Anacreon
How a mixture of minds learns to simulate individuals, not averages. Every point is a record of something a real person said.
Scroll through the pipeline
01 · Corpus
Start with what people say in public
Anacreon trains on unstructured text from many sources. Each record becomes a stimulus and a response: a condition, and what one person did under it.
Public behaviour is a strong signal for private disposition. Anacreon collects what a population makes public and infers the private traits that shape its answers.
02 · Authorship embedding
Learn to tell people apart
A seed population of individuals with rich data is chosen. A transformer is trained with a contrastive authorship objective, pulling two samples of the same person together and pushing different people apart.
The seed people end up maximally separated in the embedding space.
03 · Clustering
Every record finds its nearest person
The trained embedding is applied to the full corpus. Each record is assigned to its nearest seed individual, partitioning the corpus into clusters of people who write and think alike.
04 · Temporal split
Hold out the future
Each cluster is split by time. The newest 20% of records is held out for testing and the rest is used for training, so later information can’t leak into training.
05 · Augmentation
Harvest who they are, and how they feel
An LLM-as-judge extracts demographics and psychological traits from each cluster, with a majority vote across runs. When runs disagree, the attribute is left unknown rather than guessed.
Survey questions are mined from the text, and each is framed positively and negatively to cancel the model’s pull towards “agree”.
Every record gets a chain-of-emotion: a short trace of the appraisals a stimulus evokes, from stimulus to response.
06 · Mixture of minds
One mind per cluster
A dedicated LoRA adapter is trained for each cluster on a shared Gemma 4 12B base. Response options are reshuffled every epoch, so no mind learns to answer by position.
07 · Selection
Keep the minds that converge
Incoherent clusters, with conflicting statements, train poorly. That long tail is pruned. Each mind’s best checkpoint balances choice accuracy, ordinal alignment and chain-of-emotion perplexity:
θ* = argmax α·Amc + β·Aord − γ·PPL08 · Accuracy
State of the art on individuals
On held-out survey questions, the median mind reaches an ordinal alignment of 0.775, ahead of every published system that reports the measure: Twin-2K-500 0.717, Socrates-Qwen-14B 0.740 and the digital-twin mega-study 0.748. Top-1 accuracy on multiple choice: 0.679, against 0.657 for interview-grounded agents.
The weakest clusters aren’t concentrated in any demographic group, so performance tracks data quality, not who a cluster represents.
Evaluated on a B2B audience of SME merchants who take in-person card payments. Bars here are illustrative of the paper’s distribution.