Point of Order: Action-Aware LLM Persona Modeling for Realistic Civic Simulation

Researchers developed a reproducible pipeline to transform public Zoom recordings into speaker-attributed transcripts for training LLMs to simulate realistic civic deliberations. The method achieved 67% reduction in perplexity and nearly doubled performance metrics, with human evaluations showing simulations often indistinguishable from real government meetings.

• →New pipeline converts anonymous ASR transcripts into speaker-attributed data with persona profiles and action tags.
• →Three government deliberation datasets released covering appellate courts, school boards, and municipal councils.
• →Fine-tuned LLMs showed 67% perplexity reduction and doubled classifier-based performance metrics.
• →Turing-style evaluations demonstrated simulations often indistinguishable from real deliberations.
• →Method provides scalable approach for complex realistic civic simulation modeling.