Interactive PCP Protocol for Verifying Probabilistic Predictor Consistency

A new arXiv paper (2608.11181) presents a method to verify the consistency of probabilistic predictions made by AI models in polynomial time. The research addresses a key concern in AI safety: ensuring that a model's answers to conditional-probability queries are self-consistent, which is crucial for trust in AI systems. The authors construct an interactive probabilistically checkable proof (PCP) protocol. In this protocol, a predictive model is represented by two circuits: P, which computes probabilities, and Q, which outputs confidence levels. Together, these circuits implicitly encode exponentially many probabilistic claims. The verifier, given the circuits, evaluates them at only a few points and also accesses a proof oracle—an encoding of a probability distribution allegedly consistent with the model's predictions—reading it at a few locations. This allows the verifier to check approximate consistency efficiently. The paper is categorized as a cross announcement and is available on arXiv. The work is significant for AI safety, as it provides a theoretical foundation for verifying that AI systems' probabilistic outputs are reliable and honest, potentially preventing unwanted outcomes.