PROWL: Prioritized Regret-Driven Optimization for World Model Learning

In three linesPROWL introduces a KL-constrained adversarial curriculum to improve robustness of action-conditioned video world models. A policy exposes high-error trajectories of a diffusion-based model while a Prioritized Adversarial Trajectory (PAT) buffer re-ranks data by prediction error and learning progress. Evaluation on MineRL demonstrates improved robustness on out-of-distribution trajectories.

Summary generated by Claude — human-verified