Splits! A Flexible Dataset for Evaluating a Model's Demographic Social Inference

2025-04-06

AI Summary

This work addresses the limited capacity of large language models (LLMs) to model demographic-group–specific linguistic variations—e.g., across occupation, religion, or race—in social reasoning. We introduce “Group Theorization”: a novel task requiring models to autonomously generate interpretable, empirically verifiable theories about cross-group expressive differences from neutral-topic Reddit texts. We formally define this task and present Splits!, the first fine-grained benchmark supporting both theory generation and human validation. Splits! features structured demographic grouping, neutral-topic filtering, and alignment with population-level demographic labels; we further propose a human-evaluation–driven, multi-dimensional metric for theory quality—encompassing accuracy, fairness, and interpretability. We publicly release the Splits! dataset and evaluation framework. Empirical results reveal systematic deficiencies in current LLMs across all three dimensions, highlighting fundamental gaps in socially grounded reasoning.

Technology Category

• Machine Learning: Large Multimodal Models (LMMs)
• Natural Language Processing: (Large) Language Models
• Cognitive Modeling & Cognitive Systems: Social Cognition And Interaction

Application Category

• User Modeling, Personalization and Recommendation: Large Language Models (LLM) for user modeling and recommendation
• Search and Retrieval-Augmented AI: Large language models for search
• Semantics and Knowledge: Data modeling to support human-machine intelligence, including LLMs agents, intelligent system behavior, explanations, and user-friendly interactions

Abstract

Understanding how people of various demographics think, feel, and express themselves (collectively called group expression) is essential for social science and underlies the assessment of bias in Large Language Models (LLMs). While LLMs can effectively summarize group expression when provided with empirical examples, coming up with generalizable theories of how a group's expression manifests in real-world text is challenging. In this paper, we define a new task called Group Theorization, in which a system must write theories that differentiate expression across demographic groups. We make available a large dataset on this task, Splits!, constructed by splitting Reddit posts by neutral topics (e.g. sports, cooking, and movies) and by demographics (e.g. occupation, religion, and race). Finally, we suggest a simple evaluation framework for assessing how effectively a method can generate 'better' theories about group expression, backed by human validation. We publicly release the raw corpora and evaluation scripts for Splits! to help researchers assess how methods infer--and potentially misrepresent--group differences in expression. We make Splits! and our evaluation module available at https://github.com/eyloncaplan/splits.

Problem

• Evaluating demographic social inference in LLMs
• Developing theories for group expression differences
• Assessing bias in Large Language Models

Innovation

• Introduces Group Theorization task for demographic analysis
• Uses Reddit posts split by topics and demographics
• Provides evaluation framework with human validation