acceptodds
Withdrawn from ICLR 2027

DepthPruner: Visual Token Pruning with Pluggable Depth Guidance for Efficient Vision-Language-Action Inference

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Vision-language-action (VLA) models use visual observations and language instructions to generate robot actions, but processing visual tokens is computationally expensive. Attention-based visual token pruning reduces this cost by retaining a subset of visual tokens. However, we observe that some potentially task-irrelevant tokens can receive high attention scores and compete with task-relevant tokens for a limited token budget. Depth provides geometric cues that complement attention scores for token selection. Motivated by this observation, we propose DepthPruner, a pluggable, training-free method that augments attention-based visual token pruning methods with depth guidance. It expands the candidate set provided by a base pruner and uses mean depth, residual variance, and neighborhood contrast to select tokens within the original budget. Experiments on OpenVLA-OFT across four LIBERO suites show that DepthPruner improves the average task success rate of all three base pruners at every tested token retention ratio. By replacing at most 8 of FastV's selected tokens per view, DepthPruner increases the success rate on LIBERO-Object by 9.0 percentage points at 25% token retention. Timing measurements show that depth refinement adds little inference overhead.

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

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_92WaOpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

This paper proposes DepthPruner, a training-free and plug-and-play visual token pruning method designed to improve the inference efficiency of VLA models. The authors identify a key limitation of existing attention-based pruning methods: visual tokens with high attention scores are not necessarily task-relevant, and redundant background tokens may occupy the limited token budget, causing important visual information to be discarded.

To address this issue, DepthPruner incorporates depth-based geometric information into existing attention-based pruning frameworks. Specifically, the method first expands the candidate token set generated by a base pruner and then employs three complementary geometric cues—mean depth, residual variance, and neighborhood contrast—to refine token selection while maintaining the original token budget. This approach requires neither additional training nor modifications to the underlying VLA architecture.

The authors evaluate DepthPruner using OpenVLA-OFT across four LIBERO benchmark suites, integrating it with three existing pruning methods: FastV, SparseVLM, and ADP. Experimental results demonstrate consistent improvements in average task success rates across all three pruning methods and all evaluated token retention ratios (50%, 25%, and 12.5%). Notably, DepthPruner improves FastV's success rate on LIBERO-Object by 9.0 percentage points at 25% token retention, while introducing minimal additional inference latency.

Critical strengths

The paper leverages depth information to complement attention-based token pruning, helping identify redundant visual tokens through geometric cues. Since depth information is readily available in many robotic environments, this provides a practical approach to improving token selection in VLA models.

Critical weaknesses
  1. The experiments are limited to simulated LIBERO benchmarks, with no real-world robotic evaluation. Moreover, the qualitative visualizations are insufficient to demonstrate how depth guidance improves token selection across different scenarios. This limits the evidence for the method's practical effectiveness and generalizability.

  2. The introduction argues that attention-based pruning retains redundant background tokens at the expense of task-relevant information. However, the experiments do not sufficiently investigate or quantify this issue, nor do they convincingly demonstrate that DepthPruner resolves it. This disconnect between the motivation and experimental validation weakens the paper's central claim.

Questions for authors
  1. The experiments are limited to LIBERO simulations. Can the authors provide additional evidence demonstrating that DepthPruner remains effective in robotwin2.0 or other simulation benchmark and real-world robotic settings, particularly when depth observations contain noise or estimation errors?
  2. The paper argues that attention-based pruning over-retains redundant background tokens, but the current visualizations and quantitative analyses provide limited evidence supporting this claim. Can the authors provide more extensive visualizations and quantitative comparisons showing whether DepthPruner actually removes redundant background tokens and preserves more task-relevant information?
Reviewer_65VeOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

DepthPruner is a training-free refinement module for existing VLA token pruners. It expands the base candidate set and reranks candidates using mean depth, detrended local depth variance, and neighborhood depth contrast. On LIBERO, it improves several attention-based pruners at the same token budget with little measured latency overhead.

Critical strengths
  • The method is lightweight, modular, and does not require changing model weights
  • Well-organized paper.
  • The candidate-expansion design preserves the base budget while giving geometry a chance to replace misleading high-attention tokens.
  • Complete policy-call latency is reported, rather than only the cost of the ranking operation.
Critical weaknesses
  • More strong base methods should be considered. Perhaps the proposed method can be orthogonal to some strong VLA methods (like SpecPrune-VLA) to achieve higher performance.
  • The fixed fusion weights and candidate expansion factor are hand-designed. Sensitivity to these choices and to depth scale or camera viewpoint should be reported more thoroughly.
  • The method assumes that geometric irregularity is a useful proxy for task relevance. This may fail for flat but manipulation-critical surfaces or distant objects, and the current experiments do not test such cases.
Questions for authors

See above.

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

This paper presents DepthPruner, a training-free method that uses depth geometry to refine visual token selection in VLA models. The method expands the candidate set produced by an existing pruning method and reselects tokens using three geometric cues: mean depth, residual variance, and neighborhood contrast, while maintaining the original token budget. Experiments on OpenVLA-OFT across four LIBERO suites show improvements in average task success rates for three pruning baselines at different retention ratios, with little additional inference overhead.

Critical strengths
  1. The paper addresses a relevant limitation of attention-based visual token pruning, where highly attended regions may contain redundant background information. Using depth geometry to complement attention-based selection is well motivated.
  2. DepthPruner is training-free and preserves the original token budget without requiring changes to the underlying VLA architecture. Its compatibility with three different pruning methods is a practical advantage.
  3. The experiments cover three pruning baselines and multiple token retention ratios across four LIBERO suites, with improvements in average success rates across all tested configurations. The ablation studies, qualitative visualizations, and latency measurements provide additional support for the proposed approach.
Critical weaknesses
  1. The proposed method is relatively straightforward, relying on handcrafted depth statistics to rerank an expanded set of candidate tokens. While the results are encouraging, it is not entirely clear how much of the improvement comes from the proposed geometric scoring scheme rather than the additional candidate selection flexibility. Comparisons with simpler selection strategies, such as random replacement of a few baseline-selected tokens with additional candidates or a simpler depth-based ranking, would help clarify the contribution of the proposed design. The ablation results also suggest that mean depth plays a relatively minor role in the evaluated setting, which deserves further discussion.
  2. Although DepthPruner is evaluated with three different pruning methods, all experiments are conducted using OpenVLA-OFT on simulated LIBERO tasks. The method also relies on aligned depth maps, which are readily available in simulation but may be noisy or incomplete in real-world settings. Evaluating its sensitivity to depth quality would provide useful evidence for its practical applicability.
  3. The average performance improvements are consistent across the tested configurations, but the results are less uniform across individual suites. For example, FastV on LIBERO-Goal decreases from 96.4% to 95.4% at 25% retention after adding DepthPruner. Some of the average gains are also relatively small. Further analysis of these regressions and reporting variability across evaluation runs would help assess the reliability of the improvements.
Questions for authors

N/A

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

Related papers

Discussion (0)

Sign in to comment.