OpenAI reveals development method for 'AI that outputs accurate and easy-to-understand sentences'

A research team at OpenAI has developed a method to improve the comprehensibility of text while maintaining the accuracy of language models, and has published the results.

Prover-Verifier Games improve legibility of language model outputs | OpenAI
https://openai.com/index/prover-verifier-games-improve-legibility/

PROVER-VERIFIER GAMES IMPROVE LEGIBILITY OF LLM OUTPUTS
(PDF file) https://cdn.openai.com/prover-verifier-games-improve-legibility-of-llm-outputs/legibility.pdf

Language models used in chat AI and AI assistants are required to output not only accurate text but also easy-to-understand sentences. However, previous language model development methods had a problem in that increasing accuracy reduces readability, and increasing readability reduces accuracy, making it difficult to achieve both accuracy and readability.

The solution found by the OpenAI research team is to 'train a language model to output 'sentences that can be verified even by weaker models.'' Specifically, they prepare a powerful language model like GPT-4 and a 'weak model that cannot understand difficult sentences,' and repeat the flow of 'having the powerful model output answers to questions and having the weak model check the accuracy of the output of the powerful model' until the check is successful. This achieved both accuracy and readability.

Before applying this method, the language model was asked to solve an elementary school level arithmetic problem: 'Shauna's father is five times older than Shauna, and Shauna is three times older than Aliyah. If Aliyah is three years old, how old is Shauna's father?' The result is as follows. The correct answer, '45,' is derived, but the explanation of the calculation process is difficult to understand.

The answer after applying this method is as follows. 'First, Arya's age is 3 years. Next, Shauna's age is three times Arya's, so she's 9 years old...' The calculation process can be explained in an easy-to-understand way.

The research team claims that by utilizing this technique, 'AI systems can be made less dependent on humans' and 'the reliability and safety of AI-enabled applications can be improved.'