LLM-Driven Feature Discovery (opens in new tab)

We would often like to get a qualitative sense of a target model’s behaviors in important distributions (e.g. deployment, RL training, or evals). For example, we might want to , figure out what causes some target behavior to occur, or find between behaviors. In a recent short exploratory project, we tackled this problem via LLM-Driven Feature Discovery. Our method works as follows:Choose a dataset of model transcriptsSplit transcripts into three pieces: user turns, thoughts, and assistant res...