Compositional Generalization

Composing learned skills into novel multi-step tasks never demonstrated -- the frontier thesis of pi0.7 (2026), and the honest edge where open policies still trail proprietary ones on real-world zero-shot.