acceptodds
Withdrawn from ICLR 2027

Beyond Action Deviation: Execution-Level Quantization Sensitivity for Vision-Language-Action Models

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Post-training quantization (PTQ) offers a practical way to reduce the deployment cost of vision-language-action (VLA) models. When uniform low precision cannot preserve control fidelity, mixed-precision PTQ relies on sensitivity estimation to identify which layers should retain higher precision. Existing sensitivity criteria mainly measure internal numerical distortion or action-level deviation, overlooking how quantization perturbations propagate through robot execution and influence task performance. Consequently, mixed-precision quantization guided by these criteria can lead to substantial performance degradation, particularly when applied to the sensitive action generation modules. We identify the need for an intermediate perspective between action deviations and task outcomes that captures robot-dependent execution information. To this end, we introduce Execution-Level Quantization Sensitivity (EQS), a layer-wise sensitivity criterion for mixed-precision quantization of VLA action generators. It models how quantization-altered action commands translate into horizon-level motion discrepancy and evaluates the discrepancy's execution significance under the current robot configuration. Together, these components provide an execution-aware sensitivity signal for layer prioritization. Experiments across simulation benchmarks and real-robot manipulation tasks show that EQS exceeds the best-performing baseline by 0.44 8.06 percentage points in average task success, with gains reaching 33.82 percentage points over individual baselines.

Primary area:
applications to robotics, autonomy, planning
Keywords:
vision language action model, post training quantization, robot execution

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

The paper introduces Execution-Level Quantization Sensitivity (EQS), a criterion for allocating mixed-precision PTQ in VLA action-generation modules. EQS converts quantization-induced action differences into horizon-level SE(3) motion discrepancies, evaluates them with a robot-configuration-dependent operational-space inertia metric, balances translation, rotation, and gripper components, and uses gradient attribution to rank layers for BF16 retention. The authors evaluate EQS on π0.5/LIBERO and CogACT/SimplerEnv under multiple precision budgets, compare against action-level, drift-aware, and layer-statistics baselines, and conduct real-robot experiments on four Airbot tasks. EQS generally achieves higher task success with comparable quantization and inference costs.

Critical strengths

1.EQS addresses a meaningful gap between easily computed action deviation and expensive closed-loop task evaluation. The combination of ordered SE(3) composition and configuration-aware operational-space weighting is technically well motivated, and the ablations support the contribution of both components.

2.The empirical validation is broad and controlled: two VLA models, two simulation settings, multiple BF16 retention budgets, calibration-overhead measurements, and real-robot experiments. The method consistently outperforms the evaluated sensitivity baselines under the same retention configurations.

Critical weaknesses

1.The physical execution proxy is under-validated. EQS evaluates a horizon discrepancy using the operational-space inertia at the horizon-start configuration and keeps it fixed over the horizon, while real execution also depends on controller dynamics, contacts, joint limits, saturation, and changing configurations. The paper does not establish that the proposed cost transfers across robots or reliably correlates with closed-loop task degradation beyond the reported benchmarks.

2.The mapping from EQS scores to the final jointly quantized model is not sufficiently specified or validated. Equation (15) uses gradient magnitudes, but it is unclear how gradients pass through W4A4 quantization, whether scores are computed with one candidate layer quantized at a time, and how interactions among simultaneously quantized layers are handled. Without rank stability or sensitivity-to-final-task-loss analysis, the layer-allocation mechanism remains difficult to assess.

Questions for authors

1.How is the EQS gradient in Eq. (15) computed through W4A4 quantization? Is a straight-through estimator used, and are scores obtained by quantizing each candidate layer individually while all other layers remain full precision?

2.Does the operational-space matrix use the actual configuration at every execution step or only the horizon-start configuration? How sensitive are the rankings to the horizon length, damping coefficient, and robot/controller dynamics?

3.Are calibration chunks strictly disjoint from evaluation episodes and tasks? Please report sensitivity-ranking stability across calibration resamples and confidence intervals for the 40-trial real-robot results.

Reviewer_PdzMOpenReview ↗
Rating
2
Confidence
3
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes EQS, an execution-aware sensitivity metric for mixed-precision VLA quantization. It combines SE(3) motion composition, operational-space inertia weighting, and gradient-based layer ranking to reduce performance degradation after quantization. Experiments on LIBERO, SimplerEnv, and real robots show improvements over existing methods.

Critical strengths

The paper addresses a practical problem in VLA inference. The experiments cover multiple models, quantization settings, and real-robot tasks, with generally promising results.

Critical weaknesses
  1. Questionable motivation for motion discrepancy. The paper only considers the final pose discrepancy of an action chunk, ignoring intermediate trajectory deviations. However, for contact-rich manipulation, errors during execution may already cause task failure, even if they cancel out at the end. The StepWise ablation does not fully address this issue, cause it does not consider accumulated pose errors along the trajectory.
  2. Unclear justification for inertia weighting. It is not clear why errors along directions with higher operational-space inertia should have a greater impact on task success. The ablation shows some experimental benefits, but does not establish why inertia is a suitable measure of execution importance.
  3. Limited methodological novelty. The method mainly combines existing techniques, including SE(3) motion composition, operational-space inertia, and gradient-based sensitivity estimation. DA-PTQ has already considered motion drift caused by action errors, so the methodological contribution is incremental.
  4. Incomplete efficiency evaluation. The paper lacks a clear comparison of inference speed or memory usage against the full-precision model. Also, using the same BF16 layer-retention ratios does not guarantee equal memory budgets across methods.
Questions for authors
  1. Why only consider the final pose discrepancy? How would the method compare with a trajectory-level metric that accounts for accumulated pose errors at each step?
  2. Why is operational-space inertia a good proxy for task-level impact? Have other weighting metrics been considered?
  3. How well do the proposed sensitivity scores correlate with actual task performance degradation after quantization?
  4. What are the inference speedup and memory savings compared with the full-precision model?
Reviewer_S9XCOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes Execution-Level Quantization Sensitivity (EQS) for mixed-precision allocation in the action generators of vision-language-action models. The method composes action perturbations over an execution horizon, evaluates the resulting motion discrepancy using a configuration-dependent operational-space inertia metric, and applies component balancing and gradient-based layer attribution. The resulting scores determine which layers remain in BF16 and which are quantized to W4A4. Experiments on π0.5 with LIBERO, CogACT with SimplerEnv, and an Airbot robot report improved task success over alternative sensitivity criteria. Additional experiments examine component ablations, calibration overhead, the BCH approximation, and compatibility with HoloQ-VLA.

Critical strengths

1.The paper introduces a meaningful execution-aware perspective on VLA quantization. Action deviations of similar magnitude can have different consequences depending on their temporal composition and the robot configuration. Combining horizon-level motion discrepancy with operational-space inertial weighting provides a coherent way to incorporate these factors into precision allocation. This is the paper’s strongest conceptual contribution.

2.The empirical evaluation spans two VLA architectures, two simulation benchmarks, multiple precision-retention settings, and physical robot tasks. The reported gains on CogACT are substantial, while the ablations and calibration-overhead measurements help assess the contribution and cost of the proposed components. Appendix D also validates the BCH approximation against the exact product-log formulation, providing useful support for the computational approximation.

Critical weaknesses
  1. The connection between the stated quantization-sensitivity objective and the implemented estimator is insufficiently established. Section 3 motivates EQS using action discrepancies induced by quantizing a candidate layer. In the supplied π0.5 implementation, however, multiple action-expert linear layers receive fake-W4 perturbations simultaneously, and layer scores are obtained from gradients of a shared objective. The compared actions are single-timestep decoded estimates at a sampled noise level rather than outputs of the complete sampling procedure. The CogACT scoring path also appears to compare a reconstructed action estimate with a cached policy output without a matched full-precision forward pass at the same noisy input. These may be reasonable computational surrogates, but the manuscript does not adequately explain or validate their relationship to the stated intervention. Figure 1’s two selected layer pairs do not establish that the resulting scores reliably predict task degradation across layers. This uncertainty affects the interpretation of the central contribution.

  2. The evaluation does not yet isolate the benefit of EQS under a matched resource budget or establish the reliability of the smaller gains. Fixed MLP and attention layer-retention ratios allow different BF16 parameter counts, as acknowledged in Appendix G. Showing that EQS sometimes retains fewer parameters is useful, but does not replace matched-budget comparisons. Table 7 reports relative latency across criteria without a full-precision reference or memory measurements; moreover, the supplied action-layer wrapper uses fake quantization followed by floating-point linear operations, so native low-bit deployment benefits are not established by that implementation alone. The success-rate tables also omit uncertainty estimates, which matters for improvements such as 0.44 percentage points on LIBERO. In the real-world study, EQS exceeds QVLA by nine successes out of 160 trials, but paired outcomes or confidence intervals are not provided. Together, these limitations weaken the evidence for a reliable accuracy–efficiency advantage.

Questions for authors

1.Please clarify the exact scoring procedure for each model: which layers are quantized simultaneously, whether activation quantization is included, and how the full-precision and perturbed branches share diffusion inputs and randomness. In particular, how does the CogACT implementation separate quantization-induced error from reconstruction or sampling error?

  1. How well do the implemented gradient-based scores agree with task degradation from matched, single-layer quantization interventions? Evidence across a representative set of layers, beyond the two selected pairs in Figure 1, would materially affect my assessment of the proposed sensitivity criterion.

  2. Do the improvements persist when methods retain the same number of BF16 parameters or use the same nominal weight-storage budget? Please also clarify whether Table 7 uses native low-bit kernels or fake-quantized floating-point operations, and provide the corresponding full-precision reference if available.

  3. Please specify the rollout counts, seeds, and aggregation procedure underlying Tables 1, 5, and 6. Can seed-level results or confidence intervals be provided, together with paired outcomes for the real-world evaluation?

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

Related papers

Discussion (0)

Sign in to comment.