CHASE-VLA: Post-Training Quantization Framework for Vision-Language-Action Models with Chunk-Aware Scale Estimation

Jin Hyun; Jung Gyu Min; Gyuhyun Jung; Youngjoo Lee

cs.CVOct 2, 2026

CHASE-VLA: Post-Training Quantization Framework for Vision-Language-Action Models with Chunk-Aware Scale Estimation

Authors: Jin Hyun, Jung Gyu Min, Gyuhyun Jung, Youngjoo Lee

Organizations: KAIST, Daejeon, South Korea · POSTECH, Pohang, South Korea

Abstract

Vision-Language-Action (VLA) models map visual observations and language instructions to continuous robot actions, but a diffusion-based action expert (AE) poses a key challenge for low-bit post-training quantization (PTQ). The AE is repeatedly invoked across denoising steps and policy queries, where fixed calibration scales can be mismatched with activation ranges that vary with denoising progress and intended motion. We propose CHASE-VLA, a chunk-aware PTQ method that exploits a VLA-specific signal readily available from the policy: the generated action chunk, including its unexecuted future suffix. Rather than relying only on static scale matching for AE layers, CHASE-VLA combines the previously generated chunk as causal action context with denoising step group information to adapt AE activation scales. This enables W4A4 quantization of both MLP and attention projections in the repeated AE without modifying the pretrained policy. On LIBERO, CHASE-VLA achieves 97.3% average success rate on $π_{0.5}$ when both MLP and attention projections in the AE are quantized to W4A4, restoring FP16-level performance. CHASE-VLA also reduces the weight storage of the quantized AE linear layers by 73.4% and their single-chunk memory traffic by 70.9% and 71.2% on $π_{0.5}$ and GR00T N1.6, respectively, with a predictor overhead of at most 1.26% of the saved storage.

Figures & tables

Figure 1: VLA architecture with a diffusion-based AE for action generation.

Figure 2: Activation-scale variation in a diffusion-based VLA AE. (a) Repeated AE calls across denoising steps and policy queries. (b) Maximum absolute activation values across denoising steps and policy queries, which make fixed calibration scales prone to clipping or loss of quantization resolution.

PTQ Category	Scale-Selection Signal	Quantization Scope
LLM/VLM PTQ	Weight/activation statistics	Transformer encoder/decoder
Diffusion PTQ	Denoising step signal	Diffusion denoising blocks
VLA PTQ	Static module calibration	AE MLP only
CHASE-VLA	Previous action chunk	AE MLP
(Proposed)	+ denoising step group	+ attention projections

Table 1: Scale-selection signals and quantization scopes across PTQ categories.

Figure 3: Overview of CHASE-VLA. After each policy query, the generated action chunk, including its unexecuted future suffix, provides causal action context for precomputing activation scales for denoising step groups in the next query.

Method	Precision	AE quant. target		Spa.	Obj.	Goal	Long	Avg.
Method	Precision	MLP	Attn. proj.	Spa.	Obj.	Goal	Long	Avg.
FP16	FP16	–	–	98.5	99.0	97.5	93.5	97.1
QuantVLA	W4A8	✓	–	98.5	98.0	98.0	96.0	97.6
QuantVLA (expanded)	W4A4	✓	✓	98.5	98.5	93.5	90.5	95.3
Q-DiT	W4A4	✓	✓	97.0	97.5	95.0	88.0	94.4
PTQ4DiT	W4A4	✓	✓	97.5	94.5	94.0	88.0	93.5

Table 2: LIBERO success rates on

\pi_{0.5}

Model	Method	Precision	Spa.	Obj.	Goal	Long	Avg.
CogACT	FP16	FP16	87.5	98.0	95.5	91.5	93.1
	QuantVLA (expanded)	W4A4	89.5	57.5	73.5	81.0	75.4
	Q-DiT	W4A4	81.0	73.5	83.5	73.5	77.9
	PTQ4DiT	W4A4	87.5	77.5	80.5	73.0	79.6
	CHASE-VLA	W4A4	86.5	97.0	88.0	90.5	90.5
GR00T N1.6	FP16	FP16	99.0	99.5	97.5	96.0	98.0

Table 3: LIBERO success rates on additional VLA models.

Factor	Tested settings	Mean success (%)
Denoising groups ( $M$ )	1 / 3 / 5 / 10	95.5 / 97.3 / 97.1 / 97.4
Uncertainty margin ( $\alpha$ )	0 / 1 / 3 / 5	96.3 / 96.9 / 97.3 / 94.8
Action context	Prefix / Suffix / Full chunk	95.9 / 96.3 / 97.3
Predictor sharing	Shared / Per-layer	96.4 / 97.3
Calibration trajectories	2 / 4 / 6	96.1 / 96.8 / 97.3

Table 4: Ablation study of CHASE-VLA on

\pi_{0.5}

across the four LIBERO suites. Success rates correspond to the settings in the same order; bold entries mark the default CHASE-VLA configuration. Prefix and suffix denote the executed and unexecuted portions of the previous action chunk.

Figure 4: Covered AE linear layer weight storage and single-chunk memory traffic on $\pi_{0.5}$ and GR00T N1.6, comparing FP16, QuantVLA, and CHASE-VLA.

Model	FP16	W4A8		CHASE-VLA (W4A4)
Model	Latency (ms)	Latency (ms)	Speedup	Latency (ms)	Speedup
$\pi_{0.5}$	26.957	25.585	$1.054\times$	21.069	$1.279\times$
GR00T N1.6	43.003	41.889	$1.027\times$	37.563	$1.145\times$

Table 5: End-to-end action expert latency and speedup on LIBERO, measured on an NVIDIA A100 GPU with batch size one. CHASE-VLA applies W4A4 quantization to all MLP and attention linear projections in the action expert. Speedups are computed relative to the corresponding FP16 baseline.

Model	AE blocks	Denoising steps	Chunk $(H{\times}d_{a})$	Target layers	Breakdown
Model	AE blocks	Denoising steps	Chunk $(H{\times}d_{a})$	Target layers	MLP	Attn. proj.
$\pi_{0.5}$	18	10	$50{\times}7$	126	54	72
CogACT	12	10	$16{\times}7$	48	24	24
GR00T N1.6	32	4	$16{\times}7$	192	64	128

Table 6: AE architectures and quantized AE linear layers.

Model	Predictors	$d_{\mathrm{in}}$	$d_{\mathrm{hid}}$	$M$	$d_{\mathrm{out}}$
$\pi_{0.5}$	126	350	64	3	6
CogACT	48	112	256	3	6
GR00T N1.6	192	112	256	2	4

Table 7: CHASE-VLA predictor architecture with a two-layer MLP.

Method	$\pi_{0.5}$	CogACT	GR00T N1.6
FP16	[95.7, 98.1]	[91.2, 94.7]	[96.8, 98.8]
QuantVLA (expanded)	[93.6, 96.5]	[72.3, 78.2]	[45.9, 52.8]
Q-DiT	[92.6, 95.8]	[74.9, 80.6]	[20.1, 25.9]
PTQ4DiT	[91.6, 95.0]	[76.7, 82.3]	[17.4, 22.9]
CHASE-VLA	[95.9, 98.2]	[88.3, 92.3]	[89.4, 93.2]

Table 8: 95% confidence intervals for average LIBERO success rates (%).

Scale adjustment	Spatial	Object	Goal	Long	Avg.
Prev-query AbsMax	97.0	98.0	96.5	92.0	95.9
AbsMax EMA	97.5	98.5	95.5	92.5	96.0
Motion lookup	97.0	98.5	97.5	92.5	96.4
CHASE-VLA	97.0	99.5	98.0	94.5	97.3

Table 9: LIBERO success rates (%) for predictor-free baselines and CHASE-VLA on

\pi_{0.5}

Evaluation setup	FP16	QuantVLA	CHASE-VLA
Google Robot (VA)	46.54	29.47	46.99
Google Robot (VM)	56.00	43.67	53.33
WidowX (VM)	64.86	59.71	63.14

Table 10: GR00T N1.6 success rates (%) on SimplerEnv. VA denotes Variant Aggregation and VM denotes Visual Matching.

Explore similar work

QuantVLA: Scale-Calibrated Post-Training Quantization for Vision-Language-Action Models
Feb 23, 2026Jingxuan Zhang, Yunta Hsieh, Zhongwei Wan +5Efficient VLA Model Inference Post-Training Quantization
The Ohio State University · University of Michigan · City University of Hong Kong
VLAQuantBench: Closed-Loop Evaluation of Post-Training Quantization for Vision-Language-Action Models
Sep 21, 2026Jiuyi Xu, Qing Jin, Meida Chen +3Mixed-Precision Quantization Efficient VLA Model Inference
Colorado School of Mines · Independent Researcher · University of Central Florida +2
FoldQuantVLA: Native Low-Bit Quantization of Vision-Language-Action Models via Consistent Folding
Sep 21, 2026Hung T. Ho, Khanh D. Nguyen, Quang D. Nguyen +5VLM Quantization Efficient VLA Model Inference
VinRobotics, Vietnam · AICV Lab, EECS Department, University of Arkansas · School of Advanced Manufacturing and Robotics, Peking University, China +2

Feb 23, 2026J/K move · Enter open · S save

QuantVLA: Scale-Calibrated Post-Training Quantization for Vision-Language-Action Models

Jingxuan Zhang, Yunta Hsieh, Zhongwei Wan, Haokun Lin+4

The Ohio State University · University of Michigan · City University of Hong Kong

Efficient VLA Model Inference Post-Training Quantization VLM Quantization Vision-Language-Action Models

Open paper AskPDF arXiv

Vision-language-action (VLA) models unify perception, language, and control for embodied agents but face significant challenges in practical deployment due to rapidly increasing compute and memory demands, especially as models scale to longer horizons and larger backbones. To address these bottlenecks, we introduce QuantVLA, a training-free post-training quantization (PTQ) framework that, to our knowledge, is the first PTQ approach for VLA systems and the first to successfully quantize a diffusion transformer (DiT) action head. QuantVLA incorporates three scale-calibrated components: (1) a selective quantization layout that integerizes all linear layers in both the language backbone and the DiT while keeping attention projections in floating point to preserve the original operator schedule; (2) attention temperature matching, a lightweight per-head scaling mechanism that stabilizes attention logits and is folded into the dequantization scales at inference; and (3) output head balancing, a per-layer residual interface calibration that mitigates post-projection energy drift. The framework requires no additional training, uses only a small unlabeled calibration buffer, and supports integer kernels for low-bit weights and activations while leaving the architecture unchanged. Across representative VLA models on LIBERO, QuantVLA exceeds the task success rates of full-precision baselines, achieves about 70% relative memory savings on the quantized components, providing a practical pathway toward scalable low-bit embodied intelligence under strict compute, memory, and power constraints.

Abstract

Explore similar work

QuantVLA: Scale-Calibrated Post-Training Quantization for Vision-Language-Action Models

VLAQuantBench: Closed-Loop Evaluation of Post-Training Quantization for Vision-Language-Action Models

FoldQuantVLA: Native Low-Bit Quantization of Vision-Language-Action Models via Consistent Folding