acceptodds
Under review as a conference paper at ICLR 2027

SACQ: Structure-Aware Codebook Quantization for Ultra-Low-Bit Vision-Language-Action Models

Abstract

Vision-Language-Action (VLA) models produce remarkably capable robot behavior, but their memory footprint makes deployment on resource-constrained robotic platforms challenging. Weight quantization is a natural remedy, yet post-training quantization (PTQ) recipes inherited from large language models degrade sharply when applied to VLAs at ultra-low precision. Two limitations become particularly important: existing low-bit PTQ methods typically assess quantization error at the layer or projection level without modeling the downstream operation through which it propagates, and allocate precision without directly measuring how reducing a layer’s precision changes the policy action. We introduce SACQ, an ultra-low-bit weight-only vector-codebook quantization framework for VLAs. SACQ combines (1) structure-aware codebook fitting, which fits each layer using a role-specific quadratic metric derived from the operation its projection feeds; (2) action-aware mixed-bit allocation, which assigns precision according to decoded-action degradation as individual layers are compressed; and (3) two-stage post-quantization refinement, which first reassigns codeword indices with second-order error compensation and then calibrates codebooks and row-group scales against the decoded action. On GR00T N1.7, SACQ retains 93.85% average success across all four LIBERO suites at approximately 1.5 bits per language/action weight, only 0.25 percentage points below the 94.10% BF16 reference, while reducing resident model memory from 5.677 GB to 1.175 GB, a 4.83× reduction. On SimplER Google Robot, at approximately 2 bits per language/action weight, SACQ achieves 63.76% and 52.27% success under Visual Matching and Variant Aggregation, respectively, compared with 58.90% and 53.45% for BF16.

Primary area:
infrastructure, software libraries, hardware, systems, etc.
Keywords:
Vision-Language-Action Models, Post-Training Quantization, Codebook Quantization, Model Compression, Robotics

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

This paper studies quantization of vision-language-action models, aiming to reduce storage while preserving task success. Its motivation is that local reconstruction error does not fully capture the effect of different weights on the final action. SACQ first fits vector codebooks according to the downstream operations of each projection layer and allocates mixed bit widths based on action error per bit saved. It then refines codeword assignments with second-order error compensation and calibrates the codebooks and scales against the BF16 model’s action outputs. Experiments show that the method reduces storage while largely preserving task performance.

Critical strengths

The proposed method achieves a great trade-off between the storage and the task execution success rate of VLA, the paper is clear and easy to follow, and the results are promising, with Table 1 retaining nearly the reported BF16 success at approximately 1.5 L/A bpw with substantially lower resident memory.

Critical weaknesses
  1. Lack of experimental details: it is unclear what supervision, what calibration samples, and what optimization budget are used when adapting ActQuant to the GR00T/DiT, since the original ActQuant uses the ground truth actions while SACQ uses the output of the BF16 teacher.
  2. More analysis on the performance: the performance improvements are relatively minor when removing structured codebook fitting mixed bit width allocation, 92.60% and 93.25% respectively, compared to 93.85% of the full model. I suggest providing more analysis or per-task results to help determine whether these differences are stable.
Questions for authors
  1. What supervision, calibration data, and optimization budget are used for the adapted ActQuant baseline? How are its statistics computed for the DiT, and how does this differ from the original ActQuant?
  2. How do the 576 training requests per suite in Appendix C correspond to the 2,880 observations in Table 8? Please clarify the units and the training and validation split. It would also be helpful to report the calibration time and memory cost.
  3. Can the authors provide per-task results or uncertainty estimates for the structural codebook fitting comparison at 1.5 bpw? Results from a few different calibration seeds would also help show whether the improvement is stable.
4 comments
AuthorsOpenReview ↗

Response to Reviewer kZED

We thank the reviewer for the careful and constructive review, and for recognizing the useful low-bit operating point of SACQ and the motivation of the structural metrics. We address each point below and will add these clarifications to the revised version.

[W1 & Q1] Protocol of the adapted ActQuant baseline

The original ActQuant targets OpenVLA-OFT and π0.5. For GR00T N1.7, we used the official ActQuant code for bit allocation and quantization, and adapted the model-specific parts, such as activation collection, the Fisher loss, and the layer mapping, to GR00T. To keep the calibration data equivalent across methods, ActQuant uses the same 576 training requests per suite as SACQ and the other baselines (2,880 observations, see our answer to Q2), and uses the actions predicted by the BF16 policy as target actions instead of ground-truth demonstration actions, as in SACQ. Both the HSIC-based sensitivity and the Fisher importance are computed with these BF16 actions.

ActQuant performs HSIC-based bit allocation followed by Fisher-weighted selection of the quantization scales and quantized values, both computed over the same 576 requests. For the DiT action head, which is executed over four denoising steps, we compute the HSIC sensitivity at each step and average the four scores with equal weights. Following the original method for models with only an action-head pathway, the Fisher importance is computed from the action loss, which is the flow-matching loss of GR00T’s action head, with the noise and time step sampled once per request. The quantization scope covers the same 338 language and action linear layers as SACQ, including the attention, feed-forward, AdaLN modulation, and output projections, with the bit budget allocated jointly across the language backbone and the action head. In addition, bit budgets are matched in nominal bpw as for the other methods, and the quantized weights are evaluated in the same BF16 evaluation path as the other methods rather than in the original deployment runtime.

[W1 & Q2] Sample units and data split

In our LIBERO action-calibration dataset, each saved request contains a batch of five observations, where each observation consists of the third-person and wrist camera images, the robot state, and the language instruction. The 576 training requests per suite in Appendix C therefore correspond to 2,880 observations per suite, which is the unit used in Table 8. Accordingly, the calibration budgets of 90, 180, 360, 720, 1,440, and 2,880 observations in Table 8 correspond to 18, 36, 72, 144, 288, and 576 training requests per suite, respectively.

For each suite, we use 576 training requests (2,880 observations) and 64 separate validation requests (320 observations). The validation requests are used only to select the action-level calibration checkpoint.

The action-level calibration runs for 12 epochs with a global batch size of 6 requests, corresponding to 1,152 optimizer updates per suite. We are currently measuring the wall-clock time and peak GPU memory of each stage under a consistent setting, and will report them in a follow-up comment.

[W2 & Q3] Uncertainty of the 1.5-bpw comparison

Using the existing paired outcomes, we quantified the uncertainty of the comparison between structure-aware and layer-local fitting at 1.5 L/A bpw under the full refinement budget. Both models are evaluated on the same 2,000 episodes, where each episode uses the same initial state and policy noise for both models, so their outcomes can be paired. At the episode level, structure-aware fitting succeeds on 83 episodes where layer-local fitting fails, and layer-local fitting succeeds on 58 episodes where structure-aware fitting fails (exact McNemar test, p = 0.043). At the task level, a paired t-test over the 40 tasks gives t(39) = 2.12 and p = 0.040, with a 95% confidence interval of [+0.06, +2.44] percentage points for the mean difference of +1.25 points. The advantage of structure-aware fitting is therefore small but statistically significant at the 0.05 level.

AuthorsOpenReview ↗

Follow-up: Calibration Time and Memory Cost

As mentioned in our previous response, we measured the wall-clock time and peak GPU memory usage of SACQ's quantization and calibration stages using two NVIDIA RTX 4090 GPUs (24 GB each).

The measurement covers activation statistics collection, codebook fitting, GPTQ preparation, and mixed-bit allocation across all four LIBERO suites. The four suites are jointly used to determine a shared bit allocation, while the subsequent refinement is performed separately for each suite. Here, we measured index reassignment and action-level calibration for the Spatial suite at 1.5 L/A bpw.

Table R9. Wall-clock time and peak GPU memory of SACQ's quantization and calibration stages.

StageWall-clock timePeak GPU memory (GB)
VLM activation statistics & codebook fitting5m 15s10.13
DiT activation statistics collection5m 33s11.73
DiT codebook fitting27m 4s1.21
GPTQ preparation22m 58s12.59
Mixed-bit allocation70m 22s16.56
Index reassignment4m 52s11.04
Action-level calibration29m 12s23.38
Total2h 40m 7s23.38

Peak GPU memory represents the maximum PyTorch allocated memory on either GPU. Some stages were executed concurrently, so individual stage times do not sum to the total wall-clock time. GPTQ preparation includes both allocation candidates and fixed background weights.

These measurements are based on a single run using existing calibration data and BF16 teacher actions. Calibration dataset collection, BF16 teacher action generation, and rollout evaluation are excluded.

Thank you again for your valuable feedback.

AuthorsOpenReview ↗

Gentle Reminder: Author–Reviewer Discussion Period Ending Soon

Dear Reviewer kZED,

Thank you again for your thoughtful review and valuable feedback.

As the author–reviewer discussion period is approaching its end, we would like to briefly follow up on our rebuttal and additional results.

In our rebuttal, we provided the following clarifications and analyses:

Protocol of the adapted ActQuant baseline: We clarified that the adapted ActQuant baseline was evaluated using the same calibration requests as SACQ, with the actions predicted by the BF16 policy as targets. We also described how HSIC-based sensitivity and Fisher importance were adapted to GR00T's DiT action head.

Sample units and data split: We explained that each training request contains five observations, so 576 training requests correspond to 2,880 observations per suite, with 64 separate validation requests.

Uncertainty of the 1.5-bpw comparison: Using the existing paired outcomes, we reported an exact McNemar test (p = 0.043) and a task-level paired t-test (p = 0.040), with a 95% confidence interval of [+0.06, +2.44] percentage points.

Calibration time and memory cost: We additionally reported the measured wall-clock time and peak GPU memory for each stage of SACQ's quantization and calibration pipeline in a separate follow-up response.

If these results address your concerns, we would greatly appreciate it if you could acknowledge this and consider reassessing your rating. We would be happy to clarify any remaining questions.

Thank you again for your time and consideration.

Best regards,

The Authors

The authors’ response addresses my concerns. I maintain my score.

Reviewer_P3HZOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes SACQ (Structure-Aware Codebook Quantization), an ultra-low-bit, weight-only post-training quantization framework designed to reduce the memory footprint of VLA models. SACQ combines three ideas: projection-role-specific codebook fitting that accounts for downstream operations such as attention and feed-forward computation; mixed-bit allocation based on each layer’s effect on the decoded action; and post-quantization refinement through second-order index reassignment followed by action-level calibration of codebooks and scales. The authors evaluated SACQ primarily on GR00T N1.7 using LIBERO and SimplER, with additional experiments on OpenVLA-OFT and LIBERO-Pro. Ablation studies indicate that structure-aware fitting is particularly beneficial at lower precision and under limited refinement budgets, while post-quantization refinement is essential for recovering near-full-precision performance.

Critical strengths
  • The analysis on how quantization errors propagate through downstream operations and affect the final predicted actions. This is particularly relevant to VLA architectures, which usually have very long steps, and the pursuit of per-layer reconstruction may overlook the long tail consequence and different behaviors. Structure-aware designs sound reasonable to me in this context.

  • Strong ultra-low-bit empirical performance across multiple evaluation settings. The results also show substantial improvements over the evaluated quantization baselines at comparable precision levels. Additional evaluations on SimplER, OpenVLA-OFT, and LIBERO-Pro provide evidence that the approach is not restricted to a single benchmark or architecture.

Critical weaknesses
  • It seems to me that the refinement process contributes a lot portion in the policy persevering accuracy, according to Table 3. I am not sure but are there other vanilla refinement process so that we could ablate the contribution and effectiveness of structure-aware fitting more accurately and directly? This matters because structure-aware fitting is presented as the main methodological novelty

  • Although that the authors acknowledge that they are mainly focused on memory footprint optimization in the limitation section. I still feel that somehow the work is not complete enough for practical reasons.

Questions for authors
  • Following my first comments in the weakness section, could the authors separately report the improvements attributable to second-order index reassignment and action-level codebook/scale calibration? Additionally, how would a layer-local reconstruction baseline perform when given exactly the same refinement procedure and computational budget?
3 comments
AuthorsOpenReview ↗

Response to Reviewer P3HZ

We thank the reviewer for the constructive feedback and for recognizing our analysis of how quantization errors propagate through downstream operations, as well as the strong ultra-low-bit performance across multiple evaluation settings. We conducted the requested analyses and will include them in the revised version. We address each point below.

[W1 & Q1] Refinement and fitting ablations

Separate contributions of the two refinement stages.

Table R7 reports SACQ with index reassignment and action-level calibration each applied alone. Index reassignment alone substantially improves the quantized policy on both benchmarks without any gradient-based optimization.

Table R7. Average success (%) of SACQ by refinement stage.

BenchmarkBefore refinementIndex reassignment onlyAction-level calibration onlyBoth (full)BF16
LIBERO (1.5 L/A bpw)47.5081.8593.3593.8594.10
SimplER VM (2 L/A bpw)46.4457.9853.1163.7658.90
SimplER VA (2 L/A bpw)40.2550.9145.8752.2753.45

Action-level calibration alone is close to the full refinement on LIBERO but falls below it on SimplER, and applying both stages yields the best results on both benchmarks.

Layer-local baseline under the same refinement.

With exactly the same refinement procedure and budget, layer-local reconstruction fitting achieves 92.60%, compared with 93.85% for structure-aware fitting. This comparison is included in Appendix G (Table 8), where, at each refinement budget, only the fitting objective is changed while the mixed-bit allocation and the refinement procedure are kept identical. We reproduce it below for convenience, with an additional before-refinement row.

Table R2. LIBERO average success (%) by fitting objective and refinement budget, 1.5 L/A bpw.

Calibration obs. per suiteStructure-awareLayer-localEuclidean
Before refinement47.5015.650.00
9084.1574.5539.75
18088.8584.0557.15
36090.7088.1078.55
72091.2591.4586.45
1,44092.5592.9590.60
2,880 (full)93.8592.6092.60

Before refinement, structure-aware fitting outperforms layer-local reconstruction fitting by 31.85 percentage points, consistently across all four suites. As the refinement budget increases, the gap narrows, showing that structure-aware fitting provides a stronger initialization whose benefit is largest when refinement is limited.

[W2] Inference efficiency of SACQ

We agree that memory reduction alone does not fully establish practicality. To address this concern, we optimized SACQ inference using custom CUDA C++ and Triton kernels that decode the compressed codebook weights at inference time and, with 8-bit activations (W1.5A8), perform INT8 GEMMs, together with low-precision attention based on SageAttention and CUDA Graph execution. BF16 uses the same common runtime optimizations for a fair comparison, and latency is measured on an RTX 4090 with batch size 1.

Table R8. Inference efficiency on GR00T N1.7 (LIBERO).

ConfigurationAvg. success (%)Latency (ms)Relative speed
BF1694.1021.281.00×
SACQ (W1.5A16)93.8520.201.05×
SACQ (W1.5A8)93.5514.071.51×

SACQ reduces resident model memory by 4.83× (5.677 → 1.175 GB). With 8-bit activations, it achieves a 1.51× speedup over BF16, with success rates only 0.30 and 0.55 percentage points below W1.5A16 and BF16, respectively.

AuthorsOpenReview ↗

Gentle Reminder: Author–Reviewer Discussion Period Ending Soon

Dear Reviewer P3HZ,

Thank you again for your thoughtful review and valuable feedback.

As the author–reviewer discussion period is approaching its end, we would like to briefly follow up on our rebuttal and additional results.

In our rebuttal, we provided the following analyses and experiments:

Refinement ablation: We separately evaluated index reassignment alone (81.85%) and action-level calibration alone (93.35%), as well as their combination (93.85%) on LIBERO, with corresponding results on SimplER.

Structure-aware contribution: Under identical refinement conditions, structure-aware fitting outperforms layer-local reconstruction at full calibration (93.85% vs. 92.60%), with a larger gain under limited calibration (84.15% vs. 74.55% at 90 observations).

Inference efficiency: With optimized kernels and 8-bit activations, SACQ achieves a 1.51× speedup over BF16 while retaining 93.55% LIBERO success.

If these results address your concerns, we would greatly appreciate it if you could acknowledge this and consider reassessing your rating. We would be happy to clarify any remaining questions.

Thank you again for your time and consideration.

Best regards,

The Authors

Thanks to the authors for the detailed reply. With the newly added experiments on the refinement process, the contributions become more clear to me. I hope the authors would include this split in future versions. I will maintain my positive rating.

Reviewer_6ARjOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

This work focuses on maintaining the performance of VLA models under ultra-low-bit weight quantization. Conventional PTQ methods for LLMs primarily target reconstruction errors in local weights or layer outputs; they often fail to account for the distinct roles that various projection layers play in subsequent computations, nor do they consider how quantization errors propagate to affect the final robotic actions. To address this, the authors propose three techniques: Structure-aware codebook fitting, Action-aware mixed-bit allocation, and Two-stage post-quantization refinement. Experimental results demonstrate that this approach, termed SACQ, effectively preserves high task success rates, even in certain ultra-low-bit quantization configurations.

Critical strengths

Unlike traditional quantization methods that primarily focus on layer-wise reconstruction error, this paper considers the roles of different projection layers in subsequent computations and devises specific quantization error metrics for modules such as Attention, FFN, and AdaLN. This design is grounded in clear mathematical motivation and offers a viable research approach for structure-aware quantization in VLA models.

This paper focuses on maintaining the performance of VLA models under ultra-low-bit quantization. As VLA models continue to grow in scale, reducing their memory footprint—thereby enabling their deployment on resource-constrained robotic platforms—has become a research direction with significant practical value.

Critical weaknesses

The three main modules proposed in this paper are built upon existing techniques: Hessian-aware vector quantization, action-guided mixed-bit allocation, and post-quantization refinement. While the development of quantization metrics tailored to the distinct projection structures of VLAs offers a degree of novelty, the overall framework is more akin to a combination of existing techniques, lacking the kind of core algorithmic innovation that would constitute a significant breakthrough.

According to Table 3, at approximately 1.5 bpw, structure-aware fitting boosts the model’s success rate to 47.50%, with subsequent refinement further increasing it to 93.85%. However, when structure-aware fitting is omitted—retaining only the refinement stage—the model still achieves a success rate of 92.60%. This indicates that the final performance relies primarily on the subsequent optimization process, while the contribution of the paper’s core component—the structure-aware metric—to the overall pipeline is relatively limited, warranting more thorough analysis and justification.

The experiments are insufficient. I encourage the authors to conduct further experiments using other VLA models as well as on real-world hardware to demonstrate the effectiveness of the proposed method.

Comparisons with baselines may suffer from an unfair allocation of the optimization budget. After completing the initial quantization, SACQ performs additional gradient-based optimization of the codebooks and scales, using the decoded actions from the BF16 model as a supervisory signal. However, the paper does not explicitly demonstrate that the baselines, such as GPTQ, ActQuant, and TWLA, were granted an equivalent action-level calibration budget. Consequently, the significant performance gains reported in Tables 1 and 2 may stem, in part, from this extra training process, rather than being solely attributable to the proposed quantization method.

Questions for authors

n/a

3 comments
AuthorsOpenReview ↗

Response to Reviewer 6ARj (1/2)

We thank the reviewer for insightful comments and recognizing the novelty of the projection-specific quantization metrics. We address each concern below.

[W1] Novelty and contributions of SACQ

Our main contribution is to preserve task success under post-training quantization of VLAs at ultra-low precision on LIBERO, where the other evaluated methods degrade sharply (Table R1). We agree that SACQ builds on established components, such as second-order error compensation and codebook fine-tuning, and we do not claim these components as new. Our contributions are the following two-fold.

First, we introduce structure-aware codebook fitting, which derives each projection’s metric from the downstream operation that consumes its output, rather than from its local reconstruction error. Even before any refinement, structure-aware fitting with the mixed-bit allocation outperforms GPTQ-Ternary and ActQuant [1] at 1.5 L/A bpw (Table R1).

Second, we combine these components into a framework that works reliably at ultra-low precision for VLAs. Adding only index reassignment, which relies only on second-order statistics, substantially raises the success rate from 47.50% to 81.85%, well above all baselines at this bit budget. With the full refinement, SACQ reaches 93.85%, close to the 94.10% BF16 reference.

Table R1. Average success (%) across the four LIBERO suites at 1.5 L/A bpw.

MethodEnd-to-end action fine-tuningAvg. success (%)
GPTQ-Ternary (1.585 bpw)No31.95
ActQuantNo30.05
SACQ, before refinementNo47.50
SACQ, index reassignment onlyNo81.85
SACQ, fullYes93.85
BF16–94.10

[1] Akbari, Arash, et al. "ActQuant: Sub-4-bit Action-Guided Quantization for Vision-Language-Action Models." NeurIPS (2026).

[W2] Contribution of structure-aware fitting to the overall pipeline

We acknowledge that post-quantization refinement is an essential component of SACQ. However, the comparison in Table 3 (93.85% vs. 92.60%) is made under the full refinement budget, where refinement compensates for most of the difference in initialization. Appendix G (Table 8) examines this comparison across refinement budgets, where, at each budget, only the fitting objective is changed while the mixed-bit allocation and the refinement procedure are kept identical. We reproduce it below for convenience, with an additional before-refinement row.

Table R2. LIBERO average success (%) by fitting objective and refinement budget, 1.5 L/A bpw.

Calibration obs. per suiteStructure-awareLayer-localEuclidean
Before refinement47.5015.650.00
9084.1574.5539.75
18088.8584.0557.15
36090.7088.1078.55
72091.2591.4586.45
1,44092.5592.9590.60
2,880 (full)93.8592.6092.60

Without structure-aware fitting, the pipeline degrades sharply as the refinement budget decreases. With 90 calibration observations per suite, Euclidean fitting reaches only 39.75%, while structure-aware fitting reaches 84.15%. Euclidean fitting also needs 1,440 observations to reach 90.60%, a level that structure-aware fitting attains with 360. Structure-aware fitting therefore substantially reduces the pipeline’s reliance on refinement. Even under the full refinement budget, the gap widens as precision decreases, from 1.25 points at 1.5 L/A bpw to 19.65 points at 1 L/A bpw (Table R3).

Table R3. LIBERO average success (%) with structure-aware and Euclidean fitting under the full refinement budget. The 1 L/A bpw Euclidean result is newly evaluated.

L/A bpwStructure-awareEuclideanDifference
1.593.8592.60+1.25
1.2592.7087.65+5.05
183.0563.40+19.65
AuthorsOpenReview ↗

Response to Reviewer 6ARj (2/2)

[W3] Evaluation on other VLA models and real-world settings

To show that SACQ is not limited to a single benchmark or architecture, we additionally evaluated it on another VLA architecture (OpenVLA-OFT) and on LIBERO-Pro, alongside LIBERO and SimplER in the main experiments. These results are included in Appendix F, and we summarize them here for convenience.

Other VLA architecture.

Appendix F.1 (Table 6) evaluates SACQ on OpenVLA-OFT. At 2 V/L bpw, SACQ stays within 1.75 percentage points of BF16 with the lowest memory among the compared methods, and at 1.5 V/L bpw, it outperforms GPTQ-Ternary and ActQuant by more than 50 percentage points.

Table R4. OpenVLA-OFT on LIBERO, average success across the four suites (from Appendix F.1, Table 6). V/L bpw denotes the bits per weight over the vision encoder and language backbone.

MethodV/L bpwAvg. success (%)Memory (GB)
BF1616.0095.8015.203
SACQ3.0095.555.128
GPTQ2.0061.653.625
ActQuant2.0060.602.953
TWLA2.1194.358.960
SACQ2.0094.052.926
GPTQ-Ternary1.580.003.026
ActQuant1.5032.002.433
SACQ1.5085.702.463

Robustness under perturbations.

Appendix F.2 (Table 7) evaluates SACQ-quantized GR00T N1.7 on LIBERO-Pro, which includes five perturbation families (Language, Object, Swap, Task, and Environment) over 10,000 episodes per model.

Table R5. GR00T N1.7 on LIBERO-Pro, average success across the four suites (from Appendix F.2, Table 7).

MethodL/A bpwAvg. success (%)
BF1616.043.67
SACQ3.043.76
SACQ2.043.62
SACQ1.542.62

Real-world validation.

Regarding real-world hardware, we thank the reviewer for this valuable suggestion. We agree that real-robot validation is an important step toward practical deployment, and we plan to evaluate SACQ on real robot platforms in future work.

[W4] Optimization budget of the baselines

We understand the reviewer’s concern about the fairness of the comparison, and we will describe the baseline protocol more explicitly in the revised version. All calibration-based baselines (GPTQ, TWLA, and ActQuant) use exactly the same training requests as SACQ, namely 576 per suite on LIBERO and 512 on SimplER, and ActQuant, which requires action supervision, uses the same BF16 teacher actions as SACQ. HQQ does not use calibration data for quantizing the target weights. Each baseline follows the optimization procedure of its original work, adapted to GR00T only where necessary.

To show that the gains of SACQ do not come solely from gradient-based optimization, we evaluate SACQ with only index reassignment, which uses no gradients. On LIBERO at 1.5 L/A bpw, this setting achieves 81.85%, compared with 31.95% for GPTQ-Ternary and 30.05% for ActQuant (Table R1). On SimplER at 2 L/A bpw, it also outperforms all baselines under both protocols (Table R6). Action-level calibration then further improves the results.

Table R6. SimplER Google Robot, average success (%) at 2 L/A bpw.

MethodEnd-to-end action fine-tuningVMVA
GPTQNo25.5723.25
ActQuantNo46.6138.70
TWLA (2.06 bpw)No52.6848.16
SACQ, index reassignment onlyNo57.9850.91
SACQ, fullYes63.7652.27
BF16–58.9053.45
AuthorsOpenReview ↗

Gentle Reminder: Author–Reviewer Discussion Period Ending Soon

Dear Reviewer 6ARj,

Thank you again for your thoughtful review and valuable feedback.

As the author–reviewer discussion period is approaching its end, we would like to briefly follow up on our rebuttal and additional results.

In our rebuttal, we provided the following analyses and experiments:

Novelty and contributions: We clarified that our contributions lie in introducing structure-aware codebook fitting and integrating complementary techniques into a pipeline that preserves VLA task success at ultra-low precision. We also showed that SACQ substantially outperforms the baselines even without gradient-based action calibration.

Contribution of structure-aware fitting: We demonstrated that the benefits of structure-aware fitting become more pronounced under limited calibration budgets and at more aggressive bit widths, while keeping the subsequent refinement procedure consistent across fitting methods.

Evaluation on other VLA models: We highlighted our OpenVLA-OFT and LIBERO-Pro evaluations, demonstrating effectiveness across different VLA architectures and evaluation settings.

Optimization budget of the baselines: We specified that all calibration-based baselines use the same training data and follow their respective optimization procedures. We also presented results showing that SACQ outperforms GPTQ-Ternary and ActQuant without gradient-based action calibration.

If these results address your concerns, we would greatly appreciate it if you could acknowledge this and consider reassessing your rating. We would be happy to clarify any remaining questions.

Thank you again for your time and consideration.

Best regards,

The Authors

Comments on the paper

1 comment
AuthorsOpenReview ↗

Manuscript Revision Summary

We thank the reviewers for their constructive feedback.

We have uploaded a revised manuscript incorporating the clarifications and additional analyses discussed in our rebuttal. In particular, we have:

  • Added separate refinement-stage ablations and statistical uncertainty analysis (Appendix H).
  • Clarified the codebook-fitting configuration (Appendix A.5), calibration data units and training/validation split (Appendix C.2), and the adaptation of ActQuant to GR00T N1.7 (Appendix F).
  • Reported quantization and calibration costs (Appendix D).
  • Included inference-efficiency results (Section 4.3).

The core methodology and primary experimental results remain unchanged.

We sincerely appreciate the reviewers' time and valuable feedback throughout the discussion.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.