Virtual Biographer

With this research project, we aim to capture someone’s life story through continuous, adaptive interviews with a Large Language Model (LLM). The LLM guides the conversations based on previously learned information and detects gaps in its knowledge representation.

Project Status

Ongoing

GTC Team

Börge Scheel, Henning Metzmacher, Fraser Rothnie, Prof. Robert W. Sumner, Dr. Fabio Zünd

Collaborators

Prof. Markus Gross (ETH CGL), Chen Yang (ETH CGL)

Introduction

The Virtual Biographer project explores how people’s life stories can be captured and expressed through ongoing conversation. Instead of a single interview or static questionnaire, users engage in short, recurring conversations that gradually uncover memories, relationships, places, and emotions. Over time, these transcripts form a structured and evolving biography, one that reflects not only what someone has experienced but also how they express themselves.

Users can talk with their Virtual Biographer through text or voice. The virtual interviewer listens with empathy, adapts to each person’s preferences, and respects boundaries. The resulting transcripts provide a rich, real-world dataset that can also be used for research on long-term human-AI interaction, narrative understanding, and memory modeling.

One of the main applications of this work lies in the creation of digital humans: Virtual representations of real people that can speak and act in ways that remain faithful to the individual they are based on. Rather than generating made-up or “hallucinated” content, these systems aim to stay grounded in authentic biographical data. Achieving this requires detailed, longitudinal information about people’s experiences, personality, and way of speaking. This is precisely what the Virtual Biographer is designed to collect.

Left: The Virtual Biographer App. Right: Visualization of a person's life as a memory graph. — In the Virtual Biographer app, users can engage in text-based and voice conversations with the virtual interviewer. This gradually builds up a knowledge representation of a person's life. On the right, this dataset is visualized as a 3D memory graph, where each node represents a memory and edges highlight semantic connections between those memories.

Research Questions

In our research, we focus on three key challenges:

Learning what’s missing: How can we recognize gaps in the data we have already collected and ask questions that extend the knowledge of a person’s life?
Adaptation: How can an LLM-based interviewer adjust to the user’s conversational style, interests, and boundaries?
Active engagement: How can the LLM take the initiative, revisiting themes or introducing new topics over time, rather than staying passive in the conversation?