Think Twice, Act Once: Token-Aware Compression and Action Reuse for Efficient Inference in Vision-Language-Action Models

TL;DR

This paper introduces FlashVLA, a training-free framework that significantly accelerates vision-language-action model inference by reusing actions and pruning visual tokens, achieving over 55% reduction in FLOPs with minimal accuracy loss.

Contribution

The paper presents FlashVLA, a novel plug-and-play method for reducing inference costs in VLA models through token-aware action reuse and visual token pruning, without retraining.

Findings

Reduces FLOPs by 55.7% on LIBERO benchmark.

Decreases latency by 36.0%.

Maintains 99.3% of task success rate.

Abstract

Vision-Language-Action (VLA) models have emerged as a powerful paradigm for general-purpose robot control through natural language instructions. However, their high inference cost-stemming from large-scale token computation and autoregressive decoding-poses significant challenges for real-time deployment and edge applications. While prior work has primarily focused on architectural optimization, we take a different perspective by identifying a dual form of redundancy in VLA models: (i) high similarity across consecutive action steps, and (ii) substantial redundancy in visual tokens. Motivated by these observations, we propose FlashVLA, the first training-free and plug-and-play acceleration framework that enables action reuse in VLA models. FlashVLA improves inference efficiency through a token-aware action reuse mechanism that avoids redundant decoding across stable action steps, and an information-guided visual token selection strategy that prunes low-contribution tokens. Extensive experiments on the LIBERO benchmark show that FlashVLA reduces FLOPs by 55.7% and latency by 36.0%, with only a 0.7% drop in task success rate. These results demonstrate the effectiveness of FlashVLA in enabling lightweight, low-latency VLA inference without retraining.