Single training technique enables world models and agent self-improvement

Original: World modeling. Faster RL. Self-improvement without verifiers.

Source: x.com ↗