OpenAI has released research on developing algorithms to improve the capabilities of large-scale AI language models, or LLMs, in explaining various things more effectively to others. This method is called Prover-Verifier Games.
In this work, researchers use two AI models – one with higher efficiency as the Prover, or the one explaining the answers, and the other with lower efficiency as the Verifier, responsible for verifying the correctness of the answers received. The Prover’s task is to provide answers that make the other party believe they are correct, while the Verifier must respond whether the answers received are correct or not. In many cases, these answers are ambiguous, but the Prover uses explanations to make them seem plausible.
In the research, GPT-4 was used to enhance the abilities of both models to answer school-level math questions with clear answers. The Prover was adjusted to provide correct answers with good explanations and intentionally misleading explanations to test the Verifier, who then scored the responses.
After at least four training rounds, it was found that the Verifier performed better when faced with misleading answers, while the Prover improved in explaining things in a more easily understandable way continuously.
OpenAI stated that this testing was done to address issues with LLMs being able to provide information on complex topics but highlighted the importance of humans being able to verify whether these answers are trustworthy. Using a dual-model system to cross-check can streamline the process and reduce the risk of human involvement.
Source: OpenAI
TLDR: OpenAI research focuses on improving AI language models through Prover-Verifier Games, where one model explains answers and the other verifies them, leading to better explanations and verification processes.
Leave a Comment