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

A plug-in, phase-adaptive inference framework that accelerates Vision-Language-Action models by deciding when to spend full computation. A lightweight scheduler watches temporal representation similarity, robot-motion cues and episode progress, then allocates compute jointly across the vision encoder, the LLM and the action head — five Vision-LLM compute modes from full recomputation to multi-step temporal reuse, and three denoising modes that reuse intermediate states during stable motion while keeping full refinement for goal-sensitive stages. No retraining of the base model. Up to 2.55× speedup on GR00T and 3.77× on CogACT; on six real-world GR00T tasks it cuts computation 2.18× and lifts control frequency from 13.8 Hz to 26.3 Hz.