The Curious Case of Curiosity across Human Cultures and LLMs

TL;DR

This work introduces CUEST, a framework for evaluating culture-aware curiosity in humans and LLMs by jointly analyzing linguistic style, topic preferences, and grounding in social science constructs. Using Yahoo! Answers spanning 18 countries and 16 topics, plus LLM-generated questions with country personas, the authors reveal that LLMs tend to flatten cross-cultural diversity and align more with Western patterns. They demonstrate that adapter-based fine-tuning can narrow human–model alignment gaps by inducing curiosity, with downstream improvements on cross-cultural benchmarks NormAD, CulturalBench, and Cultural Commonsense. The results highlight practical pathways to build culturally aware, curiosity-driven NLP systems and suggest future work on expanding language coverage, refining social-science grounding, and scaling adapters for broader cultural breadth.

Abstract

Recent advances in Large Language Models (LLMs) have expanded their role in human interaction, yet curiosity -- a central driver of inquiry -- remains underexplored in these systems, particularly across cultural contexts. In this work, we investigate cultural variation in curiosity using Yahoo! Answers, a real-world multi-country dataset spanning diverse topics. We introduce CUEST (CUriosity Evaluation across SocieTies), an evaluation framework that measures human-model alignment in curiosity through linguistic (style), topic preference (content) analysis and grounding insights in social science constructs. Across open- and closed-source models, we find that LLMs flatten cross-cultural diversity, aligning more closely with how curiosity is expressed in Western countries. We then explore fine-tuning strategies to induce curiosity in LLMs, narrowing the human-model alignment gap by up to 50%. Finally, we demonstrate the practical value of curiosity for LLM adaptability across cultures, showing its importance for future NLP research.

Figures (16)

• Figure 1: Curiosity-driven question differences across countries. For a given topic, curiosity varies in style and content across countries. Here, we show examples from Argentina, India, Italy, and the Philippines. We extend this analysis to 18 countries across 16 topics.
• Figure 2: CUEST Overview: (1) Linguistic Alignment, (2) Topic Preference Alignment, and (3) Grounding to Social Science Constructs; to compare human- and LLM- questions.
• Figure 3: Linguistic Alignment. Humans and LLMs diverge: humans high in ambiguity and cohesion score, LLMs high in open-ended questions.
• Figure 4: Country-wise Linguistic Alignment: (low mean abs diff: high alignment) Western countries show the highest human–model alignment, followed by Latin American and Eastern countries.
• Figure 5: Country-wise Topic-Preference Alignment: LLaMA-3-8b correlation with humans. Although we find notable country-level differences, no region-level patterns emerge.