It also helps models learn from themselves. Swapping temperature rollouts for architectural ones in label-free test-time RL (TTRV) improves accuracy on every benchmark we tried, e.g. +15.3 on A-OKVQA…