arXiv paper: WCM: A World Critic Model for Vision-Language-Action Reinforcement Learning

A new architecture called the World Critic Model (WCM) addresses a core limitation in how reinforcement learning (RL) critics process visual information for robot control. Standard critic models in Vision-Language-Action (VLA) systems typically look at a single camera frame, losing the temporal context needed in partially observable environments. Past attempts to include observation history led to exponential complexity or failed because simply predicting a scalar reward provided too weak a signal to learn how actions unfold over time. WCM, built on a lightweight LeJEPA architecture, solves this by jointly training the critic to predict future visual states and estimate values. This explicitly forces the model's internal representation to capture temporal dynamics. The paper reports that WCM integrates with both on-policy and off-policy RL pipelines and works with leading VLA backbones like Pi0, Pi0.5, and OpenVLA-OFT. In experiments spanning 149 tasks across four benchmarks, WCM reached state-of-the-art performance, showing notable gains in generalization to out-of-distribution settings. The approach was further validated on seven real-world manipulation tasks using off-policy RL,