acceptodds
Under review as a conference paper at ICLR 2027

RAPID-WAM: Rethinking the World–Action Interface for Compact Action Decoding

63 views

Abstract

World action models (WAMs) use the rich representations of video-pretrained world models for robot control, but existing methods typically rely on large action experts and dense world–action interactions. Through feature interventions across three representative WAM formulations, we find that selectively retaining approximately one quarter of the most action-sensitive world features preserves closed-loop performance close to full access. Based on this observation, we propose RAPID-WAM with a lightweight world–action interface that uses task-conditioned queries to read and compress representations from multiple world-model depths into compact control conditions before action decoding. RAPID-WAM outperforms the corresponding Fast-WAM baselines across action-only, inverse-dynamics, and joint world–action formulations and achieves state-of-the-art performance on multiple robot manipulation benchmarks. In the default action-only setting, RAPID-WAM reduces the world-condition representation delivered to action decoding by 11,760×, improves LIBERO-Plus success from 49.7% to 70.7%, and achieves a 4.53× speedup in end-to-end policy inference.

Primary area:
applications to robotics, autonomy, planning
Keywords:
world models, world action models, robot learning

Reviews

Among:
Shape:
1234this paper, Top 30%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.67 over 3 reviews, above 70% of the others.
Reviewer_MLmyOpenReview ↗
Rating
3(+1)
Confidence
4
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

This paper aims to improve the inference latency of World Action Models (WAMs). It challenges whether action prediction requires full access to the world representation. Motivated by the case study performed by the authors, this paper proposes using a world-action interface to select and compress the video representations from the video backbone for the action decoder. With the world-action interface, the proposed RAPID-WAM can use a lightweight action module from HDP3, leading to an improvement in inference latency compared to Fast-WAM. The more complicated and heavier world-action interface will not slow down the inference speed, as all action denoising steps can share the same compressed condition, which is processed once from the world-action interface. The empirical results demonstrate that RAPID-WAM reaches improved performance with significantly reduced inference latency.

Critical strengths
  1. This paper asks a critical question for WAMs: whether action prediction really requires access to the full world representation. This question and the paper's findings will benefit the community.

  2. This paper proposes an interesting and effective solution: decoupling the world-action connection from action prediction, which are typically in the same action module in MoT-based WAMs.

  3. The empirical performance is good, especially considering the improved inference latency.

Critical weaknesses
  1. This paper verifies the proposed design on top of Fast-WAM (action-only, IDM, Joint), which should be viewed as the most important baseline. Fast-WAM is mainly evaluated on LIBERO and Robotwin in its original paper. However, this paper only shows the results on LIBERO, LIBERO-Plus, and RoboCasa. The omission of Robotwin raises some concerns about the empirical performance of RAPID-WAM.

  2. The proposed technique seems sensitive to the choice of hyperparameters, such as the choice of readout layers (Table A6) and the number of queries (Table A7).

    • For readout layers, it seems like the trends on LIBERO and LIBERO-Plus are different. And the manual selection of readout layers seems necessary, which slightly weakens the design of the proposed world-action interface.
    • For query count, a similar observation.
    • An in-depth discussion of the ablation study will significantly strengthen the paper.
  3. The statements on the action parameters and FLOPs are not accurate (Sec. 5.3). In fact, RAPID-WAM assigns two functions of the original action expert in MoT-based WAMs, world-action connection and action generation, to the world-action interface and MLP-based action decoder. Therefore, the comparison of parameters and FLOPs between the action expert in FastWAM and the MLP-based action decoder in RAPID-WAM is a bit misleading. The counts of the world-action interface should be included. This discussion will also strengthen the paper:

    • RAPID-WAM actually shifts some capacity and compute from the action expert/decoder to the world-action interface. However, the action expert/decoder needs repeated denoising steps, while the condition from the world-action interface only needs to be computed once. Such decoupling and shifting is critical to the inference speed.

(Very) Minor weaknesses

  1. The main paper can be more self-contained. For example,
    • without searching in the appendix, I cannot know which dataset the efficiency evaluation is on.
    • without searching in the appendix, I do not know what data the diagnosis of world-feature access was conducted on.

They can be made clear by adding a one-sentence description or a reference to the appendix.

  1. This paper holds a very similar high-level idea to the following paper:

Note that I do not view this as a negative point in my assessment, as: 1). this paper proposes a different solution from them; 2). these papers are contemporaneous work and/or unpublished arXiv papers.

Questions for authors
  1. Can you provide the experimental results on Robotwin under the same settings as in FastWAM? (This is the most critical)

  2. Can you provide further discussion on the choice of readout layers and query counts?

  3. Can the world-action interface benefit the more widely used DiT-based Action expert in terms of prediction performance and inference speed? The actionDiM based on HDP3 looks elegant but is less widely used at this time. I want to understand whether the gains in prediction performance and inference speed are mainly from HDP3 or the world-action interface design.

4 comments

We thank the reviewer for recognizing our research question, the decoupling of world–action interaction from action prediction, and RAPID-WAM's control performance and inference efficiency. We address the comments below.

W1 & Q1. Can you provide experimental results on RoboTwin under the same settings as Fast-WAM?

On RoboTwin 2.0 [1], RAPID-WAM improves bimanual manipulation performance and zero-shot OOD generalization over Fast-WAM [2]. The original Fast-WAM paper trains on 27,500 demonstrations (2,500 clean + 25,000 randomized) with 64 GPUs. Given our limited computational resources, we made full use of the eight GPUs available in our lab during the rebuttal period (using four GPUs per training run) and conducted the controlled comparison using only 2,500 clean demonstrations under the clean-only setting of RoboTwin 2.0. Both models use identical training and evaluation settings:

MethodClean→CleanClean→Random
Fast-WAM66.94%1.51%
Faster-WAM†68.60%2.20%
RAPID-WAM75.58%2.52%

† Faster-WAM results are from La-JEPA [3], Table 2.

On Clean→Clean, RAPID-WAM achieves 75.58%, 8.64 percentage points above Fast-WAM, also exceeding Faster-WAM's reported clean-only result (68.60%). On Clean→Random, zero-shot OOD success rises from 1.51% to 2.52%, a 66.9% relative improvement, also surpassing Faster-WAM's 2.20%. These results demonstrate improved bimanual control and generalization under a limited training budget. We have added the results and protocol details to Sec. 5.2 and Appendix C.

W2 & Q2. Could you further discuss the choice of readout layers and query count?

We thank the reviewer for this suggestion. Our readout layer positions are not manually selected for individual tasks; instead, we use a common uniform sampling strategy. DiT4DiT [4] finds that video features at different depths play different roles in control, while DeVA [5] also uses uniform sampling to access features from multiple layers. Inspired by these works, we select five evenly spaced layers (6, 12, 18, 24, and 30) from the 30-layer backbone, without selecting layer positions separately for different datasets. Tables A5 and A7 show that all configurations perform similarly on near-saturated LIBERO (97.6%–98.8%), while LIBERO-Plus is more sensitive to the layer count: success improves from 67.7% (K=1) to 70.0% (K=3) and 70.7% (K=5), but five and 30 layers perform nearly identically (70.7% vs. 70.8%). Thus, a few uniformly sampled depths suffice without elaborate layer selection. Query count trades performance against overhead: increasing Q from 2 to 8 improves LIBERO-Plus from 68.0% to 70.7%. Q=16 further raises it to 72.7%, but LIBERO falls from 98.8% to 97.7%, and readout computation increases. We therefore use K=5 and Q=8 by default to balance benchmark performance and interface cost. We have clarified this choice in Sec. 5.4.

W3. Parameter and FLOP comparisons should include the world–action interface and explain the efficiency benefits of shifting some computation from repeated decoding to a reusable interface.

We thank the reviewer for this suggestion. The original manuscript already reports the parameter counts of the decoder and readout interface separately in Appendix E, and the Action GFLOPs in the original Table 5 (Table 6 in the revision) already include interface computation. We have now further summarized the complete action-side costs in the main text and clarified how computation is reused (default action-only setting, N=10):

MetricFast-WAMRAPID-WAMReduction
Decoder parameters1.02B9.66M—
Standalone readout interface parameters—6.01M—
Total action-side parameters (incl. interface)1.02B15.68M65.1×
World–action interaction/interface GFLOPs257.261.07240.3×
Full action-stage GFLOPs (incl. interface)1,055.333.93268.6×

Although Fast-WAM caches world features, its action expert repeats dense world–action interactions at each denoising step, costing 257.26 GFLOPs across ten steps. RAPID-WAM instead reads and compresses world features once per observation (0.96 GFLOPs), then reuses the compact conditions across ten steps (about 0.011 GFLOPs per step for condition injection), for a total interface cost of 1.07 GFLOPs.

This read-once, reuse-across-steps design eliminates repeated dense interactions and enables lightweight decoding. Including the 6.01M readout parameters, RAPID-WAM reduces total action-side parameters by 65.1× and full action-stage FLOPs by 268.6×, with a 4.53× end-to-end inference speedup. We have clarified reuse in Sec. 5.3, added interface parameters to Table 6 and totals for all three formulations to its caption, and added the default model's total parameter count to Appendix E.

Q3. Can the world–action interface improve the performance and inference speed of a DiT action expert?

Yes. Our additional experiments show that RAPID's world–action interface also substantially improves control performance and inference efficiency with ActionDiT. We have combined RAPID-WAM's world–action interface with ActionDiT to construct RAPID-WAM-DiT and evaluated its control performance and inference efficiency.

Model variantWorld–action interfaceAction decoderLIBEROLIBERO-PlusTotal policy latency (speedup)
Fast-WAMNative dense MoTActionDiT97.6%49.7%384.6 ms (1.00×)
RAPID-WAM-DiTRAPID learned readoutActionDiT98.3%78.0%154.0 ms (2.50×)
RAPID-WAMRAPID learned readoutActionDiM98.8%70.7%84.8 ms (4.53×)

Compared with Fast-WAM, RAPID-WAM-DiT improves LIBERO-Plus success by 28.3 percentage points, while achieving a 2.5× end-to-end inference speedup. Notably, its success rate even exceeds that of the default RAPID-WAM (78.0% vs. 70.7%), indicating that the benefits of the RAPID interface do not depend on the lightweight ActionDiM [6] decoder.

We have added the RAPID-WAM-DiT results to Sec. 5.4 and provided the corresponding implementation details in the appendices.

Minor W1. The main text should specify the datasets used for efficiency evaluation and world-feature access diagnosis.

We thank the reviewer for pointing this out. Appendices A and E of the original manuscript already specify that both experiments are conducted on LIBERO. We have now also made this explicit in Sec. 3 / Fig. 2 and Sec. 5.3 / Table 6.

Minor W2. Please include Faster-WAM and Light-WAM in the related-work discussion.

We thank the reviewer for recommending these two valuable concurrent works. Faster-WAM [7] improves inference efficiency through cross-layer K/V fusion, while Light-WAM [8] combines cross-layer query pooling with lightweight modeling to reduce computational cost. Whereas these methods primarily focus on computational structure, RAPID-WAM further examines the more fundamental question of how much information action prediction needs to read from a world model. Through feature interventions, we find that a small subset of key world features is sufficient for effective control, motivating our compact, reusable world–action interface. This design has been validated across multiple WAM formulations and action decoders, achieving 98.8% success on LIBERO (98.5% for Faster-WAM and 97.2% for Light-WAM) and a 4.53× policy speedup (3.2× reported by Faster-WAM), while maintaining control performance on long-horizon and complex tasks. We have added the relevant discussion to Sec. 2.

References

[1] Tianxing Chen et al. RoboTwin 2.0: A Scalable Data Generator and Benchmark with Strong Domain Randomization for Robust Bimanual Robotic Manipulation arXiv preprint arXiv:2506.18088, 2025.

[2] Tianyuan Yuan et al. Fast-WAM: Do World Action Models Need Test-time Future Imagination? arXiv preprint arXiv:2603.16666v2, 2026.

[3] Mateo Clémente et al. La-JEPA: Compact Robot Policies with Latent World Modeling arXiv preprint arXiv:2611.08156, 2026.

[4] Teli Ma et al. DiT4DiT: Jointly Modeling Video Dynamics and Actions for Generalizable Robot Control arXiv preprint arXiv:2603.10448, 2026.

[5] Mengqi Zhang et al. DeVA: Decoupled Video-Action Model with physical guidance for robot policy learning arXiv preprint arXiv:2607.24159, 2026.

[6] Jinhao Zhang et al. Hyper-DP3: Frequency-Aware Right-Sizing of 3D Diffusion Policies for Visuomotor Control arXiv preprint arXiv:2605.01581v4, 2026.

[7] Liheng Ma et al. Faster-WAM: Do World Action Models Need Deep Action Modules? arXiv preprint arXiv:2608.02365v1, 2026.

[8] Ziang Li et al. Light-WAM: Efficient World Action Models with State-Fusion Action Decoding arXiv preprint arXiv:2606.08242, 2026.

Thanks for the rebuttal.

I believe the authors have addressed most of my concerns.

One remaining point is that FastWAM reports the results on full Robotwin training (Clean+Random training --> Clean+Random eval) . I believe the reported results aligning with the most important baseline in the paper will greatly strengthen the paper.

However, regarding the current rebuttal from the authors, I will raise my score accordingly.

We sincerely thank the reviewer for carefully considering our rebuttal and for the positive reassessment. We are glad that our additional experiments and clarifications have addressed most of your concerns. We greatly appreciate your constructive feedback, which has helped us further strengthen the paper.

Reviewer_iToaOpenReview ↗
Rating
3
Confidence
3
Informativeness
3
Presentation
3
Soundness
3
Read review
Paper summary

WAMs expose rich world representations to action prediction through dense interactions and large action experts. The paper first diagnoses this access: replacing world features one position at a time shows action prediction depends highly unevenly on them, and keeping about a quarter of the most action-sensitive positions preserves near-full closed-loop success across action-only, IDM, and joint formulations (95.0/98.0/97.5 versus 96.0/96.5/99.5 full access), while random quarters lose 4 to 21 points. On this observation the paper builds RAPID-WAM: task-conditioned queries read five world-model depths and compress them into compact conditions (1,536 scalars versus 18M delivered to the action decoder), decoded by a 9.66M ActionDiM instead of a 1.02B expert. Results: LIBERO 98.8 versus 97.6, LIBERO-Plus 70.7 versus 49.7, RoboCasa +13 points, real-world +8.75, action-stage FLOPs down 268x, end-to-end latency down 4.53x. Reporting is careful: baseline provenance is documented per table, and the local Fast-WAM reproduction (45.6 on LIBERO-Plus) is disclosed beside the published 49.7.

Critical strengths
  • The diagnostic and mechanism is structured scientifically. The intervention protocol (spatial-mean replacement of one position's K/V across all layers, action RMSE as sensitivity, paired closed-loop episodes with hierarchical bootstrap intervals) isolates how much each world feature matters, and the retention control proves the unevenness is real: a sensitivity-selected quarter holds 95.0/98.0/97.5 while random quarters collapse to 74.5/86.5/93.5. The loop is then closed on the method side: interface attention correlates with independently measured action sensitivity, and high-attention retention matches full access (99.0 versus 93.5 random). Each claim about the mechanism is checked with its own experiment, which is why the interface design is credible rather than lucky.
Critical weaknesses
  • The OOD gain has no mechanism and no capacity-matched control. The diagnostic shows world features are redundant, which explains why compression can be free, not why it gains 21 points out of distribution. The obvious alternative account, that a 9.66M decoder simply overfits less than a 1.02B expert, is never ruled out: there is no Fast-WAM-small baseline and no random-masking-trained baseline, so "the interface selects control-relevant information" and "a small decoder regularizes" are confounded. The paper's own table further complicates the headline: Fast-WAM-IDM beats the default RAPID-WAM overall (71.4 versus 70.7) and wins language, background, and layout outright, and the IDM and Joint gains are only +0.6 and +4.2, so the robustness story lives almost entirely in the action-only formulation.
Questions for authors
  • Can you run a capacity-matched control, Fast-WAM with a ~10M action expert or Fast-WAM trained with random 25% world-feature masking, on LIBERO and LIBERO-Plus? This separates "the interface selects better information" from "a small decoder regularizes," which are currently confounded, and it would tell us whether the OOD gain is a mechanism or a capacity effect. If the control recovers most of the 21 points, the contribution narrows to efficiency; if it does not, the interface claim is much stronger.
4 comments

We thank the reviewer for recognizing our feature intervention analysis, mechanistic validation, and world–action interface design. To address the two concerns, we have conducted additional controlled experiments and analyzed robustness and efficiency across different WAM formulations.

W1.1 & Q1. Can Fast-WAM controls with matched capacity or 25% random world-feature masking during training clarify the source of the OOD gains?

RAPID-WAM's OOD robustness gains primarily stem from the learned world–action interface's effective selection and compression of control-relevant world information. We have evaluated two additional control variants under the same training and evaluation settings. Fast-WAM-small retains the native dense world–action interactions while reducing the action expert to 15.59M parameters, closely matching RAPID-WAM's total action-side parameter count (15.68M), including its readout interface and decoder. Fast-WAM-Mask25 randomly masks access to 25% of world features during training.

Model variantWorld–action interfaceTotal action-side parametersLIBEROLIBERO-Plus
Fast-WAMNative dense MoT1.02B97.6%49.7%
Fast-WAM-smallNative dense MoT15.59M97.1%58.7%
Fast-WAM-Mask2525% random masking during training1.02B97.3%56.2%
RAPID-WAMRAPID learned readout15.68M98.8%70.7%

RAPID-WAM achieves 70.7% on LIBERO-Plus, outperforming the parameter-matched Fast-WAM-small by 12.0 percentage points and Fast-WAM-Mask25 by 14.5 percentage points. In particular, it surpasses Fast-WAM-small by 42.8 and 24.0 percentage points under camera-viewpoint and sensor-noise perturbations, respectively. Together with the attention and feature-retention analyses in Fig. 4(c,d), these results further support the role of the learned readout in effectively selecting and compressing control-relevant world features before decoding.

We have included the new controlled experiments in Sec. 5.4, with implementation details and per-category results provided in Appendices B–D.

W1.2. Are the robustness gains mainly concentrated in the action-only formulation?

Selective world-feature readout in RAPID-WAM improves OOD robustness and substantially accelerates action generation across all three WAM formulations, action-only, IDM, and Joint.

FormulationFast-WAMRAPID-WAMLIBERO-Plus ΔCamera viewpoint ΔSensor noise ΔRoboCasa-24 ΔAction-generation speedup
Action-only49.7%70.7%+21.0+47.7+40.2+13.58.73×
IDM71.4%72.0%+0.6+20.0+17.5+13.18.70×
Joint68.1%72.3%+4.2+27.5+19.6+12.24.69×

When comparing within the same formulation, although the overall LIBERO-Plus gains for IDM and Joint are relatively modest, RAPID-WAM improves performance under camera-viewpoint perturbations by 20.0 and 27.5 percentage points, and under sensor-noise perturbations by 17.5 and 19.6 percentage points, respectively. On the more challenging RoboCasa-24 benchmark, all three formulations consistently achieve improvements of 12.2 to 13.5 percentage points, while IDM and Joint also deliver action-generation speedups of 8.70× and 4.69×, respectively.

Comparisons across formulations also demonstrate RAPID-WAM's efficiency advantage. Fast-WAM-IDM achieves 71.4% success on LIBERO-Plus by generating future video, with a policy latency of 785.9 ms. The default RAPID-WAM does not need to generate future frames and achieves a comparable success rate of 70.7% with only 84.8 ms latency, delivering a 9.3× end-to-end speedup over Fast-WAM-IDM. It also outperforms Fast-WAM-IDM by 18.0 and 16.3 percentage points under camera-viewpoint and sensor-noise perturbations, respectively.

AuthorsOpenReview ↗

Sorry to bother you

Dear Reviewer iToa,

As the discussion period is coming to a close, we wanted to check back to see whether you have any remaining questions about our response. We would be happy to clarify further and would be grateful for any additional feedback you may have. We really appreciate your time and effort during the review and rebuttal process.

Thank you very much, and we look forward to hearing from you!

Best regards,
Authors of Paper 40402

The rebuttal experiment showed ablation that the paper's contributions were not purely introduced by reduced action network size or general regularization, the method's effectiveness seems real. Despite the authors' efforts on showing effectiveness, this paper fundamentally does seem incremental on the novelty side, since it is mainly a marginal architectural adjustment, which itself is not novel, although one can argue that applying that to this task setting is novel. I choose to maintain my score.

Thank you for your careful review and for recognizing the effectiveness of our method. The controls you suggested, together with our feature-intervention and closed-loop analyses, further support the important role of the learned world–action interface in the robustness gains. We appreciate your time and constructive feedback throughout the discussion.

Reviewer_HQa3OpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes RAPID-WAM, a compact world-action interface that selectively reads and compresses world-model representations to enable efficient action decoding while maintaining or improving robotic control performance.

Critical strengths

This paper reframes WAMs from the perspective of the world–action interface and reveals that action prediction depends highly unevenly on world features.

This paper introduces a task-conditioned compact readout mechanism that compresses multi-level world representations and enables lightweight action decoding.

This paper demonstrates improved performance and inference efficiency across three WAM formulations and multiple robotic benchmarks, showing strong generality.

Critical weaknesses

The paper simultaneously introduces a compact world–action interface and a new lightweight ActionDiM decoder, but does not isolate the interface’s effect while holding the decoder fixed, making it unclear whether the performance gains can be attributed to the proposed interface design.

Questions for authors

See the Weaknesses.

5 comments

We thank the reviewer for recognizing the value of studying WAMs from the perspective of the world–action interface, as well as the generality of RAPID-WAM across different WAM formulations and tasks.

W1 & Q1. It is unclear whether the performance gains can be attributed to the proposed interface design.

Our additional experiments further validate the important contribution of the world–action interface itself to performance improvements. We have combined RAPID-WAM's world–action interface with an ActionDiT decoder to construct RAPID-WAM-DiT and evaluated its control performance and inference efficiency.

Model variantWorld–action interfaceAction decoderLIBEROLIBERO-Plus
Fast-WAMNative dense MoTActionDiT97.6%49.7%
RAPID-WAM-DiTRAPID learned readoutActionDiT98.3%78.0%
RAPID-WAMRAPID learned readoutActionDiM98.8%70.7%

With both models using ActionDiT decoders, RAPID-WAM-DiT increases LIBERO-Plus success from 49.7% to 78.0%, an improvement of 28.3 percentage points, while achieving a 2.50× end-to-end inference speedup. Notably, this result even exceeds that of the default RAPID-WAM with ActionDiM (70.7%), indicating that the interface's benefits do not depend on a lightweight decoder.

We have added the RAPID-WAM-DiT results to Sec. 5.4 and provided the corresponding comparisons and implementation details in the appendices.

AuthorsOpenReview ↗

Sorry to bother you

Dear Reviewer HQa3,

As the discussion period is coming to a close, we wanted to check back to see whether you have any remaining questions about our response. We would be happy to clarify further and would be grateful for any additional feedback you may have. We really appreciate your time and effort during the review and rebuttal process.

Thank you very much, and we look forward to hearing from you!

Best regards,
Authors of Paper 40402

Reviewer_HQa3OpenReview ↗

Response to the rebuttal

Thanks for the additional experiments. Some uncertainty remains about how closely the decoders are matched. Please clarify whether decoder depth, width, and parameter count are unchanged, and which conditioning components differ. This would help assess how cleanly the comparison isolates the proposed interface’s contribution.

Comments on the paper

1 comment
AuthorsOpenReview ↗

General Response

We sincerely thank the AC and all reviewers for their time and constructive feedback throughout the review process.

All three reviewers recognized RAPID-WAM's research perspective of focusing on the world–action interface: using a learnable interface to select and compress control-relevant world information, enabling lightweight action decoders to improve both performance and efficiency across different WAM formulations and benchmarks. Reviewer iToa acknowledged our additional capacity-matched and random-masking control experiments, noting that these results further validated the effectiveness of our method, and maintained a positive assessment of the paper. Reviewer MLmy recognized the value of this research question to the community, as well as our additional RoboTwin experiments, mechanism analysis, and ActionDiT validation, and provided a positive assessment of the work.

Reviewer HQa3 also recognized our research perspective, task-conditioned readout, and generalizability across WAM formulations. The reviewer's primary concern was whether the performance gains could indeed be attributed to the interface itself. To address this concern, we conducted an additional controlled experiment: RAPID-WAM-DiT-L retains the original 1.02B-parameter ActionDiT from Fast-WAM, along with its temporal modulation, language, and robot-state conditioning pathways, and replaces only the world–action interface. This change alone improves the LIBERO-Plus success rate from 49.7% to 75.7% (+26.0 percentage points). This result demonstrates that the proposed interface itself delivers substantial performance gains even when the original action decoder is kept unchanged, directly addressing the reviewer's central concern.

We once again thank the AC and all reviewers for their thoughtful feedback, which has guided us toward more targeted experimental validations and helped us clarify and strengthen both the contributions and the supporting evidence of our paper.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.