ANCORA: Self-Play Framework for Verifiable Reasoning Without Human Supervision

Researchers propose ANCORA, a novel framework that shifts from learning to answer to learning to question. The system alternates between a Proposer that generates novel specifications and a Solver that produces verified solutions, enabling self-improvement without human supervision. Key mechanisms include a two-level group-relative update coupling advantages, iterative self-distilled SFT projecting onto a valid-output manifold, and a UCB-guided Curriculum DAG that grows only through verified specifications. These stabilizers prevent Proposer collapse under sparse verifier feedback. The work is detailed in arXiv:2604.27644.