Behavioral State World Modeling

Motivation

This project builds on emerging literature which explores the idea that evolving human mental states can be inferred from observable inputs and represented graphically as partially observable HMMs. My masters thesis attempts to do this with multimodal data (audio and text for now). I am also working on a second version of this project with Yige Yuan, who is using ARPA-H grant-provided health sensor, location, text data, etc. to predict a user’s behavioral or mental state. This work falls under the broader umbrella of AI safety for mental health. We hope that the research conducted here will inform safer LLM-evaluation across multiple turns, especially as LLMs interact with vulnerable users and manifest many different kinds of harm.