Mind the Gaps: Mixture-of-Minds for Human Simulation

Technical Summary: Anacreon – Mixture-of-Minds for Human Simulation

Problem Statement
The paper addresses the persistent "individual–aggregate gap" in human behavior prediction. While statistical methods excel at predicting population-level trends (e.g., insurance risk, mortality rates), they fail to capture individual heterogeneity. Conversely, Large Language Model (LLM) simulators often inherit this limitation: they recover a population's central tendencies but flatten the diversity, uncertainty, and minority positions that define real individuals. Furthermore, standard LLM simulators suffer from two critical failure modes:

1. Prompt Brittleness: Simulated responses shift based on exact wording, framing, or option ordering rather than reflecting stable preferences.
2. Bias: Models tend to over-represent dominant groups, caricature marginalized identities, and exhibit a "sycophantic" positive bias due to Reinforcement Learning from Human Feedback (RLHF).

Existing approaches, such as "silicon samples" or agent-based models, often prioritize aggregate plausibility over individual fidelity, leading to compressed variance and weak correlations with human responses on novel stimuli.

Methodology: The Anacreon Pipeline
Anacreon is an audience simulation model designed for narrow, well-specified domains. It moves away from a single dense model toward a "Mixture-of-Minds" architecture. The pipeline consists of four key stages:

1. Authorship Embedding and Clustering:

   • A seed population is selected from a corpus of public text based on data richness.
   • A transformer is trained using a contrastive authorship-representation objective (InfoNCE) to generate embeddings where individuals are maximally separated.
   • The full corpus is partitioned into clusters based on the nearest seed individual. This creates distinct groups of similar behavioral profiles.

2. Data Augmentation:

   • Attribute Harvesting: An LLM-as-judge extracts demographics and psychological traits (e.g., Big Five) from the text. A proprietary harness uses majority voting across multiple runs to minimize hallucinations, leaving attributes undefined if consensus is low.
   • Survey Mining: Standard survey questions are mined from the text, converting free-text responses into known-answer multiple-choice or Likert-style items.
   • Chain-of-Emotion (CoE): Each record is augmented with a short trace of emotional states and appraisals evoked by the stimulus. This follows a constructive-choice view where people weight context-dependent subsets of attributes.
   • Bias Mitigation: To counter LLM positivity bias, non-neutral questions are included twice in the dataset: once framed positively and once negatively.

3. Mixture-of-Minds Training:

   • Instead of one global model, a dedicated high-rank LoRA adapter is trained for each cluster on top of a Gemma 4 12B base model.
   • During training, the order of response options for survey questions is randomly shuffled to prevent the model from learning positional biases.
   • The system effectively learns approximately 110B parameters across 420 clusters (400 retained after pruning incoherent clusters).

4. Checkpoint Selection:

   • The best model checkpoint is selected by maximizing a composite score balancing three metrics on held-out data: Top-1 accuracy (multiple choice), Ordinal Alignment (Likert scales), and the Perplexity of the Chain-of-Emotion (ensuring the model converges on the emotional trace).

Key Contributions

• Mixture-of-Minds Architecture: The paper introduces a method to cluster a corpus around seed individuals and train specialized adapters for each cluster, preserving heterogeneity rather than averaging it out.
• Chain-of-Emotion Augmentation: It integrates a structured emotional trace (stimulus $\to$ emotion $\to$ response) to ground predictions in the psychological evolution of the individual.
• Robustness Mechanisms: It explicitly addresses prompt brittleness via option shuffling and reduces positive bias via dual-framing (positive/negative) of survey questions.
• Domain Restriction: The work argues for and demonstrates the efficacy of restricting simulation to narrow, well-specified populations (e.g., SME merchants) to ensure predictions are grounded in real data rather than speculative generalization.

Results
Evaluated on a B2B audience model of SME merchants using a large, externally sourced survey:

• Ordinal Alignment: Anacreon achieved a state-of-the-art score of 0.775 on the ordinal alignment metric ($1 - \text{MAD}/R$). This surpasses prior systems including Twin-2K-500 (0.717), Socrates-Qwen-14B (0.740), and the digital-twin mega-study (0.748).
• Top-1 Accuracy: The median cluster reached 0.679 on multiple-choice preference questions, comparable to the 0.657 achieved by interview-grounded agents in prior work.
• Bias and Variance:
  • The model retains a small positive mean error (+0.437 semantic scale positions), significantly reduced from base LLMs but not eliminated.
  • The standard deviation of errors within clusters is approximately 1.85 scale positions. The author notes this is moderately higher than human test-retest reliability (approx. 1.27) but argue that a faithful simulator should reproduce human variance rather than collapse to a point.
  • Poor performance was not concentrated in any specific demographic group, suggesting limitations stem from data quality within clusters rather than demographic bias.
• Convergence: The Chain-of-Emotion perplexity median was near 6.0, indicating the models reliably reconstruct the emotional chains leading to answers within the constrained domain.

Significance and Claims
The paper claims that Anacreon represents a step toward "drawing aggregate insight from faithfully simulated individuals." By grounding a language model in real qualitative data for a narrow domain, it simulates a population well enough to set a new state-of-the-art for individual-level accuracy.

The author emphasizes that the value of this approach lies in its ability to recognize the "edge of its own competence." Because the model is trained on specific clusters, it can identify when a query falls outside the support of its training data and decline to answer, avoiding the "overreach" of general-purpose simulators. The work posits that individual fidelity is the primary objective; once individuals are simulated faithfully, aggregate population distributions can be derived through post-stratification weighting, rather than forcing a single model to learn population distributions directly.

The paper concludes modestly: Anacreon does not dissolve the boundary between aggregate and individual prediction, but it "moves it" by demonstrating that narrow, grounded simulation can achieve high fidelity where broad, ungrounded simulation fails.