Prover-Verifier Deliberation Improves LLM Selective Prediction

A novel protocol for inference time, termed prover-verifier deliberation (PVD), has been developed to enhance selective prediction in extensive language models. This approach is based on interactive proof theory, where a prover substantiates a proposed answer with verifiable sub-claims, while a verifier poses specific challenges and responds with Accept, Challenge, or Reject. This system enables the reporting of high-confidence responses while allowing for abstention in uncertain scenarios. The primary experiment employs Claude Sonnet 4.6 as the prover and Claude Haiku 4.5 as the verifier, utilizing the GPQA dataset. The protocol is empirically assessed through its coverage-precision dynamics, as formal guarantees of soundness and completeness do not apply due to the limitations of models operating over a noisy channel.