Modern large language models (LLMs) are capable of generating high-quality text fora variety of tasks. However, evaluating whether a text is truly "high-quality" is extremely difficult,and searching for such texts is an even more challenging problem. In this talk, we introducemethods such as Best-of-N sampling and Minimum Bayes Risk (MBR) decoding, which approachtext generation as an optimization problem to search for better texts. We will discuss how theuniversal challenges of search and evaluation are addressed in the context of LLMs.
抄録全体を表示