VLA-InfoEntropy: A Training-Free Vision-Attention Information Entropy Approach for Vision-Language-Action Models Inference Acceleration and Success

Chuhang Liu; Yayun He; Zuheng Kang; Xiaoyang Qu; Jianzong Wang

cs.CVApr 7, 2026

VLA-InfoEntropy: A Training-Free Vision-Attention Information Entropy Approach for Vision-Language-Action Models Inference Acceleration and Success

Authors: Chuhang Liu, Yayun He, Zuheng Kang, Xiaoyang Qu, Jianzong Wang

Organizations: Ping An Technology (Shenzhen) Co., Ltd., Shenzhen, China · Tsinghua Shenzhen International Graduate School, Tsinghua University, Shenzhen, China

Abstract

Vision-Language-Action (VLA) models integrate visual perception, language understanding, and action decision-making for cross-modal semantic alignment, exhibiting broad application potential. However, the joint processing of high-dimensional visual features, complex linguistic inputs, and continuous action sequences incurs significant computational overhead and low inference efficiency, thereby hindering real-time deployment and reliability. To address this issue, we use image entropy to quantify the grayscale distribution characteristics of each visual token and introduce attention entropy to capture the distribution of attention scores over task-related text. Visual entropy identifies texture-rich or structurally informative regions, while attention entropy pinpoints semantically relevant tokens. Combined with timestep information, these metrics enable a dynamic transition strategy that shifts the model's focus from global visual features to attention-guided local informative regions. Thus, the resulting VLA-InfoEntropy method integrates spatial, semantic, and temporal cues to reduce redundancy while preserving critical content. Extensive experiments show that our method reduces inference parameters, accelerates inference speed, and outperforms existing approaches.

Figures & tables

Fig. 1: The motivation of VLA-InfoEntropy . In traditional VLA inference mode, all tokens are computed indiscriminately. Although this ensures full coverage, many tokens are clearly redundant. The VLA-cache mode [ 27 ] addresses this issue, but its token selection strategy requires further optimization. To this end, we propose the VLA-InfoEntropy method, which identifies the most important tokens in an image, excludes them from the static token set, and then reuses the key-value (KV) cache for efficient computation.

Fig. 2: The overview of the proposed method. Visual Entropy : Compute visual entropy for each image token to quantify its intrinsic information content. Tokens from texture-rich and edge-dense regions are selected as visually salient candidates. Attention Entropy : Extract text–vision attention scores from the cross-attention matrix, compute attention entropy, and identify tokens whose attention is strongly concentrated on the instruction, marking them as task-relevant. VLA-InfoEntropy : Using the timestep $t$ , dynamically adjust the proportions of tokens selected by visual and attention entropy, gradually shifting the focus from global visual context at early steps to task-driven local evidence at later steps. Reuse token : important tokens are removed from the static token set, and the remaining low-information tokens are treated as reusable KV-cache tokens, enabling efficient inference while suppressing redundant computation.

Method	Success Rate (%, $\uparrow$ )					Latency ( $\downarrow$ )	FLOPs ( $\downarrow$ )	Speed up ( $\uparrow$ )
Method	LIBERO-Spatial	LIBERO-Object	LIBERO-Goal	LIBERO-Long	Average	Latency ( $\downarrow$ )	FLOPs ( $\downarrow$ )	Speed up ( $\uparrow$ )
OpenVLA [ 14 ] (baseline)	84.4 $\pm$ 0.9	86.6 $\pm$ 0.8	75.6 $\pm$ 1.0	53.2 $\pm$ 1.4	75.0 $\pm$ 0.6	51.91	1.864	1.00
SparseVLM [ 32 ]	79.8 $\pm$ 0.6	67.0 $\pm$ 1.2	72.6 $\pm$ 1.4	39.4 $\pm$ 0.8	64.7 $\pm$ 1.0	83.39	1.407	1.32
FastV [ 5 ]	83.4 $\pm$ 1.2	84.0 $\pm$ 0.6	74.2 $\pm$ 1.4	51.6 $\pm$ 0.8	73.3 $\pm$ 1.1	53.28	1.864	1.00
VLA-Cache [ 27 ]	83.8 $\pm$ 0.9	85.8 $\pm$ 0.7	76.4 $\pm$ 0.6	52.8 $\pm$ 1.3	74.7 $\pm$ 0.8	34.38	1.355	1.38
SP-VLA [ 15 ]	84.4 $\pm$ 1.0	85.6 $\pm$ 1.6	75.4 $\pm$ 1.8	54.2 $\pm$ 1.4	74.9 $\pm$ 1.3	-	3.100	1.35

TABLE I: Experimental results of different models on the LIBERO benchmark. Boldface indicates the best results.

Fig. 3: The visualization results of the LIBERO . For representative tasks from LIBERO-Spatial, LIBERO-Object, LIBERO-Goal, and LIBERO-Long, the figure illustrates how VLA-InfoEntropy highlights task-relevant regions while suppressing redundant background areas. Important tokens (purple) tend to cluster around manipulable objects and interaction points, whereas reusable tokens (green) are distributed over uninformative regions. These results demonstrate that the proposed entropy-based token selection effectively concentrates computation on task-relevant visual evidence, thereby reducing overall computational cost while accelerating inference.

Method	Success Rate (%, $\uparrow$ )					Latency ( $\downarrow$ )	FLOPs ( $\downarrow$ )	Speed up ( $\uparrow$ )
Method	LIBERO-Spatial	LIBERO-Object	LIBERO-Goal	LIBERO-Long	Average	Latency ( $\downarrow$ )	FLOPs ( $\downarrow$ )	Speed up ( $\uparrow$ )
OpenVLA(baseline)	84.4	86.6	75.6	53.2	75.0	51.91	1.864	1.00
Visual-Entropy	82.0	82.6	73.8	42.6	69.8	33.82	1.378	1.42
Attention-Entropy	83.4	87.6	76.4	50.2	73.9	33.24	1.311	1.39
Visual+Attention-Entropy	85.4	86.6	79.2	51.8	75.8	32.26	1.244	1.49
Ours	86.4	87.6	79.4	52.2	76.4	31.25	1.214	1.53

TABLE II: Ablation Study. The importance of visual entropy, attention entropy, and timestep are evaluated separately.

Fig. 4: Hyperparameter Sensitivity. (a) The success rate initially rises and then declines as the hyperparameter $T$ increases. (b) The success rate also increases at first and then decreases as the ratio of the hyperparameters $k_{1}/k_{2}$ increases.

Tokens	Success Rate (%, $\uparrow$ )	Latency ( $\downarrow$ )	FLOPs ( $\downarrow$ )
60	71.6	30.89	1.108
80	75.2	31.00	1.157
100	76.4	31.25	1.205
120	75.4	34.49	1.251
140	76.8	36.64	1.297

TABLE III: Comparison of different total number of important tokens

Explore similar work

ATA: Bridging Implicit Reasoning with Attention-Guided and Action-Guided Inference for Vision-Language Action Models
Mar 2, 2026Cheng Yang, Jianhao Jiao, Lingyi Huang +8Vision-Language-Action Models Reasoning for VLA Models
Rutgers University · University College London · Rice University +1
StreamingVLA: Streaming Vision-Language-Action Model with Action Flow Matching and Adaptive Early Observation
Mar 30, 2026Yiran Shi, Dongqi Guo, Tianchen Zhao +8Vision-Language-Action Models Efficient VLA Model Inference
Tsinghua University · Lenovo Group Ltd.
VOTE: Vision-Language-Action Optimization with Trajectory Ensemble Voting
Jul 7, 2025Juyi Lin, Amir Taherin, Arash Akbari +11Robotic Control Efficient VLA Model Inference
Northeastern University, Boston, USA · EmbodyX,San Mateo,USA

Mar 2, 2026J/K move · Enter open · S save

ATA: Bridging Implicit Reasoning with Attention-Guided and Action-Guided Inference for Vision-Language Action Models

Cheng Yang, Jianhao Jiao, Lingyi Huang, Jinqi Xiao+7

Rutgers University · University College London · Rice University+1

Vision-Language-Action Models Reasoning for VLA Models Efficient VLA Model Inference Robustness of VLA Models

Open paper AskPDF arXiv

Vision-Language-Action (VLA) models rely on current observations, including images, language instructions, and robot states, to predict actions and complete tasks. While accurate visual perception is crucial for precise action prediction and execution, recent work has attempted to further improve performance by introducing explicit reasoning during inference. However, such approaches face significant limitations. They often depend on data-intensive resources such as Chain-of-Thought (CoT) style annotations to decompose tasks into step-by-step reasoning, and in many cases require additional visual grounding annotations (e.g., bounding boxes or masks) to highlight relevant image regions. Moreover, they involve time-consuming dataset construction, labeling, and retraining, which ultimately results in longer inference sequences and reduced efficiency. To address these challenges, we propose ATA, a novel training-free framework that introduces implicit reasoning into VLA inference through complementary attention-guided and action-guided strategies. Unlike CoT or explicit visual-grounding methods, ATA formulates reasoning implicitly by integrating attention maps with an action-based region of interest (RoI), thereby adaptively refining visual inputs without requiring extra training or annotations. ATA is a plug-and-play implicit reasoning approach for VLA models, lightweight yet effective. Extensive experiments show that it consistently improves task success and robustness while preserving, and even enhancing, inference efficiency.

Abstract

Explore similar work

ATA: Bridging Implicit Reasoning with Attention-Guided and Action-Guided Inference for Vision-Language Action Models

StreamingVLA: Streaming Vision-Language-Action Model with Action Flow Matching and Adaptive Early Observation

VOTE: Vision-Language-Action Optimization with Trajectory Ensemble Voting