Prompts to Proxies: Emulating Human Preferences via a Compact LLM Ensemble

Technical Summary: Prompts to Proxies (P2P)

Problem Statement
Large Language Models (LLMs) are increasingly utilized as proxies for human subjects in social science research to simulate survey responses and model community dynamics. However, the external validity of these synthetic agents is compromised by a fundamental alignment challenge: ensuring that the aggregate preferences of a synthetic ensemble faithfully reflect the heterogeneity of a target human population. Existing approaches face significant trade-offs. Supervised Fine-Tuning (SFT) and Reinforcement Learning from Human Feedback (RLHF) often collapse toward central tendencies, obscuring preference diversity. Conversely, demographic conditioning methods require sensitive personal data and labor-intensive profile construction, yet still fail to guarantee distributional fidelity. Furthermore, many researchers lack the computational resources or data access required for these heavy-weight approaches.

Methodology: Preference Reconstruction Theory
The paper introduces Preference Reconstruction Theory, which formalizes population preference alignment as a representation learning problem inspired by revealed preference theory in economics. The core insight is that population-level preferences can be reconstructed not by finding a single "representative agent," but by constructing a functional basis of diverse proxy agents and recovering population preferences through weighted aggregation.

The authors operationalize this theory via Prompts to Proxies (P2P), a modular, two-stage inference-based system that requires no fine-tuning and no access to sensitive demographic data:

1. Stage 1: Active Endowment Generation (Diverse Agent Pool Construction)

   • Structured Prompting: The system utilizes a hierarchical "attribute bank" containing interpretable factors (e.g., ideology, personality traits, economic dispositions) derived from social science disciplines.
   • Adaptive Sampling: Instead of exhaustive enumeration, P2P employs an entropy-based adaptive sampling strategy. It tracks "question entropy" (response variability) across different attribute modes.
   • Question Patching: To address low-diversity questions, the system dynamically infers new attributes specific to low-entropy questions and generates "mixed modes" to inject diversity.
   • Goal: To construct a diverse pool of agents (endowments) that spans the latent preference space of the topic domain, maximizing response variability.

2. Stage 2: Regression-Based Aggregation (Compact Ensemble Selection)

   • Variable Selection: Given the diverse pool of agents, the system employs L1-regularized regression (Constrained Lasso) to select a compact subset of agents.
   • Optimization: The algorithm solves a constrained optimization problem to find weights for the selected agents such that their aggregate response distributions minimize the loss against observed ground-truth survey data (training set).
   • Generalization: The selected ensemble is validated on a held-out test set to ensure the learned weights generalize to unseen questions, preventing overfitting.

Key Contributions

• Problem Formulation: The paper formalizes population preference alignment as a representation learning problem, introducing the concept that learnable population preferences can be represented by a functional basis of proxy agents with appropriate aggregation weights.
• Modular Alignment System (P2P): A two-stage system combining entropy-guided adaptive sampling for agent generation and constrained Lasso for ensemble selection. It operates without fine-tuning, demographic data, or significant computational resources.
• Empirical Validation: The authors validate P2P across 14 waves of the American Trends Panel (ATP) and the World Values Survey (WVS), demonstrating robust performance across diverse topics and locales.

Results

• Performance on ATP: Across 14 survey waves, P2P achieved an average test Mean Squared Error (MSE) of 0.014, outperforming both a "Vanilla" prompting baseline (0.026) and a demographic-conditioned "PERSONA" baseline (0.029). This represents a 43% improvement over the best baseline.
• Diversity and Cost: The P2P agent pool exhibited higher average entropy (0.53) compared to baselines (0.43–0.45), confirming that response diversity supports preference reconstruction. The cost per survey was approximately $0.80, comparable to the Vanilla baseline and significantly cheaper than PERSONA ($1.30+).
• Cross-Locale Generalization: When tested on the World Values Survey (US, Great Britain, Hong Kong), P2P achieved comparable performance to US results on Western locales (MSE ~0.014–0.015). Performance on Hong Kong was lower (MSE ~0.024), which the authors attribute to the training data bias of LLMs toward Western populations. Error analysis showed that while cross-locale transfer leverages diversity, locale-specific generation captures culturally distinct preferences that a foreign basis cannot fully reconstruct.
• Stress Testing: In a stress test against a distribution-calibrated SFT model trained on pooled data from 46 countries, P2P achieved competitive Jensen-Shannon Divergence (1-JSD) and Majority-Class Prediction Accuracy (MCPA) using less than 3% of the training data and no fine-tuning.

Significance and Claims
The paper claims to offer a practical, resource-efficient platform for interdisciplinary research in social science simulation. By treating alignment as a reconstruction problem rather than a single-agent optimization task, P2P addresses the "invisible resource barrier" faced by academic institutions and independent researchers. The authors position the work as a step toward pluralistic alignment, demonstrating that synthetic agents can faithfully reproduce population-level heterogeneity without relying on sensitive individual data or expensive model retraining. The study suggests that while LLMs have inherent biases, a properly constructed functional basis can mitigate these issues to emulate human choice patterns effectively. The authors modestly note that future work is needed to address non-stationary preferences over time, extend the framework to freeform text outputs, and further investigate generalization across different survey designs to avoid "shortcut learning."