45 and +0.40 points) are each within about one standard error. This undercuts contributions 2 and 3. Because the scoring idea is supported but two of the three claimed contributions are not, the paper sits just below the acceptance bar.

Questions for authors

Request to verify an existing result: Did the authors run all Table 1 baselines under 10 × 50 rollouts? If so, how were the SP-VLA values of 96.30% and 93.70% on OFT obtained?

Request for new analysis: Please report seeds and confidence intervals for Tables 1–2.

Clarification: Why do the 50% rows in Tables 2(a) and 2(d) differ? Table 2(d)'s fixed 50% row already matches DART-VLA's latency (54.05 vs 54.06 ms) within 0.2 points; what does AR-DB add beyond it? Does the 1.26× speedup include the vision encoder and the amortized backward pass at full-scoring queries?

Request for a new experiment: Please compare AMC-VA against a fixed 50/50 view split.

Reviewer_t81rOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper proposes DART-VLA, a training-free visual-token pruning framework for efficient VLA inference. It estimates token relevance using gradients of an action-conditioned objective combined with attention weights, and introduces dynamic token budgeting and motion-conditioned view allocation. The method is evaluated on OpenVLA-OFT and OpenPI's π0.5 across the four LIBERO suites and on a real-robot platform. The reported results show improved inference efficiency while largely preserving task success. However, the justification of the action-relevance objective, the distinction from existing action-aware pruning methods, and the reproducibility of the reported efficiency gains require further examination.

Critical strengths

Action-oriented token selection. DART-VLA combines action-gradient information with attention weights to estimate the relevance of visual tokens to action prediction. This provides a more directly decision-oriented criterion for token pruning than relying solely on semantic relevance or attention magnitude. The ablation results support the benefit of combining these signals.

Systematic and carefully conducted ablation studies. The paper provides a relatively comprehensive set of ablations covering individual framework components, token-scoring criteria, gradient-propagation depth, and pruning ratios. These experiments help disentangle the contributions of different design choices and provide useful insights into the trade-off between task performance and inference efficiency.

Critical weaknesses

Limited evaluation across VLA architectures and action-generation objectives. The proposed action-relevance score relies on gradients derived from a scalar objective over the predicted action chunk and the model's action-to-vision attention weights. However, the evaluation is limited to OpenVLA-OFT and π0.5, and the paper does not establish whether the approach generalizes to substantially different VLA architectures, such as autoregressive action-generation models, or to models trained with different action-prediction objectives or loss functions. Since the availability and interpretation of action gradients may depend on the action-generation paradigm and objective formulation, evaluating the method across more diverse VLA architectures and training objectives would be important for establishing its generality.

Limited benchmark diversity and insufficient evaluation of generalization under challenging conditions. The simulation evaluation focuses exclusively on LIBERO, where strong VLA policies can already achieve high success rates, potentially limiting the ability to assess the benefits and limitations of the proposed pruning method. Additional evaluation on more challenging or diverse benchmarks, such as RoboTwin 2 and RoboDojo, would provide stronger evidence of its robustness. Moreover, if the central claim is that DART-VLA preserves task-relevant visual information while removing irrelevant tokens, it would be particularly valuable to evaluate whether this property translates into improved robustness under distribution shifts, for example through LIBERO-Plus, LIBERO-Pro, or RoboTwin 2 clean-to-random generalization settings. Without such evaluations, it remains unclear whether the method offers benefits beyond maintaining performance on the evaluated benchmarks.

Questions for authors
  1. Why is the L1 norm of the normalized predicted action chunk chosen as the scalar objective for action-relevance scoring? Have alternative objectives been evaluated, and how sensitive are the selected tokens to this choice?

  2. How does DART-VLA differ technically from existing action-aware pruning methods, particularly those that use action-related signals for token selection or budget adaptation?

  3. Does the reported end-to-end latency include the full-scoring forward and backward passes, token selection, and all other runtime overheads? Could the authors report the latency of full-scoring and pruned queries separately?

  4. How many real-robot rollouts were conducted per task, and can the authors provide uncertainty estimates or repeated-trial results to support the reported success-rate improvements?

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.