[ICLR 2026] Learning to Reason without External Rewards - View it on GitHub