Recurrent-Depth VLA: Implicit Test-Time Compute Scaling of Vision–Language–Action Models via Latent Iterative Reasoning

Technical Analysis "Recurrent-Depth VLA: Implicit Test-Time Compute Scaling of Vision–Language–Action Models via Latent Iterative Reasoning" is an architecture that achieves computational adaptivity through latent iterative refinement instead of explicit token generation. RD-VLA employs a recurrent action head with weight-tied layers, enabling arbitrary depth with a constant memory footprint . High-level Analogy: Imagine you're trying to solve a complex puzzle. Instead of writing down every single step and option on a piece of paper (which takes a lot of time and space, like a robot using 'token-based reasoning'), you work it out mostly in your head. You start with a rough idea, then refine it, thinking deeper and longer when the puzzle is tough, and quickly arriving at a solution for simpler parts. This 'thinking in your head' where you iteratively refine your internal understanding until you're confident is what RD-VLA does. It repeatedly processes its internal 'thoughts' (latent representations) to improve its understanding and plan, spending more time when a robotic task is difficult and less when it's easy, all without having to 'write down' those intermediate thoughts explicitly. Motivation of the Work Current Vision-Language-Action (VLA) models for robots typically operate with a fixed computational budget. This means they spend the same amount of effort on simple tasks, like adjusting a grip, as they do on complex ones, like navigating a cluttered environment, making…