World Models
AMD article finds world-model RL gains depend on repair and recipe
AMD says failure-data fine-tuning and PPO fixes improved reproducibility in robot world-model RL, while action-fork errors remained.