acceptodds
Under review as a conference paper at ICLR 2027

FAVP: Future-Aware Vision Token Pruning for Unified World Action Models

Abstract

Vision token pruning can accelerate inference in unified World Action Models (WAMs), yet existing approaches primarily use semantic, temporal, or action-related cues without explicitly exploiting the future-prediction knowledge learned during world modeling. To bridge this gap, we propose **FAVP**, a training-free **F**uture-**A**ware **V**ision Token **P**runing framework that reuses this capability to guide vision token selection during action inference. FAVP introduces three modules: (1) **Future-Prediction Query Augmentation** to efficiently extract future-aware visual information without full WM inference, (2) **Future-Aware Attention Masking** to preserve the original action-generation process and to block the queries from directly accessing the original non-visual tokens, and (3) **Action-Future Token Selection** to jointly exploit action-aware and future-aware importance to prune redundant and unimportant tokens. Extensive experiments across WAM backbones (i.e., RynnVLA-002 and UniVLA) in both simulation and real-world manipulation tasks demonstrate that FAVP achieves up to a 1.63× inference speedup on simulation benchmarks and a 1.67× speedup in real-world manipulation, with relatively small reductions in success rate.

Primary area:
applications to robotics, autonomy, planning
Keywords:
Vision-Language-Action, World Action Models, Vision Token Pruning, 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_KUg9OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper introduces a training-free visual token pruning method to speed up World Action Models (WAMs). It has three main ideas: (1) Future-Prediction Queries, which help identify visual tokens useful for predicting future states; (2) Future-Aware Attention Masking, which keeps the added queries from interfering with the original action generation; and (3) Action-Future Token Selection, which combines action-related and future-related importance scores to select the most useful visual tokens. The method is tested on UniVLA and RynnVLA-002 across LIBERO, SimplerEnv-WidowX, and three real-world robot tasks, achieving up to 1.63× speedup in simulation and 1.67× on real robots.

Critical strengths
  1. The authors cleverly exploit the intrinsic characteristics of the WAM architecture, using the future-state prediction capability that the WAM has acquired during training to perform visual token pruning.

  2. The authors design a structured attention mask that ensures the consistency of the action branch's forward propagation, thereby keeping the original action branch unaffected by the newly added queries. This allows the method to be integrated into existing WAMs as a training-free plugin.

Critical weaknesses
  1. The reported gains over VLA-Pruner are 0.5–1.1 pp across all benchmarks, which is within run-to-run evaluation noise, and no seeds or variance are reported. Without significance evidence, the claimed advantage over the strongest baseline remains unconvincing.

  2. As a training-free method, its deployment hinges on hyperparameters. The paper only studies sensitivity to the pruning layer (Table 7); sensitivity to the pruning ratio and other key settings is not reported, although the per-suite tuned ratios in A.3 suggest the results depend on them.

Questions for authors
  1. Could the authors report 3-seed LIBERO results on at least one backbone to show whether the 0.5–1.1 pp margins over VLA-Pruner are statistically meaningful?
  2. As only the pruning layer is currently studied, could the authors add sensitivity analysis on the pruning ratio, future queries, and trigger-free settings on at least one benchmark?

I will consider raising my score if my questions are answered satisfactorily.

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

This paper proposes FAVP, which is a training-free vision token pruning framework for unified World Action Models (WAMs). Existing pruning approaches usually emphasise semantic, temporal or action-related signals and overlook visual information, which can be crucial for future state prediction and action generation. FAVP addresses this weakness by focusing on vision-related pruning using the WAM's future-predictive knowledge acquired during world-model training to guide token selection.

The framework consists of three components: (1) Future prediction query augmentation: it utilises the WAM's future prediction queries to provide future-aware information for action generation (2) Future-aware attention masking: it applies an attention mask so that the future prediction queries can only access the visual context (3) Action-Future Token Selection: it selects the most important tokens at the intersection of action and future importance, then supplements them with the most diverse tokens from their union. All remaining tokens are discarded after the selected pruning transformer layer.

The method is evaluated on RynnVLA-002 and UniVLA using LIBERO, SimplerEnv, and real-world robotic manipulation tasks. The reported results demonstrate inference speedups of up to 1.63× in simulation and 1.67× in real-world experiments, with relatively small reductions in average task success rates.

Critical strengths
  1. Effective exploitation of future-predictive knowledge for efficient inference

The paper presents an interesting approach that reuses future-prediction queries learned during WAM training for vision token selection. The idea of integrating both action-aware and future-aware information is well motivated, as their overlap highlights the most relevant tokens through Top-K selection. For the retained set, it is also sensible to select the most diverse tokens that are dissimilar to the overlapped set.

  1. Promising efficiency-performance trade-off across multiple backbones

FAVP achieves speedups on LIBERO, SimplerEnv with both UniVLA and RynnVLA-002 and real-world manipulation tasks. It achieves inference speedups of up to 1.63× in simulation and 1.67× in the real world, with only minor reductions in success rate of 0.2 and 0.6 percentage points respectively. These drops are small relative to the gains in inference speed. The ablation studies also provide evidence supporting the contribution of each of the three proposed components.

Critical weaknesses
  1. Limited scope of real-world evaluation

Although FAVP is evaluated on LIBERO and SimplerEnv, its real-world validation is limited to three tabletop pick-and-place tasks on a single robotic platform. These tasks provide limited evidence of the method's generalisability to more challenging real-world manipulation scenarios, including precision-sensitive and contact-rich operations (e.g. cable insertion and assembly) or long-horizon execution. While the experiments introduce distractor objects, robustness under more challenging visual conditions like occlusions and substantial scene variations, remains insufficiently explored. A broader real-world evaluation would strengthen the evidence for FAVP's practical applicability and robustness.

  1. Insufficient analysis of baseline comparison fairness and efficiency-performance trade-offs

Although FAVP achieves promising inference speedups while largely preserving task success rates, the fairness and consistency of its comparisons with existing pruning methods could be better established.

Appendix A.3 states that pruning ratios between 30% and 50% were evaluated, but the complete results and specific criteria for selecting the final configurations are not reported. For instance, VLA-Pruner is evaluated at a pruning ratio of 0.45, whereas FAVP uses 0.50. While different methods may have different optimal pruning ratios, the original VLA-Pruner paper evaluates multiple pruning budgets (50%, 75%, and 87.5%) and reports performance across these settings. The official GitHub repository also sets 50% as the default pruning ratio. Providing success-rate versus latency curves under matched pruning ratios or computational budgets would strengthen the evidence that FAVP consistently offers a superior efficiency–performance trade-off.

Furthermore, the real-world evaluation only compares FAVP against the unpruned backbone without including competing pruning methods. This limits the evidence for its relative advantages in real-world deployment.

Questions for authors

Q1. Performance preservation on challenging manipulation tasks

Since FAVP aims to reduce inference latency while preserving task performance, could the authors discuss whether its efficiency–performance trade-off is expected to remain stable on precision-sensitive or contact-rich manipulation tasks? In particular, could pruning fine-grained visual tokens lead to greater performance degradation relative to the unpruned backbone in such scenarios, due to less accurate or less efficient trajectories towards the goal? Even if per-step inference is faster, the overall task completion time could become longer.

Q2. Pruning-ratio selection and comparison fairness

Appendix A.3 states that pruning ratios between 30% and 50% were evaluated, but the complete results and specific selection criteria are not reported. Why was a pruning ratio of 0.45 selected for VLA-Pruner, compared with 0.50 for FAVP? Were these ratios tuned independently for each backbone? Could the authors clarify whether each method was evaluated at its best-performing setting, and whether the comparison is fair across backbones and computational budgets?

Q3. Real-world efficiency and competing baselines

Table 3 reports a 1.67× inference speedup over the unpruned RynnVLA-002 model but does not include competing pruning methods. Could the authors clarify whether comparisons with other pruning methods were conducted on the real robot?

Q4. Clarification of the architecture diagram

Although Figure 2 and its caption describe the augmentation and removal of future-prediction queries, their integration into the main Transformer pipeline could be illustrated more explicitly. In particular, could the authors indicate that these queries are reused from the pretrained WAM, show the augmented token sequence entering the early Transformer layers, and make their removal after the pruning layer more visually explicit? This would help readers follow the complete information flow without relying on the accompanying text.

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

FAVP is a training-free visual token pruning method that combines action-related and future-prediction attention signals. It reuses pretrained future-prediction interfaces without generating future observations. Experiments on RynnVLA-002 and UniVLA demonstrate inference speedups in simulation and real-world manipulation.

Critical strengths

The method makes practical use of pretrained future-prediction capabilities without additional training. Evaluation on two backbones and real-world tasks supports its potential for inference acceleration. The reported latency measurements also include the additional query and token-selection overhead.

Critical weaknesses
  1. The experimental comparison needs stronger support. SimplerEnv uses only 96 evaluation episodes. The improvement over VLA-Pruner, from 74.0% to 75.0%, corresponds to one additional successful episode, making its reliability difficult to assess without repeated runs or confidence intervals. Methods are also compared at different pruning ratios selected using success rate and latency, but the selection procedure is unclear. Success-rate–latency Pareto curves, particularly comparisons at matched success rates or allowable performance drops, would provide stronger evidence of an efficiency advantage.

  2. The practical applicability beyond the evaluated backbones remains unclear. The paper explicitly targets unified WAMs with a shared backbone. However, other architectures, such as OpenWAM-α, use separate video and action Transformers coupled through shared attention. It is unclear whether FAVP’s query construction and pruning procedure can be applied to such designs without substantial modification. Clarifying this boundary would help assess how broadly useful the method is beyond RynnVLA-002 and UniVLA.

Questions for authors
  1. Can you provide repeated-run results or confidence intervals for the SimplerEnv comparison?
  2. Were pruning configurations selected on a separate validation set? Do the efficiency gains persist when methods are compared at matched success rates?
  3. Can FAVP be adapted to OpenWAM-α’s separate video and action Transformers? What modifications and additional computation would be required? If adaptation is not straightforward, what are the limiting factors?
2 comments

Thank you for the additional experiments. The LIBERO comparison at comparable success rates and the OpenWAM-α results address my main concerns about inference efficiency and applicability beyond shared-backbone models. The gain on SimplerEnv remains modest. Based on the new evidence, I am raising my rating from 2 to 3.

We thank the reviewer for the careful re-evaluation and for raising the rating to 3. We are glad that the LIBERO comparison at matched success rates and the OpenWAM-α results clarified the two concerns about inference efficiency and applicability beyond shared-backbone models.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.