Beyond Demographics: Fine-tuning Large Language Models to Predict Individuals' Subjective Text Perceptions

2025-02-28

AI Summary

This work investigates whether large language models (LLMs) can systematically model associations between demographic attributes and subjective text annotation behavior. Using datasets from five standardized subjective annotation tasks, the study employs fine-grained supervised fine-tuning, demographic-conditioned prompting, cross-task generalization evaluation, and ablation analysis. Results show that performance gains primarily stem from memorizing and fitting individual annotator behavior—not from learning generalizable demographic regularities. Crucially, across all five tasks, LLMs exhibit no robust or interpretable group-level patterns in response to demographic prompts. This is the first systematic examination challenging the feasibility of treating LLMs as “demographically aware annotators.” It reveals a critical risk of individual-level overfitting in current paradigms for subjective annotation modeling and provides both methodological caution and an empirical benchmark for trustworthy subjective modeling.

Technology Category

• Machine Learning: Large Multimodal Models (LMMs)
• Natural Language Processing: (Large) Language Models
• Cognitive Modeling & Cognitive Systems: Analogy

Application Category

• Economics, Online Markets and Human Computation: Humans versus LLMs for data annotation and labeling
• User Modeling, Personalization and Recommendation: Large Language Models (LLM) for user modeling and recommendation
• Semantics and Knowledge: Data modeling to support human-machine intelligence, including LLMs agents, intelligent system behavior, explanations, and user-friendly interactions

Abstract

People naturally vary in their annotations for subjective questions and some of this variation is thought to be due to the person's sociodemographic characteristics. LLMs have also been used to label data, but recent work has shown that models perform poorly when prompted with sociodemographic attributes, suggesting limited inherent sociodemographic knowledge. Here, we ask whether LLMs can be trained to be accurate sociodemographic models of annotator variation. Using a curated dataset of five tasks with standardized sociodemographics, we show that models do improve in sociodemographic prompting when trained but that this performance gain is largely due to models learning annotator-specific behaviour rather than sociodemographic patterns. Across all tasks, our results suggest that models learn little meaningful connection between sociodemographics and annotation, raising doubts about the current use of LLMs for simulating sociodemographic variation and behaviour.

Problem

• Improving LLMs to predict subjective text perceptions
• Training LLMs with sociodemographic data for better accuracy
• Assessing LLMs' ability to simulate sociodemographic variation

Innovation

• Fine-tuning LLMs for subjective text perception prediction
• Training models with sociodemographic data for improved accuracy
• Analyzing annotator-specific behavior over sociodemographic patterns