As AI models continue to develop powerful capabilities, it becomes critical to ensure that their outputs align with our intentions. Recent research focuses on verification via debate, a model of interactive proofs where two competing powerful provers (or AI models) debate each other to convince a weaker verifier (or human) of the correctness of their claims. However, debate assumes that the two AI models possess equal abilities and that one of them is truthful, which may not be realistic.
In this work, we demonstrate how to avoid debate: we initiate the study of single-prover interactive proofs for AI safety. Prior results in single-prover interactive proofs do not immediately transfer to the AI safety setting; for example, they fail when the computation has access to an oracle, such as human judgment or an external database.
We present doubly-efficient single-prover interactive proofs and arguments for oracle-aided computations (also known as relativizing proofs) that are effective in scenarios where (1) the computation is robust, meaning the output does not change if at most a small fraction of the answers to oracle queries are incorrect, or (2) the oracle is a low-degree polynomial. These results suggest that interactive verification is possible even without debate, under structured or noise-tolerant oracle access.
Blogger's Review: This article offers a fresh perspective in the field of AI safety by introducing the single-prover model, circumventing the limitations of traditional debate. It highlights the potential of interactive proofs in practical applications, especially when dealing with AI models of asymmetric capabilities. This research lays a foundation for future AI verification systems and is worth following closely.