acceptodds
Under review as a conference paper at ICLR 2027

BridgeWAM: Does Shared Attention Align Physical World Representations with Actions?

Abstract

World Action Models (WAMs) commonly couple video and action experts through dense, layer-wise Mixture-of-Transformers (MoT) attention, exposing action prediction to irrelevant visual features while enforcing matched expert depths. This coupling hinders selective visual-feature use and introduces redundant computation. We introduce Bridge-of-Experts (BoE), instantiated as BridgeWAM, which aggregates visual, language, and proprioceptive priors through 32 Latent Bridge Queries (LBQs) into a shared, limited latent space. Joint video–action supervision lets action gradients guide the Video Expert toward representations better suited for action prediction, while the Action Expert learns to selectively retrieve task-relevant information through cross-attention. Decoupling expert depths and reusing LBQs during action denoising yield a model with a 30-layer Video Expert and a two-layer Action Expert, achieving % success on LIBERO and % pooled success on LIBERO-Plus, demonstrating effective task execution and strong generalization with an ultralight action head. Ablations support joint representation adaptation and unified LBQ conditioning. Relative to Fast-WAM, BridgeWAM reduces single-sample inference FLOPs by % in simulation profiling and real-robot latency to ms ( speedup). Qualitative experimental results further indicate preferential attention to target objects and task destinations across perturbations. Our project page is available at https://bridgewam.github.io/Bridge-Wam-show/ .

Primary area:
applications to robotics, autonomy, planning
Keywords:
World Action Model, Embodied AI, Multimodal Representation Learning

Reviews

Among:
Shape:
1234this paper, Top 17%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.75 over 4 reviews, above 83% of the others.
Reviewer_UKQkOpenReview ↗
Rating
3
Confidence
2
Informativeness
3
Presentation
2
Soundness
3
Read review
Paper summary

This paper presents BridgeWAM, a cross-modal interaction architecture named Bridge-of-Experts (BoE) for World Action Models (WAMs). The method addresses two limitations of conventional depth-symmetric Mixture-of-Transformers (MoT) designs: dense full-image features competing for attention, and action-side parameter redundancy caused by binding the two experts to the same depth. The method has three core designs: (1) a set of compact Latent Bridge Queries that aggregate visual, language, and proprioceptive priors within a limited latent space, (2) an asymmetric attention mask that prevents causal leakage from future video-prediction features into the bridge queries, and (3) decoupling the depths of the video and action experts, using a lightweight alternating cross–self action head whose query representations are reused throughout the denoising process. The method is evaluated on the LIBERO suites, RoboTwin 2.0, and real robot-arm tasks, where it improves robustness under out-of-distribution perturbations and reduces inference latency while maintaining action execution accuracy.

Critical strengths
  1. The authors identify the limitations of symmetric MoT architectures in current WAMs. By introducing a fixed number of learnable bridge queries as an information bottleneck, the new design decouples the depth dependency between the perception backbone and the control action head, reducing the computational overhead of closed-loop denoising.

  2. The paper's attention mask design accounts for both causality and inference consistency: by restricting the bridge queries' visibility to future video features, training-time future prediction supervision can still shape the underlying representations, while test-time single-frame feature extraction remains consistent with training.

Critical weaknesses
  1. The paper does not study sensitivity to the number of LBQs. All cross-modal conditioning passes through the 32 LBQs, but no ablation over the query count M (e.g., M ∈ {8, 16, 32, 64, 128}) is reported. It is unclear whether M = 32 is a tuned optimum or one point within a robust range, and how the query count trades off between filtering background distractors and preserving contact-relevant details.

  2. The authors state in the conclusion that real-robot long-horizon tasks "indicate limited performance", but do not analyze these failures. Without failure-case attribution, it is unclear whether the 32 queries saturate in representational capacity when switching between subtask stages, or whether the action chunk drifts under long-horizon error accumulation.

Questions for authors
  1. Could you report an ablation over the number of LBQs (e.g., M = 16, 32, 64 on LIBERO-Plus), to show whether 32 queries form a stable and transferable information bottleneck?

  2. Could you describe the failure cases in the real-robot long-horizon tasks? Do the 32 LBQs exhibit interference across stages or states during long-horizon planning?

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

The paper addressed the problem of how to improve computation efficiency of WAMs. The authors propose ‘BOE’ and ‘BRIDGEWAM’, which aggregate visual, language, and proprioceptive priors through 32 Latent Bridge Queries (LBQs) into a shared, limited latent space. The method could keep the performance of WAM and dramatically reduce the inference time. The authors verified their method on different benchmarks as well as real world experiments.

Critical strengths
  1. The authors introduce ‘BOE’ and ‘BRIDGEWAM’, which construct a limited latent space that aggregates visual, language, and proprioceptive priors into a unified latent representation tailored to the Action Expert. The method reduces the parameter reduncdancy in previous WAM and decreases inference cost while deployed on real world.
  2. The authors demonstrated their advancement on several simulation benchmarks such as LIBERO, Libreo-plus, Robo-twin 2.0 and four tasks on real-world experiments. Clear demos can be found in their project pages.
Critical weaknesses
  1. Even though the authors show improvement compared with Faster-WAM by Zhao etal., there is another paper that also names Faster-WAM by Ma etal., showing a dramatic decrease on inference latency. The authors need to give a comparison with Fater-WAM by Ma etal. and prove the benefit of BRIDGEWAM.
Questions for authors

1)In Line 102, BRIDGEWAM pairs a 30-layer Video Expert with a two-layer alternating cross–self-attention Action Expert for LBQ readout, any ablation study for the choice of two layers? Can it be four or more? 2)In Line 297, the ‘built’ and ‘can maintain’ are not consistent in grammar and give ambiguity. 3)Can the method be used on VLA models with the VLM backbone?

Reviewer_1p7jOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
3
Read review
Paper summary

The paper introduces 32 Latent Bridge Queries to connect the Video Expert and a lightweight Action Expert, enabling decoupled expert depths. The idea is reasonably novel, and the experiments cover LIBERO, LIBERO-Plus, RoboTwin, and real-robot tasks. BridgeWAM replaces dense layer-wise video-action coupling with 32 Latent Bridge Queries. A 30-layer Video DiT aggregates visual, language, and proprioceptive information into this compact interface, which conditions a two-layer action head and is reused across denoising steps.

Critical strengths

A substantial efficiency gain with competitive task performance. Under the matched profiling conditions in Table 6, action computation falls from 1.043 to 0.028 TFLOPs and latency improves by 4.61 times relative to Fast-WAM.

A strong LIBERO-Plus result with informative interface variants.

Critical weaknesses

The main concerns are that the latency and speedup numbers in the abstract are inconsistent with those in Table 6 and Figure 4. There are also no sufficient ablations on the number and placement of LBQs or comparisons with strong query-bottleneck baselines. The real-robot evaluation is small and uses staged scoring, which is not fully equivalent to complete-task success. Moreover, BridgeWAM is not consistently better than Faster-WAM on LIBERO-Pro, LIBERO-Plus, or RoboTwin.

The principal representation-learning explanation is not isolated. Frozen-Video-DiT changes initialization and freezes shared projections involved in bridge processing, so its degradation cannot be assigned specifically to removing action gradients from visual learning.

Questions for authors

The idea is promising, but the authors should first clarify the efficiency measurements, provide the missing appendices, and strengthen the fairness of the baselines and ablations.

With matched initialization and optimization, what happens when action-loss updates to the Video DiT are selectively blocked while video-loss training and the intended bridge/action parameters remain trainable? Please specify gradient routing through shared projections.

Reviewer_Hm1fOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes Bridge-of-Experts (BoE), which replaces the dense, layer-wise MoT video K/V coupling in WAMs with 32 learnable Latent Bridge Queries (LBQs). The LBQs aggregate the current frame, language, and proprioceptive state inside the Video DiT, and their final states serve as the sole conditioning input to a two-layer cross/self-attention action head, reused across all denoising steps. The two experts are trained jointly with equally weighted video and action flow-matching losses. BridgeWAM (5.06B) achieves 98.3% on LIBERO and 72.2% on LIBERO-Plus, with a real-robot latency of 77.1 ms, which is 2.47x faster than Fast-WAM.

Critical strengths

The interface design is simple and the efficiency gains are practical. The action head has only two layers, FLOPs are reduced by 28.7%, real-robot latency is 77 ms, and LIBERO-Plus improves by 22.7 points over the direct baseline Fast-WAM. For closed-loop control, this performance and latency trade-off is the most valuable result of the paper.

The ablations are fairly complete. Table 4 sweeps the action-head depth, and Table 5 examines a frozen video expert, MoT readout, text and state injection paths, and Joint/IDM coupling under the same backbone. The Frozen-Video-DiT result in particular shows that backpropagating action gradients into the video expert matters.

Critical weaknesses

The central claim lacks direct validation. The method section argues that joint supervision aligns world representations with actions, but there is no ablation that keeps the LBQ interface while removing the video loss. It is therefore impossible to tell whether the gains come from video supervision or simply from a cross-attention bottleneck combined with a trainable 5B backbone. The explanation in Table 4 that deeper action heads are more prone to memorization is also only a hypothesis, and the differences across depths may lie within noise.

Statistical rigor and presentation are problematic. The paper reports no seed counts or standard deviations, yet several conclusions rely on differences of 1 to 3 points. BridgeWAM is below Faster-WAM on every aggregate benchmark and clearly behind OpenWAM on the Background and Robot dimensions of LIBERO-Plus, which the paper does not discuss. The real-robot experiments cover only 4 short-horizon tasks, the overall score is essentially tied with π0.5, and the conclusion acknowledges limited long-horizon real-robot performance. A more accurate positioning of the paper is a much faster WAM with competitive but not best success rates, and the abstract and Fig. 1 should be revised accordingly.

Questions for authors

Please add an ablation that keeps the LBQ interface but removes the video loss, and report whether LIBERO-Plus performance drops. How many seeds underlie the main tables? Please report means and standard deviations for Table 4 and Table 5. In Table 6, the video-branch FLOPs of BridgeWAM are 42% higher than those of Fast-WAM. What is the reason, and which computations are skipped at inference? The real-robot latency of Fast-WAM is 190 ms, while Table 6 reports 406 ms, a large discrepancy. Were the real-robot hardware and deployment settings identical? How many trials were run per real-robot condition? Were the baselines fine-tuned on the same data and with the same budget? Why was M = 32 chosen? Was a sweep over the number of LBQs performed? Equations 1 to 4 and the Table 3 header contain several symbol rendering errors. The Related Work should compare against bottleneck designs such as Perceiver resampler and Q-Former. If the authors can address most of my concerns, I will raise my rating to Weak Accept.

Comments on the paper

1 comment
AuthorsOpenReview ↗

Gentle Follow-up on Our Rebuttal

Dear Reviewers Hm1f, 1p7j, and DhJe,

Thank you very much for your time and effort in reviewing our submission. We sincerely appreciate your valuable comments and constructive suggestions, which have helped us improve our work.

As the author–reviewer discussion period will close in approximately 10 hours, we would be grateful if you could kindly revisit our rebuttal and let us know whether your concerns have been adequately addressed. We would be happy to provide any further clarifications if needed.

Thank you again for your time and consideration.

Best regards, The Authors

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.