acceptodds
Under review as a conference paper at ICLR 2027

Q-WAM: 4-Bit Quantization of World Action Models with Action-Subspace Protection

Abstract

World Action Models (WAMs) jointly generate video and robot actions through iterative diffusion and perform strongly in robotic manipulation. However, their prohibitive compute and memory costs pose substantial deployment challenges. Post-training quantization (PTQ) can reduce these costs, but existing PTQ methods such as smoothing and rotation are insufficient to maintain the precision of action generation. To overcome this limitation, we propose Q-WAM, a new 4-bit weight-activation quantization for WAMs that preserves the actions the model generates. Specifically, we introduce the Action Observability Gramian (AOG), which measures how much rounding errors in each weighted combination of a layer's input channels change the final action through all denoising steps. We also develop Action-Subspace Protection (ASP), which keeps the few most action-sensitive channel combinations in a tiny 16-bit branch and quantizes the remaining weights and activations to 4 bits. Finally, we aggregate the AOG-derived action mass across the layers of each expert to identify the experts that matter most for the generated action, and apply ASP only to those experts, thereby preserving action quality with minimal overhead. We evaluate Q-WAM on three WAMs, both in simulation and in real-world deployment. On the RoboTwin 2.0 benchmark, it reaches 89.6–93.0% average success rate, within 1.1 percentage points of the 16-bit models, while reducing the memory of the quantized blocks by 3.1–3.4. Our method outperforms the strongest baseline, SVDQuant, by 2.5–8.7 percentage points. In real-world deployment on a Unitree G1 humanoid and a UR3 bimanual robot, our quantized Fast-WAM and ImageWAM reach 56.0% and 76.0% average success rate, 12.8 and 17.6 percentage points above SVDQuant.

Primary area:
applications to robotics, autonomy, planning
Keywords:
World Action Models, Post-training quantization, Efficient inference, Robotic manipulation

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

This paper proposes Q-WAM, a post-training weight-and-activation quantization method for world action models. It estimates an Action Observability Gramian (AOG) through the unrolled denoising process to measure how activation perturbations affect the final action. Action-Subspace Protection (ASP) preserves the leading action-sensitive directions in a rank-32 16-bit branch while quantizing the complementary computation to 4 bits. For multi-expert models, aggregated sensitivity guides which experts receive protection. Evaluation covers three models on 50 RoboTwin 2.0 tasks under clean and randomized conditions, together with five physical-robot tasks. For success rates averaged over clean and randomized conditions, Q-WAM remains within 1.1 percentage points of bf16 and improves over SVDQuant by 2.53–8.69 points. The paper also reports 3.1–3.4× compression of targeted blocks and 1.06–1.56× speedups for the measured quantized action-token layers.

Critical strengths
  1. An informative connection between action sensitivity and precision allocation. The paper measures sensitivity through the complete denoising process and uses its dominant directions to define a small protected subspace. Figure 3a provides direct motivation: in the Fast-WAM single-layer quantization study, AOG-based scores correlate with action damage at 0.95, compared with 0.04 for local activation error. The connection from this observation to ASP and expert selection gives the method a coherent rationale. This is consequential because it identifies a useful criterion for preserving behavior when compressing iterative action-generation models.

  2. Consistent empirical gains with favorable storage costs. Table 1 reports improvements over SVDQuant across all three evaluated models, including both clean and randomized settings. Q-WAM uses fewer bits per weight on the two multi-expert models and the same BPW and reported memory on LingBot-VA, so its gains cannot be explained simply by a larger high-precision storage budget. Physical-robot results provide additional evidence of utility, with average improvements of 12.8 and 17.6 percentage points over SVDQuant. The actual 4-bit implementation also includes the transformation and protected-branch overheads in its block-level timings. Together, these results support practical value for the proposed method.

Critical weaknesses
  1. The task-level contribution of AOG-based direction selection is not fully isolated. Table 3 adds the entire ASP component, including an additional 16-bit branch, while Table 5 retains ASP and adds SVDQuant on the weight side. Appendix A.5 compares AOG and random subspaces using local residual statistics and AOG-weighted damage, but does not report the corresponding closed-loop task success. Table 1 provides useful overall performance and storage comparisons; however, a comparison that holds preprocessing, protected experts, and low-rank budget fixed while changing only the protected subspace would more directly establish the value of the central direction-selection criterion. This limits the strength of the attribution, although the overall empirical gains remain supported.

  2. The finite-probe subspace-estimation argument is incomplete for the sketch actually used. Appendix A.4 acknowledges that the cited dense Gaussian range-finder guarantee does not directly apply to its block-diagonal probe construction, but then relies on unbiasedness and the large total count NP to justify subspace recovery. These properties alone do not ensure accurate recovery: for example, if sensitivity were concentrated in one input and one token-step node, the estimate would have rank at most P regardless of how many zero-sensitivity inputs were added. Appendix C.3 provides useful stability evidence across calibration sets, but does not directly isolate error from the finite probe count. An applicable assumption or direct convergence evidence would strengthen confidence in the estimated subspace underlying ASP. This is a gap in the justification, rather than evidence that the reported estimates have failed in practice.

Questions for authors
  1. Evidence for the central attribution: Do you have task-success results that replace the AOG subspace with a random or weight-SVD subspace while keeping smoothing/rotation, protected experts, rank, and precision budget fixed? If so, please clarify the exact matching conditions. This comparison would help determine how much of the task improvement comes specifically from AOG-based direction selection, beyond the benefit of a protected low-rank branch.

  2. Clarification of estimator reliability: What assumptions make the subspace-recovery argument applicable to the block-diagonal sketch in Appendix A.4? Alternatively, with the calibration inputs fixed, can you provide evidence that the protected subspace and relevant action-error or task metrics stabilize as the probe count increases beyond P = 12? This would address uncertainty in the central sensitivity estimate without requiring a new application setting.

1 comment
AuthorsOpenReview ↗

Follow up on our response

Thank you again for your careful review, in particular the close reading of App. A.4. We wanted to kindly follow up on our response:

  • Central attribution: as you suggested, we ran the closed-loop comparison that changes only the protected subspace (random or channel-based), with smoothing, rotation, protected experts, rank (r=32r = 32) and precision held fixed. The AOG subspace accounts for the gain.
  • Estimator reliability: we stated the assumption under which the block-diagonal estimate applies, with measured block statistics on ImageWAM, and added a convergence study over the probe count (P=4,12,48P = 4, 12, 48), showing that the subspace has stabilized by P=12P = 12 (mean subspace overlap 0.9900.990 vs. 0.9950.995 at P=48P = 48).

We would welcome your thoughts on whether these address your two questions, and we are glad to clarify anything further.

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

This paper introduces Q-WAM, a mixed-precision (W4A4) post-training quantization method designed for World Action Models (WAMs). The authors employ random probes during the full denoising process to perform backpropagation, estimating a layer-wise Action-Observability Gram (AOG) matrix. Through a mechanism called Action Subspace Preservation (ASP), the most sensitive activation directions are retained within a low-rank 16-bit branch, while the remaining weights and activations are subjected to 4-bit quantization. The method is evaluated across three WAMs, the RoboTwin 2.0 benchmark, and two real-world robotic platforms, demonstrating robust performance.

Critical strengths
  1. To address the high computational and memory overhead of world-model-based action models, the authors propose a straightforward and elegant approach: they identify the activation directions most critical to the final action, preserving these in a small, high-precision branch, while applying 4-bit quantization to the remainder. The idea is both effective and interesting.
  2. Validation on real robots across different architectures and platforms demonstrates the practical value of this work. Q-WAM significantly reduces the memory footprint of target modules while maintaining strong performance on robotic tasks, offering a valuable technical pathway for alleviating the deployment burden of world-action models.
Critical weaknesses
  1. Prior work has explored precision allocation guided by action sensitivity and low rank branches in high precision. The distinction of this paper lies in its identification of dense, motion-sensitive subspaces through the full denoising process. However, whether selecting subspaces through the full denoising process outperforms simpler channel selection or local subspace methods is not clear. Tab. 3 introduces ASP while simultaneously increasing the high-precision budget, and the comparison with SVDQuant further differs in preprocessing and protection strategy, making it difficult to attribute the observed gains specifically to the proposed subspace selection.

  2. Can the protection subspace be reused across tasks? AOG relies on action responses derived from calibration data; if recalibration is required for every new task, the deployment cost would increase significantly. However, the paper does not present results for uncalibrated tasks, nor does it report the time and VRAM overhead associated with recalibration.

  3. The paper validates the speedup for the quantized action-token layers using a genuine 4-bit kernel; however, it lacks comparisons with W4A4, SVDQuant, and Q-WAM—implemented under identical conditions but without ASP—as well as data on end-to-end latency (from observed input to action output) and peak VRAM usage. Notably, the speedup reported for the Fast-WAM component is only 1.06×, leaving it unclear whether the full strategy yields a meaningful improvement in speed.

Questions for authors

As noted above, My main concerns are the following:

  1. Is there any evidence demonstrating that full denoising of the AOG is necessary?
  2. Can protected subspaces be reused across tasks? How do they perform on tasks not included in the calibration process, and how many resources are required for recalibration?
  3. Compared to a W4A4 configuration without ASP, how do the latency and peak VRAM usage change when the full strategy is applied?

The issues explored in this paper are quite interesting; however, some of the methods have already appeared in other related work, and the current methodology and experiments are insufficient to support the paper's arguments. Whether the authors can address these concerns will directly influence my overall assessment of the paper.

3 comments
AuthorsOpenReview ↗

Follow up on our response

Thank you again for your detailed review. We wanted to kindly follow up on our response, as your questions led to several new experiments. In brief:

  • Attribution: at a matched 16-bit budget and identical preprocessing, both a random subspace and a simpler channel selection fall well below Q-WAM's action-sensitive subspace (2.2 to 9.2 points lower), so the gain comes from which directions are protected, not from the extra budget or the preprocessing.
  • Full denoising: it is not required; a single-step AOG reaches the same success and reduces Fast-WAM's total calibration from about 12.2 to about 5.2 H100 GPU-hours.
  • Reuse across tasks: a subspace calibrated on 25 tasks stays within 0.7 points of the original on 25 unseen tasks, and recalibration is a short, one-time offline step.
  • Latency and VRAM: we report end-to-end latency and peak memory for bf16, W4A4 without ASP, SVDQuant and Q-WAM under identical conditions; the protected subspace adds at most 0.25 GB of peak memory.

We would greatly appreciate your feedback on whether these results resolve your concerns, and we are happy to provide any further clarification. If they do, we would be grateful if you would consider them in your final assessment.

Thank you for your rebuttal, which addresses some of my concerns. However, this work shares methodological similarities with prior works (e.g., QVLA/SVDQuant), and its novelty is relatively limited. I will keep my current rating.

Thank you for reading our rebuttal and for your reply. We are glad that it addressed some of your concerns. Since the main remaining point is the relation to SVDQuant and QVLA, we would like to clarify how Q-WAM differs from both.

SVDQuant and QVLA contribute on the weight side; neither proposes a novel method for activation quantization.

  • SVDQuant's contribution is a 16-bit low-rank branch obtained from the SVD of the weights. For activations, it relies on standard smoothing (as in SmoothQuant), and then they apply a generic group-wise quantization. Nothing in the activation is protected or selected.
  • QVLA's contribution is an action-sensitive bit allocation over weight output channels. For activations, it uses a single uniform bit width across the model.

In both methods, every activation direction is quantized in the same way, regardless of its effect on the action.

Q-WAM is an activation quantization method for world action models, with three contributions:

  1. Action-Observability Gramian (AOG): for each layer, a matrix over the full activation space that captures how combinations of activation directions jointly affect the generated action, estimated with random probes through the denoising process. Unlike per-channel scores, it captures interactions across channels, and its eigenvectors directly give the directions to protect.
  2. Action-Subspace Protection (ASP): the leading AOG directions, which are dense and mix many channels, are kept in 16 bits, and the rest of the computation is quantized to 4 bits.
  3. Expert selection: for multi-expert models, the aggregated AOG sensitivity determines which experts receive protection.

To our knowledge, none of these has been explored in prior quantization work.

The rebuttal experiments confirm that this formulation drives the gain:

  • Channel-level sensitivity, the view QVLA is built on, is not enough. Protecting the most action-sensitive individual channels stays 2.2 to 7.2 points below Q-WAM: on Fast-WAM, 32 channels cover only 22.5% of the expected action error, versus 97.9% for Q-WAM's 32-dimensional subspace.
  • The choice of directions matters. A random subspace with the same budget loses 2.6 to 9.2 points.
  • State-of-the-art W4A4 results. Q-WAM stays within 1.1 points of bf16 on all three models and outperforms SVDQuant by 2.5 to 8.7 points at similar or lower memory.

If there is a specific aspect that you find overlaps with prior work, we would be grateful if you could point us to it, so that we can address it directly in the revision.

Reviewer_D7vGOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

Q-WAM proposes a method for quantizing world action models to 4-bit precision. This approach estimates how perturbations in intermediate activation values ​​affect the final generated actions, and preserves the most sensitive components within a separate, small-scale, high-precision branch. The remaining operations are subjected to 4-bit quantization. Evaluations conducted using three backbone models in the RoboTwin 2.0 environment, as well as two models deployed on a physical robot, demonstrate improved task success rates compared to existing quantization techniques, including SVDQuant.

Critical strengths
  1. The quantization objective provides a sound rationale for robot control.

Prioritizing the preservation of directions based on their impact on the final action is a more meaningful approach than simply maintaining weight structures or minimizing local reconstruction error. Sensitivity analysis supports this, and ablation studies demonstrate that, beyond smoothing and rotation, preserving the action subspace yields further performance gains.

  1. Experimental results substantiate the efficacy of this method.

Across three different backbones, Q-WAM improves simulation success rates by approximately 2.5 to 8.7 percentage points compared to SVDQuant, while maintaining similar or lower storage costs. Significant performance improvements over SVDQuant are also observed in real-world robotic experiments.

Critical weaknesses
  1. Substantial efficiency gains have not been fully demonstrated at the system level.

Reductions in memory usage pertain specifically to quantized blocks, and latency measurements are taken at the block level rather than end-to-end. Estimating action sensitivity also requires backpropagation through the denoising process; the appendix notes that Fast-WAM calibration necessitates approximately 130,000 backpropagation steps. Providing data on total inference latency, peak memory consumption, and the actual computational cost of calibration would help clarify the deployment advantages.

  1. It would be possible to more directly isolate the contribution of specific protected subspaces.

While the paper compares the method against SVDQuant and includes an ablation study of its components, further downstream analysis e.g., replacing the action-sensitivity-based approach with random or weight-based directions within the same pipeline, would provide greater clarity regarding the underlying mechanisms. Although Appendix A.5 offers a local comparison against random directions, it does not report the corresponding differences in task success rates.

Questions for authors
  1. What are the end-to-end inference latency, peak GPU memory, and offline calibration time for Q-WAM compared with BF16 and SVDQuant?

  2. With rank, precision, smoothing, rotation, and protected layers held fixed, how does task success change when the action-sensitive basis is replaced by a random or weight-derived basis?

1 comment
AuthorsOpenReview ↗

Follow up on our response

Thank you again for your review. We wanted to kindly follow up on our response. Following your two questions, we added:

  • System-level cost: end-to-end latency, peak GPU memory and offline calibration time for bf16, SVDQuant and Q-WAM.
  • Isolating the protected subspace: a matched-budget experiment that replaces the action-sensitive basis with a random one while keeping rank, precision, smoothing, rotation and protected layers fixed.

Your question on calibration cost also led us to a single-step estimate that reduces Fast-WAM's total calibration from about 12.212.2 to about 5.25.2 H100 GPU-hours.

We would be grateful to hear whether these results address your concerns, or whether there is anything further we can clarify.

Comments on the paper

1 comment

Dear Area Chair and Reviewers,

We thank you again for your time and constructive feedback. We have posted detailed responses to all three reviews, with new experiments addressing each question. To make the discussion easier, we summarize the main concerns and how we addressed them below. We would greatly value your feedback and are happy to clarify anything further.

Shared concern (D7vG, TXNp, ctjZ): does the gain come from the action-sensitive subspace, or from the extra 16-bit budget and preprocessing? We kept rank 32, the 16-bit branch, smoothing, rotation, protected layers and kernels identical, and changed only which directions are protected. A random subspace loses 2.6 to 9.2 points across all three models, and selecting the 32 most action-sensitive channels stays 2.2 to 7.2 points below Q-WAM. The gain therefore comes from the choice of directions, not from the budget or the preprocessing.

Reviewer D7vG: system-level efficiency.

  • End-to-end latency and peak memory: measured from observation to action chunk with real 4-bit kernels on one L40S, Q-WAM runs at the speed of the bf16 model while reducing peak GPU memory by 1.8x to 2.9x.
  • Calibration cost: reported for every component.

Reviewer TXNp: necessity of the full denoising process, reuse across tasks, and latency without ASP.

  • Full denoising: not required. A single-step AOG matches the full-process estimate in closed loop, and reduces Fast-WAM's total calibration from about 12.2 to about 5.2 H100 GPU-hours.
  • Reuse across unseen tasks: a subspace calibrated on 25 RoboTwin tasks stays within 0.7 points of the original on 25 unseen tasks.
  • Latency without ASP: compared with W4A4 without ASP, the protected subspace adds 4.5 to 8.9% latency and at most 0.25 GB of peak memory, while raising average success by 3.2 to 5.6 points.

Reviewer ctjZ: reliability of the probe-based estimate.

  • Assumption: we stated the assumption under which the block-diagonal estimate applies, and verified it with measured block statistics.
  • Convergence: a study over the probe count (P = 4, 12, 48) shows that the protected subspace has stabilized by P = 12, with no meaningful change in closed-loop success beyond it.

All of these results are included in the revised version of our paper in blue text. Thank you again for your engagement.

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.