Anacreon: A Mixture-of-Minds Model for Individual-Level Audience Simulation
A recent paper published on arXiv presents Anacreon, a model for audience simulation that aims to forecast individual-level reactions to inquiries, tackling the shortcomings of large language models (LLMs) that often oversimplify population diversity. The study, titled 'Mind the Gaps: Mixture-of-Minds for Human Simulation' (arXiv:2608.06115), introduces a technique that integrates authorship embeddings, qualitative data clustering, and tailored adapters to reflect individual variances. Built on the Gemma 4 12B base model, Anacreon organizes a real qualitative dataset around seed individuals, developing a specific adapter for each cluster. It gathers demographic data, psychological characteristics, and survey feedback from public texts, enhancing each entry with a chain-of-emotion to minimize prompt brittleness. This model focuses on a specific domain to enhance individual-level accuracy, addressing a persistent challenge in social science and AI.
Key facts
- Paper title: 'Mind the Gaps: Mixture-of-Minds for Human Simulation'
- arXiv ID: 2608.06115
- Announcement type: new
- Introduces Anacreon, an audience simulation model
- Targets individual-level predictions within a narrow domain
- Uses authorship embeddings to separate individuals
- Clusters a real qualitative corpus around seed people
- Trains dedicated adapters for each cluster (mixture of minds)
- Based on Gemma 4 12B base model
- Harvests demographics, psychological traits, and survey responses from public text
- Augments records with chain-of-emotion
- Reduces prompt brittleness by shuffling response options
Entities
Institutions
- arXiv