Mitigating Legibility Tax with Decoupled Prover-Verifier Games

The paper introduces a decoupled prover-verifier game (DPVG) framework aimed at mitigating the legibility tax in large language models by separating the tasks of correctness and checkability. It proposes training a "translator" model that converts a solver model's output into a checkable format without compromising accuracy, enhancing the overall checkability of model outputs. This approach is significant for practitioners as it provides a method to improve the reliability of model outputs in applications where verification by less capable systems is essential.