acceptodds
Under review as a conference paper at ICLR 2027

DyQ-VLA: Dynamic Quantization and Compensation for VLA models with Runtime Error Awareness

Abstract

Vision-language-action (VLA) models achieve favorable task performance, yet runtime errors in closed-loop execution evolve with alternating updates of actions and observations. Existing VLA quantization methods mainly trade off inference speed and model performance while rarely investigating how quantization alters runtime errors. By comparing error distributions between full-precision and quantized models, we identify distinct evolutionary patterns for the two types of errors: quantization enlarges the variance of translation error distributions and increases their dispersion, whereas the distribution center of rotation errors gradually shifts across execution steps, demonstrating cumulative drift. Motivated by this observation, we rethink the optimal quantization strategy for VLA models and propose DyQ-VLA, a runtime-error-aware quantization framework, which dynamically selects activation precision according to execution steps and tracks as well as compensates rotation errors via accumulated quantization residuals. Corresponding operators and runtime adaptation strategies are devised within the framework to enable dynamic-precision execution. Experiments show that DyQ-VLA achieves a 1.88 to 1.93 inference speedups while maintaining comparable or higher average task success rates. Moreover, it reduces mean execution steps by 17.7% to 19.8%. Our code is here: https://anonymous.4open.science/r/DyQ-VLA-7F51/.

Primary area:
infrastructure, software libraries, hardware, systems, etc.
Keywords:
Dynamic Quantization,Error Control,VLA Models,Compensation

Reviews

Among:
Shape:
1234this paper, Top 15%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.00 over 3 reviews, above 85% of the others.
Reviewer_NBTdOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper presents DyQ-VLA, a runtime-error-aware quantization framework for Vision-Language-Action models. Rather than optimizing only the conventional accuracy–latency trade-off, the paper studies how quantization affects closed-loop execution errors over time. The authors identify different behaviors for translational and rotational errors: quantization increases the dispersion and response sensitivity of translation errors, while rotation errors exhibit relatively concentrated distributions but progressive drift across execution steps. Based on these observations, DyQ-VLA dynamically selects activation precision during execution to control translation errors and uses accumulated quantization residuals to compensate for rotational drift. The method is evaluated on π0.5 and GR00T-N1.5 across LIBERO, RoboTwin 2.0, and real-world manipulation tasks, showing competitive success rates together with reduced execution steps and improved rollout efficiency.

Critical strengths

The main strength of this work is that it introduces a meaningful and relatively novel perspective for VLA quantization. Instead of treating each inference step independently, the paper explicitly considers the closed-loop interaction between execution errors, observations, and subsequent actions, and further discovers distinct evolution patterns for translational and rotational errors. This analysis directly motivates the two components of DyQ-VLA, making the method well connected to the empirical findings rather than being an ad hoc combination of techniques. The paper also provides reasonably comprehensive evaluations across multiple VLA backbones, simulation benchmarks, and real-world tasks, and the reduction in rollout steps is particularly relevant because it shows that per-step acceleration alone is insufficient for evaluating VLA deployment efficiency.

Critical weaknesses

My main concern is the definition of runtime error and, consequently, the interpretation of the paper’s central observations. The paper measures runtime error as the deviation between the executed end-effector trajectory and a matched demonstration/reference trajectory. However, many robotic manipulation tasks admit multiple valid trajectories, and a trajectory that deviates substantially from the demonstration may still be equally correct or even more efficient. Therefore, the measured translation and rotation errors are only proxies for undesirable execution behavior rather than direct measures of task error. Since the key motivation of DyQ-VLA relies on the observed distributional patterns of these errors, it would be important to better justify why deviation from a reference trajectory is an appropriate optimization target. In addition, the reported execution steps and errors are averaged only over successful episodes, which may introduce selection bias when different methods succeed on different subsets of rollouts. 44616_DyQ_VLA_Dynamic_Quantiza

Questions for authors

How sensitive are the proposed error patterns and the resulting dynamic quantization policy to the choice of reference trajectory? In particular, if a task has multiple valid execution trajectories, would DyQ-VLA still make the same precision-selection and compensation decisions when a different successful demonstration is used as the reference? It would be useful if the authors could clarify whether the observed translation-dispersion and rotation-drift phenomena persist under alternative valid references, or provide a reference-independent evaluation of execution quality.

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

This paper studies low-bit post-training quantization (PTQ) for closed-loop Vision-Language-Action (VLA) models. Existing methods overlook how quantization-induced runtime errors trigger extra online corrections, increasing total execution steps and offsetting single-step speedups. Analyzing perturbed action–error distributions, the authors find that quantization amplifies high-variance divergence and single-step sensitivity in translation errors, while causing unidirectional cumulative drift in rotation errors. To address this, they propose DyQ-VLA with fixed 4-bit weights (W4): it dynamically selects {2,4,8,16}\{2, 4, 8, 16\}-bit activation precision via online sensitivity estimation to suppress translation variance, and projects activation quantization residuals through a calibrated Jacobian to compensate for rotational drift. Combined with weight-sharing dynamic-precision CUDA kernel fusion and CUDA Graph replay, DyQ-VLA achieves a 1.59×∼1.93×1.59\times \sim 1.93\times task-level speedup and reduces execution steps by 17.7%∼19.8%17.7\% \sim 19.8\% across LIBERO, RoboTwin 2.0, and real-world Agilex Piper-X experiments.

Critical strengths

1.Insightful shift from single-step latency to closed-loop efficiency: The paper captures a key nuance of embodied control: even when task success rates hold, quantization-induced trajectory deviations trigger extra corrective steps that negate single-step speedups—as shown in Table 2, where ActQuant and HoloQ-VLA drop from 1.10×∼1.17×1.10\times \sim 1.17\times per-step speedups to 0.92×∼0.99×0.92\times \sim 0.99\times across full rollouts. Decoupling action errors into translation variance and rotational drift with targeted remedies yields a clear, compelling design.

2.Solid algorithm-to-hardware execution: To avoid weight-repacking and kernel-launch overheads in dynamic quantization, the authors develop custom A2 (4:8 sparse INT4 mma.sp)/A4/A8/A16 kernels sharing a single W4 layout, alongside epilogue fusion and CUDA Graph replay (Sec. 4.3, App. D.2–D.3). Evaluations on π0.5\pi_{0.5} and GR00T-N1.5 across 6 real-world tabletop tasks verify a 1.59×1.59\times end-to-end speedup.

Critical weaknesses

1.Unexplained speed–accuracy tradeoff in Table 5 and missing bit-distribution / W4A2 baselines: In Table 5, switching from static W4A4 to dynamic W4/AdynA_{dyn} ({2,4,8,16}\{2, 4, 8, 16\} bits, without compensation) reduces steps by only 9.6%9.6\% (171.5→155.1171.5 \to 155.1) yet boosts rollout speedup from 1.44×1.44\times to 1.89×1.89\times (~20%20\% faster per step than static W4A4). Achieving faster single-step inference than pure A4 implies Adyn\text{A}_{\text{dyn}} predominantly selects A2. However, this raises a contradiction: why does a mostly 2-bit model achieve much lower translation error (61.8461.84 mm) and higher SR (97.2%97.2\%) than static W4A4 (91.4791.47 mm, 94.1%94.1\%) and W4A8 (74.9474.94 mm, 94.4%94.4\%), even outperforming full-precision BF16 (97.0%97.0\% SR, 159.7159.7 steps)? The paper omits both the selection ratios across {2,4,8,16}\{2, 4, 8, 16\} bits and a static W4A2 baseline.

2.Theoretical inconsistency in outperforming BF16 via quantization-residual compensation: Sec. 3.1–3.2 state that unidirectional rotational drift is inherent to full-precision BF16 (Figs. 2c, 9) and only weakly affected by quantization. Yet the compensation signal δk,n,i=QDQ(xk,n,i)−xk,n,i\delta_{k,n,i} = QDQ(x_{k,n,i}) - x_{k,n,i} (Sec. 3.3, App. B.4) only captures quantization truncation error relative to BF16 (which is identically zero for BF16). If quantization barely alters rotational drift and RtRR_t^R only tracks quantization loss, why does compensating for it allow DyQ-VLA to systematically beat the BF16 baseline in both rotational error and steps (e.g., improving GR00T-N1.5 from 86.1%86.1\% to 88.5%88.5\% SR and reducing steps from 237.2237.2 to 219.8219.8 in Table 1)?

3.Confounded system-vs-algorithm speedups in Table 2 and unclear implementation details: In Table 2, static baselines HoloQ-VLA (W4A4) and QuantVLA (W4A8) achieve only 1.10×∼1.17×1.10\times \sim 1.17\times and 1.23×∼1.32×1.23\times \sim 1.32\times single-step speedups, whereas the authors' own static W4A4 in Table 5 reaches a 1.44×1.44\times rollout speedup (~1.55×1.55\times per step). This indicates that DyQ-VLA's reported speed gains over prior methods largely stem from custom fused epilogue kernels and CUDA Graph replay (Sec. 4.3) rather than the quantization algorithm itself. Furthermore, the paper fails to clarify the granularity of dynamic bit-switching bb (whole-network, action-expert-only, or layer-wise), which layer provides the monitored activations xk,n,i\mathbf{x}_{k,n,i} in Eq. (9), and how unquantized activations are obtained inside fused kernels without extra FP16 overhead.

Questions for authors

1.Bit-selection distribution and static W4A2 baseline (Critical Clarification): What are the exact step-wise selection percentages for {A2, A4, A8, A16} on LIBERO and RoboTwin 2.0, and what are the SR, step count, and speedup of static W4A2 under the same kernel framework? Please explain how W4/Adyn\text{A}_{\text{dyn}} (without compensation) in Table 5 simultaneously achieves faster single-step inference than static W4A4 and higher accuracy than W4A8 and even BF16.

2.Granularity of dynamic bit-switching and residual monitoring (Clarification): Does configuration b∈Bb \in \mathcal{B} set the activation bit-width uniformly across the entire model (VLM backbone + Action Expert) or only for the Action Expert (flow steps)? In addition, which specific layer(s) are monitored for xk,n,ix_{k,n,i} in Eq. (9), and does computing and projecting $ QDQ(x_{k,n,i}) - x_{k,n,i}$ at runtime require writing residuals from registers back to HBM—and what percentage of single-step latency does this overhead add?

3.How quantization residuals correct inherent full-precision drift (Conceptual Clarification): Given that rotational cumulative drift is equally prominent in BF16 and barely altered by quantization (Observation 2), why does compensating purely for activation quantization truncation residuals QDQ(x)−x\text{QDQ}(\mathbf{x}) - \mathbf{x} reduce rotational error and step counts well below the BF16 baseline? Would applying the same compensation via fake QDQ residuals on top of the BF16 model similarly improve BF16 performance? 4.Fair speed comparison under aligned kernels (Fairness of Comparison): When QuantVLA (W4A8) and HoloQ-VLA (W4A4) are deployed under your exact fused-kernel and CUDA Graph backend, what are DyQ-VLA's net single-step and full-rollout speedups over them?

1 comment
Reviewer_GKLaOpenReview ↗

Response to the authors

Thank you for the thorough response and additional clarifications. All of my concerns have been addressed, and I have no further questions.

Reviewer_mnnEOpenReview ↗
Rating
3
Confidence
4
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes DyQ-VLA, a runtime-error-aware quantization framework for vision-language-action models. It combines dynamic activation precision selection with signed residual-based compensation for rotational drift, while keeping quantized weights fixed at 4 bits. The framework is evaluated in simulation and on a real robot, reporting faster inference and fewer execution steps while largely preserving task success rates.

Critical strengths
  • The method is clearly presented and easy to follow. The distinction between dynamic precision selection for translation error control and residual compensation for rotational drift makes the design and its motivation understandable.
  • The evaluation covers both simulation benchmarks and real-world robotic tasks.
Critical weaknesses
  • The real-world evaluation includes only ten trials per task. At this sample size, the reported success-rate differences may be sensitive to trial-to-trial variability, making it difficult to distinguish consistent improvements from noise. Additional rollouts on a few representative tasks would strengthen the empirical evidence.
  • Although dynamic activation precision selection is central to the method, the paper does not report how frequently each precision is selected. In particular, the proportion of queries using A16 versus lower-bit activations remains unclear, limiting understanding of how much the method relies on high-precision computation to maintain performance.
Questions for authors
  • Could the authors increase the number of real-world rollouts for a few representative tasks and report confidence intervals for success rates, together with variability in task completion time? This would help assess whether the observed differences are consistent across trials.
  • Could the authors provide a per-query activation precision trace for representative episodes, along with aggregate selection frequencies for A2, A4, A8, and A16 across rollouts? Please also clarify which configuration is used as the fallback when no candidate satisfies the constraints and how often this occurs.
  • The paper introduces execution-feedback-based precision selection and rotational residual compensation for VLA models. Could the authors clarify which aspects of the precision-selection mechanism are specifically required for VLA execution? This would help isolate the benefit of the VLA-specific feedback mechanism from that of dynamic precision switching itself.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.