Astrolabe: Steering Forward-Process Reinforcement Learning for Distilled Autoregressive Video Models - Explained Simply | ArXiv Explained