Mitigating Legibility Tax with Decoupled Prover-Verifier Games

Researchers propose a new approach to address 'legibility tax' in AI systems by decoupling solver and verification functions. They introduce a translator model that converts correct solutions into checkable forms, maintaining accuracy while improving verifiability through decoupled prover-verifier games.

• →Current prover-verifier games suffer from accuracy degradation called 'legibility tax' when models are trained for both correctness and checkability.
• →The proposed solution uses a separate translator model to convert correct solutions into verifiable forms without compromising accuracy.
• →The approach allows solver models to focus purely on correctness while translator models handle verification requirements.
• →Decoupled prover-verifier games create equilibria that produce faithful and checkable translations.
• →This method could improve AI safety by making powerful model outputs more verifiable by less capable systems.