ElegantVLA Learning When to Think for Efficient Vision-Language-Action Models

Ye Li1, Huanan Liu1, Kangye Ji1, Yuan Meng1*, Jiajun Fan2, Yuansong Wang1, Shiyu Qin1, Chenglei Wu1, Shu-Tao Xia1, Zhi Wang1*
1Tsinghua University   2University of Illinois Urbana-Champaign
*Corresponding authors
NeurIPS 2026

Abstract

Vision-Language-Action (VLA) models have emerged as a powerful paradigm for generalist robotic control. However, their high computational cost and limited control frequency hinder real-time manipulation, especially when large vision-language backbones and iterative action heads are executed at every control step. Existing acceleration methods mainly rely on static compression or input-level reduction, overlooking the non-uniform reasoning demands of sequential embodied control.

ElegantVLA is a plug-in phase-adaptive inference framework that accelerates VLA models through intra-model dynamic compute scheduling. A lightweight scheduler observes temporal representation similarity and robot states, then jointly allocates computation across the vision encoder, LLM, and action head. For perception-language reasoning, it selects among full computation, partial recomputation, and temporal reuse; for action generation, it reuses intermediate action-head states during stable motion while retaining full computation for precision-sensitive stages.

Extensive experiments on GR00T, CogACT, and real-world tasks show that ElegantVLA preserves or improves task success while substantially accelerating inference. It achieves up to 2.55x average speedup on GR00T and 3.77x average speedup on CogACT. In GR00T-based real-world experiments across six tasks, it reduces computation by 2.18x and increases control frequency from 13.8 Hz to 26.3 Hz.

Framework

ElegantVLA framework overview.

Experimental Results

Detailed task-level results are grouped by benchmark for easier inspection.

Cell format: success rate / FLOPs speedup.

CogACT Visual Matching results

Method PickCan MoveNear Drawer PutDrawer Avg.
CogACT 91.30 / 1.00 85.00 / 1.00 71.80 / 1.00 50.90 / 1.00 74.80 / 1.00
Random Dropping 9.70 / 1.20 20.40 / 1.20 53.50 / 1.20 0.00 / 1.20 20.90 / 1.20
FastV 82.67 / 1.31 72.92 / 1.31 66.67 / 1.28 46.30 / 1.30 67.14 / 1.30
VLA-Cache 87.33 / 1.47 77.92 / 1.46 69.44 / 1.44 47.22 / 1.47 70.48 / 1.46
MoLe-VLA 86.40 / 1.53 80.20 / 1.53 70.60 / 1.53 50.40 / 1.53 71.90 / 1.53
ElegantVLA 89.67 / 3.32 80.42 / 3.05 89.35 / 4.10 50.93 / 4.41 77.59 / 3.72

CogACT Variant Aggregation results

Method PickCan MoveNear Drawer PutDrawer Avg.
CogACT 89.60 / 1.00 80.80 / 1.00 28.30 / 1.00 46.60 / 1.00 61.30 / 1.00
Random Dropping 4.00 / 1.20 16.10 / 1.20 15.60 / 1.20 0.00 / 1.20 8.90 / 1.20
FastV 76.00 / 1.29 71.67 / 1.29 24.34 / 1.26 47.62 / 1.26 54.91 / 1.28
VLA-Cache 80.00 / 1.47 75.00 / 1.45 30.69 / 1.43 43.39 / 1.43 57.27 / 1.45
MoLe-VLA 89.20 / 1.49 79.50 / 1.49 29.90 / 1.49 46.20 / 1.49 61.20 / 1.49
ElegantVLA 82.30 / 3.32 77.17 / 3.26 73.54 / 4.34 57.14 / 4.15 72.54 / 3.77

Cell format: success rate / FLOPs speedup.

GR00T Google Robot main comparison results

Method CloseD MoveNear OpenD PickCan PickObj PlaceD Avg.
GR00T 82.00 / 1.00 92.00 / 1.00 49.50 / 1.00 99.50 / 1.00 92.50 / 1.00 11.00 / 1.00 71.08 / 1.00
FastV 80.00 / 1.13 88.00 / 1.13 64.00 / 1.13 94.00 / 1.13 88.00 / 1.13 12.00 / 1.13 71.00 / 1.13
VLA-Cache 72.00 / 1.13 90.00 / 1.13 62.00 / 1.13 94.00 / 1.13 84.00 / 1.13 10.00 / 1.13 68.67 / 1.13
ElegantVLA 81.00 / 2.39 95.50 / 1.90 74.50 / 2.31 98.00 / 2.35 88.50 / 2.59 12.50 / 2.39 75.00 / 2.35

GR00T WidowX+Bridge main comparison results

Method Carrot CloseD OpenD Basket Sink Spoon Stack Avg.
GR00T 60.50 / 1.00 98.50 / 1.00 93.00 / 1.00 37.50 / 1.00 37.50 / 1.00 71.50 / 1.00 7.00 / 1.00 57.93 / 1.00
FastV 58.00 / 1.13 98.00 / 1.13 84.00 / 1.13 22.00 / 1.13 30.00 / 1.13 74.00 / 1.13 12.00 / 1.13 54.00 / 1.13
VLA-Cache 72.00 / 1.13 98.00 / 1.13 84.00 / 1.13 24.00 / 1.13 24.00 / 1.13 72.00 / 1.13 12.00 / 1.13 53.14 / 1.13
ElegantVLA 63.50 / 2.58 99.50 / 2.54 85.00 / 2.56 31.50 / 2.52 41.00 / 2.57 72.50 / 2.55 13.50 / 2.55 58.07 / 2.55

Trial-level success counts

Method Phone Pen Bowls Pineapple Toast Chocolate Total Rate
Full GR00T 6/10 7/10 4/10 7/10 7/10 6/10 37/60 61.67%
ElegantVLA 6/10 7/10 4/10 8/10 8/10 6/10 39/60 65.00%
Improvement +0 +0 +0 +1 +1 +0 +2 +3.33 pp

Latency breakdown

Task Success Full GR00T ElegantVLA Reduction Latency speedup FLOPs speedup
Phone-on-Stand 60.0% 72.31 ms 37.70 ms 47.9% 1.92x 2.18x
Pen-to-Holder 70.0% 71.74 ms 36.20 ms 49.5% 1.98x 2.24x
Stack-Bowls 40.0% 71.92 ms 41.30 ms 42.6% 1.74x 2.01x
Pineapple Bun 80.0% 72.74 ms 35.70 ms 50.9% 2.04x 2.31x
Toast 80.0% 72.88 ms 37.40 ms 48.7% 1.95x 2.21x
Chocolate Bun 60.0% 73.05 ms 39.70 ms 45.7% 1.84x 2.11x
Average 65.0% 72.44 ms 38.00 ms 47.5% 1.91x 2.18x

Visualizations

Real-world video comparisons

Phone on Stand

GR00T Original ElegantVLA

Pen Holder

GR00T Original ElegantVLA

Stack Bowls

GR00T Original ElegantVLA

Pineapple Bun

GR00T Original ElegantVLA

Toast

GR00T Original ElegantVLA

Chocolate Bun

GR00T Original ElegantVLA

Task-level simulation scheduling rollouts

Google Robot Place in Closed Drawer scheduling visualization.
Google Robot: Place in Closed Drawer
Google Robot Close Drawer scheduling visualization.
Google Robot: Close Drawer
Google Robot Move Near scheduling visualization.
Google Robot: Move Near
Google Robot Open Drawer scheduling visualization.
Google Robot: Open Drawer
Google Robot Pick Coke Can scheduling visualization.
Google Robot: Pick Coke Can
Google Robot Pick Object scheduling visualization.
Google Robot: Pick Object
WidowX Carrot on Plate scheduling visualization.
WidowX+Bridge: Carrot on Plate
WidowX Close Drawer scheduling visualization.
WidowX+Bridge: Close Drawer
WidowX Open Drawer scheduling visualization.
WidowX+Bridge: Open Drawer
WidowX Eggplant in Basket scheduling visualization.
WidowX+Bridge: Eggplant in Basket
WidowX Eggplant in Sink scheduling visualization.
WidowX+Bridge: Eggplant in Sink
WidowX Spoon on Towel scheduling visualization.
WidowX+Bridge: Spoon on Towel
WidowX Stack Cube scheduling visualization.
WidowX+Bridge: Stack Cube

Code

The GR00T-N1.6 implementation, a trained scheduler for the SimplerEnv Google Robot tasks, and the evaluation and two-stage training scripts are available at github.com/ChildTang/ElegantVLA. Quick start:

git clone --recurse-submodules https://github.com/ChildTang/ElegantVLA.git
cd ElegantVLA
# set up the GR00T + SimplerEnv environment (see README), then:
bash scripts/install_gr00t_integration.sh
pip install -e .
MODEL_PATH=/path/to/GR00T-N1.6-fractal bash elegantvla/gr00t/scripts/eval_google_robot.sh

BibTeX

@article{li2026elegantvla,
  title={ElegantVLA: Learning When to Think for Efficient Vision-Language-Action Models},
  author={Li, Ye and Liu, Huanan and Ji, Kangye and Meng, Yuan and Fan, Jiajun and Wang, Yuansong and Qin, Shiyu and Wu, Chenglei and Xia, Shu-Tao and Wang, Zhi},
  journal={arXiv preprint arXiv:2605.29438},
  year={2026}
}