DIAL Framework Decouples Intent and Action in VLA Models

Researchers have introduced DIAL, a framework for Vision-Language-Action (VLA) models that separates high-level decision making from low-level motor execution. Unlike existing end-to-end VLAs that treat Vision-Language Models (VLMs) primarily as multimodal encoders, DIAL uses a VLM-based System-2 for latent world modeling, synthesizing visual foresight within the VLM's native feature space. This foresight encodes intent and acts as a structural bottleneck. A lightweight System-1 policy then decodes this intent into actions. The approach addresses training instability and underutilization of VLM semantic representations. The paper is available on arXiv under identifier 2603.29844.