OpenAI | Prover-Verifier Games improve legibility of language model outputs

Research from OpenAI has shown that training strong language models to produce text that weak models can easily verify also makes the text easier for humans to evaluate. This technique, called “Prover-Verifier Games,” involves two players (a prover and a verifier) checking the correctness of solutions. This ensures that model outputs are not only accurate but also easy to understand and verify.

Source: Prover-Verifier Games improve legibility of language model outputs, July 17, 2024