Mind the Gaps: Mixture-of-Minds for Human Simulation
Anacreon is an audience simulation model introduced in the paper 'Mind the Gaps: Mixture-of-Minds for Human Simulation'. It targets individual-level prediction within a narrow domain, using an authorship embedding, clustering a qualitative corpus, and training adapters on a Gemma 4 12B base. It achieves state-of-the-art ordinal alignment of 0.775 on a large external survey.
Development
- First ReportMind the Gaps: Mixture-of-Minds for Human SimulationarXiv cs.AI
- Current AssessmentThis work indicates a trend toward more granular, individual-level simulation in AI, moving beyond aggregate predictions. It could impact market research, social science, and personalized AI applications. The use of public text to harvest traits raises privacy and ethical considerations. Next signal: adoption by market research firms or integration into survey platforms.Agent Pulse · analysis
The paper 'Mind the Gaps: Mixture-of-Minds for Human Simulation' introduces Anacreon, an audience simulation model designed to predict individual-level responses to survey questions, addressing the gap where statistical methods and LLM simulators capture central tendencies but flatten heterogeneity. Anacreon learns an authorship embedding to separate individuals, clusters a real qualitative corpus around seed people, and trains a dedicated adapter for each cluster (a mixture of minds) on a Gemma 4 12B base. It harvests demographics, psychological traits, and survey responses from public text, augments records with chain-of-emotion, reduces prompt brittleness by shuffling response options, and reduces positive bias by balancing training distribution. On a large external survey, Anacreon achieves state-of-the-art ordinal alignment of 0.775, the individual-level accuracy measure.
Anacreon's approach suggests that individual-level simulation can be improved by combining authorship embeddings with cluster-specific adapters, effectively creating a mixture of experts for different persona types. The use of chain-of-emotion and balanced training distribution addresses common biases in LLM-based simulation. The reported ordinal alignment of 0.775 indicates a significant improvement over central-tendency methods, but the paper notes it is within a narrow domain, so generalization remains uncertain. Next signal: replication on broader domains or benchmarks.
This work indicates a trend toward more granular, individual-level simulation in AI, moving beyond aggregate predictions. It could impact market research, social science, and personalized AI applications. The use of public text to harvest traits raises privacy and ethical considerations. Next signal: adoption by market research firms or integration into survey platforms.
Anacreon could enable more accurate audience simulation for product testing, marketing, and political polling, reducing the need for expensive surveys. It may also enhance personalized user modeling in recommendation systems. The approach could be commercialized as a service or API, but requires careful handling of privacy and bias.
Future work may extend Anacreon to broader domains and larger populations, potentially improving individual-level predictions in various applications. The method could also be combined with other techniques like reinforcement learning from human feedback. However, scalability and generalization remain open questions.