Faithful LLM Persona Synthesis with Multi-Distribution Matching

Using sets of LLM personas as stand-ins for real human populations is a key challenge in LLM-based social simulation. Recent methods use LLMs to generate personas from real social media histories or via structured sampling, and align them to a reference distribution such as Big Five psychometrics. But imagine a synthetic population whose Big Five personality profile is statistically indistinguishable from that of a real survey population. Ask these personas for their opinions on a policy, or observe how they actually behave in a simulated decision, and the match may fall apart: aligning a persona set to one reference distribution guarantees nothing about the others, and recent evidence shows that a persona's self-reported traits often dissociate from its behaviour. Instead of aligning the persona set to a single psychometric distribution, we aim to align it to multiple reference distributions simultaneously — spanning psychometric, opinion, and behavioural responses — and study whether jointly aligned persona sets generalize better to distributions and simulation tasks they were never aligned on. The team is co-led by Dr. Aditya Joshi, a Senior Lecturer in Natural Language Processing (NLP), and Haokai Zhao, a PhD student, in the UNSW-NLP research group.

The ideal student will have strong programming skills in Python. A good grounding in statistics would be highly regarded.