This is part 1 of a series based on the paper Chatting Up Attachment: Using LLMs to Predict Adult Bonds.

TLDR on the paper: the researchers use LLMs to generate synthetic Adult Attachment Interview transcripts, use the transcripts to train classifiers, and then evaluate the classifiers against genuine human responses to the same questions.

Why would they want to do this? Human data is expensive and maybe ethically & legally risky.

My goal here is to replicate the research.

(just want to see the code? here you go)

attachment styles

the adult attachment interview

paper outline

setting up the llm agents

interviewing the agents

do the interviews seem good?

next steps

  • qualitative assesment of data
    • does it form clusters?
    • does it cluster with human data? after adjustment?
  • i need the embeddings for the human data - email them