Abstract:

Vision-Language-Action (VLA) models have advanced rapidly, yet most gains still come from larger backbones, larger embodied datasets, or expensive iterative decoders, while action representation remains under-explored. We argue that action space design is a primary bottleneck and present LEAP (Latent-Enhanced Action Prototyping)-VLA, a two-stage framework that learns a structured action latent space from demonstrations via Multi-level Soft Residual Quantization (MSRQ), then trains a lightweight VLM-external aligner to predict actions by prototype-aware soft selection in that space. This design replaces discrete code assignment and multi-step denoising with fully differentiable single-pass latent prediction, preserving pretrained visionlanguage alignment without modifying the VLM. Empirically, LEAPVLA outperforms the state-of-the-art methods with substantially fewer trainable parameters and no embodied pretraining, showing that a wellstructured continuous action latent space can offset model scale and data requirements.