acceptodds
Withdrawn from ICLR 2027

MARK-VLA: Mixed-Precision Action-Risk-Guided Quantization with Mixed-Weight Kernels for Vision-Language-Action Models

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

20 views

Abstract

Vision-Language-Action (VLA) models couple visual perception, language reasoning, and action generation within a unified policy, but their large model size and computationally intensive inference impose substantial memory and latency overheads. We present MARK-VLA, a post-training quantization framework that enables efficient VLA inference through channel-wise mixed-precision quantization and tailored GPU kernel support for mixed-precision weights. Instead of pursuing extreme compression for individual modules, MARK-VLA applies W2/W4/W8 weight quantization with A8 activation quantization across the major perception-to-action stages to reduce the overall memory footprint. We propose a per-module precision allocation method that automatically determines channel-wise precision based on gradient-based action risk evaluated under a quantized reference policy. To account for the discrepancy between the risk estimated for individual quantization decisions and their combined effect, we adopt a progressive precision reduction scheme that gradually lowers the quantization precision. Finally, to translate these fine-grained precision assignments into practical acceleration, we develop fused mixed-weight GPU kernels together with trace-guided per-layer specialization to optimize the complete execution path of quantized Linear layers. Experiments on LIBERO across six VLA policies and three GPU platforms demonstrate up to compression and up to end-to-end speedups, while the largest decrease in average success rate is only 0.82 percentage points. These results demonstrate that action-aware precision allocation and deployment-oriented kernel co-design can jointly achieve compact and efficient VLA inference with minimal accuracy degradation.

Primary area:
applications to robotics, autonomy, planning
Keywords:
Vision-Language-Action Models, Post-Training Quantization, Mixed-Precision Quantization, Efficient Inference

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_je4mOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

MARK-VLA presents a post-training quantization framework for VLA models that combines action-sensitive channel-wise precision allocation with specialized GPU kernels. Experiments cover six VLA configurations on LIBERO and three GPUs, reporting up to 3.62× smaller weight storage and 1.69× faster inference, with a maximum decrease of 0.82 percentage points in average success rate.

Critical strengths
  • The paper addresses activation preprocessing, weight unpacking, mixed-precision computation, and kernel selection, demonstrating measured inference improvements.
  • Evaluation across six VLA configurations and three GPUs supports the applicability of the approach beyond one particular architecture.
Critical weaknesses
  • The algorithmic contribution is incremental and needs more precise positioning. Action-guided channel-wise quantization, adjacent precision reductions, and marginal sensitivity per saved bit are already present in QVLA.
  • In Table 3, replacing the diagonal approximation with the full estimator improves success from 90.20% to 90.60% on GR00T and from 94.65% to 94.75% on Qwen2.5-OFT under per-module progressive allocation. These gains are small, and uncertainty estimates are absent. The W8A8 reference is also not separately ablated against a full-precision reference. The current evidence more strongly supports the allocation strategy than the claimed estimator improvement.
  • The method relies on empirically selected storage budgets for the vision, language, and action modules. Their values, selection procedure, and tuning cost are important for reproducibility and for assessing comparisons against a baseline with a single global budget. Budget sweeps and comparable tuning opportunities for competing methods would strengthen the results.
  • ActQuant is discussed but not evaluated. Preserving selected modules in floating point does not necessarily imply an inferior whole-policy trade-off. A uniform-W4A8 baseline using the same smoothing, rotation, and optimized execution would also help establish the necessity of channel-wise mixed precision. Comparisons should report success, complete-policy storage, and latency across multiple budgets.
Questions for authors
  1. How were module budgets selected, and what validation data and tuning cost were involved?
  2. Which gains remain when QVLA receives the same module budgets and strict precision-staging constraint?
  3. Do reported storage sizes include the actual hardware-specific packed representations?
  4. Can the authors provide repeated evaluations and comparisons against ActQuant and uniform W4A8?
Reviewer_855VOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes MARK-VLA, a post-training quantization framework for Vision-Language-Action models. Activations are quantized to INT8 after smoothing and Hadamard rotation, and each weight channel is assigned W2, W4 or W8 by a gradient-based action risk computed on a quantized reference policy. Precision is lowered progressively inside each module under module-wise storage budgets, and fused mixed-precision GEMM kernels give end-to-end speedup. On LIBERO with six VLA policies and three GPUs, the paper reports up to 3.62× weight reduction and 1.69× speedup over BF16, with at most 0.82 points drop in average success rate.

Critical strengths
  1. The whole policy, including the vision, language and action modules, is quantized. And the quantization is reasonable for VLA in the litmitted resource conditions.

  2. The paper goes beyond simulated compression and provides fused mixed-precision kernels with end-to-end speed on three GPUs.

Critical weaknesses
  1. Limited novelty. Smoothing, Hadamard rotation, gradient-based sensitivity with greedy bit allocation and mixed-precision INT8 kernels are established PTQ components, and action-guided channel-wise allocation comes from QVLA. The changes over QVLA are small, and the full gradient changes little in the deployed setting.
  2. The allocation is not automatic. The per-module storage budgets are set by hand and their values are not reported, yet they matter more than the risk estimator: global allocation drops GR00T N1.5 from 90.6 to 75.3. The comparison with QVLA, which only gets one global budget, is not fair.
  3. Missing simple baselines. There is no uniform W4A8 or W8A8 accuracy with the same smoothing, rotation and kernels, and no simple channel ranking (e.g., weight magnitude). The benefit of action-risk-guided mixed precision over plain quantization is unclear, while the speed is only on par with uniform W8A8.
  4. The evaluation does not support the deployment claim. The paper is motivated by edge robots, but speed is only measured on RTX 4090, RTX 6000 Ada and A800, with no edge device. Accuracy is only on LIBERO in simulation, where the policies are already above 90% and calibration uses LIBERO's own demonstrations, without confidence intervals or real-robot tests.
Questions for authors

na

Reviewer_8hQuOpenReview ↗
Rating
1
Confidence
2
Informativeness
1
Presentation
3
Soundness
2
Read review
Paper summary

The authors propose MARK-VLA mixed precision post-training framework - the key idea and motivation being quantization can be action-guided where mixed-precision choices (2,4,8) bits can be assigned per output channel for the weights and 8 bit for the activation. For some application of the quantization profile across the weights, the pertubatino introduced by the quantization compared to the 8 bit variant is measured which is the risk of quantization introduced at that weight-channel and the gradient accordingly. The risk is then weighted by how much memory it would save. Finally, fused activation preprocessing and specialized mixed-weight GEMM kernels are used with the trace-guided kernel selection to accelerate quantized inference. The authors then show results comparing with standard BF16 training and their proposed method (and QuantVLA) and show the saving in memory (weights) and also the speedup.

  1. The authors do a very poor job of discussing related work - in quantization in general, specifically in LLMs/VLMs/VLAs. Introduction + Related work has a total of 3 relevant citations to this paper - the novelty and the approach DO NOT justify the absence of important relevant work.
  2. A simple search shows papers like MixLLM: LLM Quantization with Global Mixed-precision between Output-features and Highly-efficient System Design from 2024 which highlights similar principles adn techniques, yet are completely missing from the paper.
  3. Table 1 is poorly presented - without any CIs or SDs for accuracy, the results is hard to understand, evaluate and contextualize.
  4. Simple baselines like Quantization aware training which are present in a majority of post-training frameworks are missing - so the gain from BF16 might be obvious but without such results, any contribution stated here are standalone.
  5. Contributions are overstated - the novelty might just be a different domain, but these techniques have largely exist and the lack of a clear discussion makes this look worse.
Critical strengths

N/A

Critical weaknesses

Stated in the summary

Questions for authors

N/A

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

Related papers

Discussion (0)

Sign in to comment.