✅ ICLR 2026
SP-VLA: A Joint Model Scheduling and Token Pruning Approach for VLA Model Acceleration
Unified framework for real-time VLA inference — model scheduling + token pruning
Ye Li, Yuan Meng, Zewen Sun, Kangye Ji, Chen Tang, Jiajun Fan,
Xinzhu Ma, Shu-Tao Xia, Zhi Wang, Wenwu Zhu
Tsinghua University & UIUC
TL;DR — VLA models are powerful but too slow for real-time robotics. SP-VLA jointly addresses temporal redundancy (via action-aware model scheduling) and spatial redundancy (via spatio-semantic token pruning), achieving significant speedup while maintaining or improving task performance.
⚙️ Two Complementary Mechanisms
🤖 Action-Aware Model Scheduling
Categorizes VLA actions into deliberative (complex, needs full VLA) and intuitive (routine, handled by lightweight generator). Dynamically switches between models to reduce temporal redundancy — inspired by human intuition vs deliberation.
🔍 Spatio-Semantic Token Pruning
Classifies tokens by both spatial importance and semantic relevance, then prunes redundant ones before VLA inference. Reduces spatial redundancy in visual input without harming task-critical information.
📊 Results
1.5×
Lossless speedup
on LIBERO benchmark
+6%
Avg performance gain
over baseline VLA
2.2×
Inference frequency
improvement (SimplerEnv)
📅 Publication Journey
Oct 2025
Submitted to ICLR 2026
Submitted to The Fourteenth International Conference on Learning Representations (Submission #944).
Jan 2026
✅ Accepted at ICLR 2026 (Poster)
Apr 2026
Presented at ICLR 2026 · Rio de Janeiro
📖 Cite This Paper
BibTeX (OpenReview):
@inproceedings{li2026spvla,
title={{SP}-{VLA}: A Joint Model Scheduling and Token Pruning
Approach for {VLA} Model Acceleration},
author={Ye Li and Yuan Meng and Zewen Sun and Kangye Ji and
Chen Tang and Jiajun Fan and Xinzhu Ma and
Shu-Tao Xia and Zhi Wang and Wenwu Zhu},
booktitle={The Fourteenth International Conference on
Learning Representations},
year={2026},
url={https://openreview.net/forum?id=RwdGIIjPlC}
}