Beyond Marginal Distributions: A Framework to Evaluate the Representativeness of Demographic-Aligned LLMs

TL;DR

The paper tackles the challenge of achieving population-representative value alignment in LLMs by arguing that aligning marginal response distributions is insufficient to capture real-world heterogeneity. It introduces a framework that jointly evaluates marginal distributions and multivariate correlation patterns against World Values Survey data, enabling assessment of both item-level fit and higher-order cognitive structures. Through a systematic comparison of persona prompting and demographic fine-tuning (OpinionGPT), the study finds that while fine-tuning improves marginal similarity, neither approach fully preserves latent correlation structures, highlighting representativeness as a distinct alignment axis. The results call for alignment methods that explicitly incorporate population-level dependencies to avoid optimistic conclusions based solely on marginals, and point to future work leveraging richer dependency-aware approaches and broader demographic coverage.

Abstract

Large language models are increasingly used to represent human opinions, values, or beliefs, and their steerability towards these ideals is an active area of research. Existing work focuses predominantly on aligning marginal response distributions, treating each survey item independently. While essential, this may overlook deeper latent structures that characterise real populations and underpin cultural values theories. We propose a framework for evaluating the representativeness of aligned models through multivariate correlation patterns in addition to marginal distributions. We show the value of our evaluation scheme by comparing two model steering techniques (persona prompting and demographic fine-tuning) and evaluating them against human responses from the World Values Survey. While the demographically fine-tuned model better approximates marginal response distributions than persona prompting, both techniques fail to fully capture the gold standard correlation patterns. We conclude that representativeness is a distinct aspect of value alignment and an evaluation focused on marginals can mask structural failures, leading to overly optimistic conclusions about model capabilities.

Figures (7)

• Figure 1: Overview of our suggested framework. Given a set of survey questions $Q$ (top), we compare marginal responses of one human subgroup $s \in S$ and a steered model $m \in M$ for each question (left) as well as the correlation structures across subgroups $S$ and across models $M$ over all questions (right).
• Figure 2: Process for constructing the marginal distributions of human opinions (green) and different steered models (yellow).
• Figure 3: Process for constructing a correlation matrix from human opinions (green) and comparing it with a simulated correlation matrix (yellow).
• Figure 4: Evaluation metrics by demographic subgroup and dimension. Left: mean dissimilarity (lower = better). Right: mean response variance (closer to true data = better).
• Figure 5: System prompt used for all simulations