English

Beyond Attention Magnitude: Leveraging Inter-layer Rank Consistency for Efficient Vision-Language-Action Models

Computer Vision and Pattern Recognition 2026-03-27 v1 Computation and Language

Abstract

Vision-Language-Action (VLA) models excel in robotic manipulation but suffer from significant inference latency due to processing dense visual tokens. Existing token reduction methods predominantly rely on attention magnitude as a static selection. In this work, we challenge this assumption, revealing that high-attention tokens are task-dependent and can even degrade policy performance. To address this, we introduce \textbf{TIES} (\textbf{T}au-guided \textbf{I}nter-layer \textbf{E}fficient \textbf{S}election), a dynamic framework guided by inter-layer token ranking consistency. By adaptively balancing attention magnitude with ranking consistency, TIES ensures robust token selection without requiring additional training. On the CogACT + SIMPLER benchmark, TIES improves average success rates by 6\% while reducing token usage by 78\%, and demonstrate strong generalization across diverse decoders and benchmarks.

Keywords

Cite

@article{arxiv.2603.24941,
  title  = {Beyond Attention Magnitude: Leveraging Inter-layer Rank Consistency for Efficient Vision-Language-Action Models},
  author = {Peiju Liu and Jinming Liu and Xipeng Qiu and Xuanjing Huang},
  journal= {arXiv preprint arXiv:2603.24941},
  year   = {2026}
}

Comments

10 pages, 7 figures, preprint

R2 v1 2026-07-01T11:38:19.221Z