How to Avoid Debate: Scalable AI Safety via Doubly-Efficient Interactive Proofs

This paper studies single-prover interactive proofs for AI safety, aiming to verify powerful-model outputs without relying on two equally capable AI systems or assuming that one debater is truthful. The authors develop doubly-efficient proofs and arguments for oracle-aided computations. Their results apply when the computation is robust to a small fraction of incorrect oracle answers, or when the oracle is a low-degree polynomial. The work provides a theoretical route to interactive verification without debate, but only under structured or noise-tolerant oracle assumptions.