Virtual Personas for Language Models via an Anthology of Backstories

Anthology is a method to enhance large language models (LLMs) by conditioning them with rich, detailed backstories to create representative and diverse virtual personas. This approach allows LLMs to simulate individual human responses more accurately, which is valuable for user research and social science surveys. Anthology's generated backstories help capture nuanced personal identities, leading to virtual personas that closely approximate real human demographics and opinions. The work shows promising results in improving the fidelity of approximating public opinion polls, but also highlights challenges such as potential biases and privacy issues.