Mind the Gaps: Mixture-of-Minds for Human Simulation introduces Anacreon, a model designed to simulate individual human behavior within specific domains. While traditional statistical methods and large language models (LLMs) often capture population averages, they frequently struggle to represent the heterogeneity of individuals, often collapsing distinct perspectives into a single, biased output. Anacreon addresses this by using a "mixture of minds" approach to simulate diverse, independent individuals, aiming to provide more accurate aggregate insights.
How the Approach Works
Anacreon functions by partitioning a population into clusters based on authorship. The process begins by training a transformer to create authorship embeddings, which separate individuals based on their public text. The model then assigns records to the nearest seed individual, creating clusters that represent different segments of the population.
For each cluster, the researchers train a dedicated LoRA adapter on a Gemma 4 12B base model. To improve the model's reasoning, each record is augmented with a "chain-of-emotion"—a trace of the emotional states and appraisals triggered by a stimulus—before the model generates a response. During training, the system shuffles response options to reduce prompt brittleness and balances the training distribution to mitigate the positive bias commonly found in LLMs.
Measuring Individual Fidelity
The paper evaluates Anacreon using "ordinal alignment," a metric that measures how closely a model’s predicted responses on Likert-style scales match ground-truth survey data. This metric is calculated as one minus the normalized mean absolute deviation.
On a large, externally sourced survey of SME merchants, Anacreon achieved an ordinal alignment score of 0.775. According to the paper, this result represents the state of the art for this measure, outperforming previous systems such as Twin-2K-500 (0.717) and the digital-twin mega-study (0.748).
Limitations and Observations
The research notes that while Anacreon reduces the positivity bias inherent in base LLMs, a small residual bias remains, with a pooled signed prediction error of +0.437.
The authors also observed that the model’s performance is tied to the quality of the data within each cluster. When clusters contain conflicting statements, the model struggles to determine whether a response should be positive or negative, leading to a "long tail" of clusters that failed to converge and were subsequently pruned. Analysis of the 50 lowest-scoring clusters showed that poor performance was not concentrated in any specific demographic group, suggesting that the model's limitations are driven by internal data quality rather than demographic bias.
Comments (0)
to join the discussion
No comments yet
Be the first to share your thoughts!