Token Expand-Merge: Training-Free Token Compression for Vision-Language-Action Models

2025-12-10

AI Summary

To address high latency and computational overhead in real-time deployment of large-scale vision-language-action (VLA) models, this paper proposes TEAM-VLA—a training-free, dynamic token compression framework. TEAM-VLA introduces a novel two-stage “expand-merge” paradigm: first, spatially aware token expansion guided by attention hotspots; second, adaptive token merging steered by action semantics via hierarchical clustering and parameter-free differentiable fusion. This design preserves contextual integrity while ensuring action-semantic consistency, requiring zero additional training and enabling plug-and-play integration. Evaluated on the LIBERO benchmark, TEAM-VLA achieves up to 2.3× inference speedup with maintained or improved task success rates (+1.8%), significantly advancing the practical deployment of VLA models on resource-constrained robotic systems.

Technology Category

• Computer Vision: Language and Vision
• Machine Learning: Large Multimodal Models (LMMs)
• Intelligent Robots: Multimodal Perception & Sensor Fusion

Application Category

• Search and Retrieval-Augmented AI: Large language models for search
• Semantics and Knowledge: Methods to enhance, augment, integrate or synergize semantic models such as knowledge graphs and LLMs
• Economics, Online Markets and Human Computation: Cost models of using LLMs in production systems

Abstract

Vision-Language-Action (VLA) models pretrained on large-scale multimodal datasets have emerged as powerful foundations for robotic perception and control. However, their massive scale, often billions of parameters, poses significant challenges for real-time deployment, as inference becomes computationally expensive and latency-sensitive in dynamic environments. To address this, we propose Token Expand-and-Merge-VLA (TEAM-VLA), a training-free token compression framework that accelerates VLA inference while preserving task performance. TEAM-VLA introduces a dynamic token expansion mechanism that identifies and samples additional informative tokens in the spatial vicinity of attention-highlighted regions, enhancing contextual completeness. These expanded tokens are then selectively merged in deeper layers under action-aware guidance, effectively reducing redundancy while maintaining semantic coherence. By coupling expansion and merging within a single feed-forward pass, TEAM-VLA achieves a balanced trade-off between efficiency and effectiveness, without any retraining or parameter updates. Extensive experiments on LIBERO benchmark demonstrate that TEAM-VLA consistently improves inference speed while maintaining or even surpassing the task success rate of full VLA models. The code is public available on href{https://github.com/Jasper-aaa/TEAM-VLA}{https://github.com/Jasper-aaa/TEAM-VLA}

Problem

• Accelerates VLA inference while preserving task performance
• Reduces token redundancy without retraining or parameter updates
• Improves inference speed while maintaining task success rate

Innovation

• Training-free token compression for Vision-Language-Action models
• Dynamic token expansion near attention-highlighted regions
• Action-aware token merging to reduce redundancy