Research map / Large language models / AI and society

CulturePark: Boosting Cross-cultural Understanding in Large Language Models

NeurIPS 2024 · Large language models › AI and society · approach: Critiques of LLM claims and judging · 12 citations (OpenAlex)

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Introduces CulturePark, an LLM multi-agent cross-cultural dialogue framework that generates 41K cultural samples to fine-tune culture-specific GPT-3.5-Turbo models evaluated on content moderation, Hofstede cultural alignment, and situated cultural education.

Problem

Cultural bias in LLMs stems from scarce culture-representative data, and existing cultural datasets rely on costly real-world collection or human annotation, or on semantic augmentation with limited diversity.

Approach

Simulate cross-cultural human communication with GPT-3.5-based agents (an English main contact plus culture delegates) that debate seed survey questions under self-calibration and self-guidance prompts, then refine the dialogues via opinion extraction, verification, and K-means diversification to build culture-specific fine-tuning data.

Key results

• GPT-3.5-based culture-specific models match or outperform GPT-4 on content moderation across 41 datasets (outperforming GPT-4 on 5 of 8 cultures per the text).
• In the situated learning study, participants learning with the CulturePark models had lower average Euclidean distance to Hofstede ground truth (175.75 vs 215.59 for GPT-4) and higher average satisfaction (4.33 vs 3).
• Ablation shows Generate+Verify+Diversify improves Arabic average F1 from 0.370 (GPT-3.5-turbo) to 0.514.

Measured effects reported in the paper

Surprising findings

• Data generated by GPT-3.5 agents via communication yields fine-tuned models that outperform GPT-4 on content moderation despite GPT-4 being the stronger generator.
• Participants reported GPT-4 responses to be vague and neutral even when prompted to role-play a culture.

Limitations

• Only GPT-3.5-Turbo used as the main agent backbone; other LLMs not systematically explored.
• Fine-tuned models mainly cover high-resource cultures due to scarce low-resource benchmarks.
• Relies on language as a proxy for culture boundaries.

Experimental setup

Datasets: World Values Survey Pew Global Attitudes Surveys OSACT5 TRAC2020 GermEval2018 K-MHaS HateBR ToLD-Br DETOXIS SemEval-2020 CValues BIG-Bench Hard

Benchmarks: 41 content moderation datasets across 8 cultures Hofstede VSM 13 VSM 2013 situated learning exam BIG-Bench Hard

Metrics: average F1 Euclidean distance to Hofstede scores user satisfaction score (1-5)

Baselines: GPT-3.5-turbo GPT-4 Gemini-pro SeaLLM TaiwanLLM Synatra-7B-v0.3-dpo EEVE-Korean-10.8B-v1.0 CultureLLM CultureBank

Builds on: CultureLLM Cognitive Conflict Theory Social Cognitive Theory

Closest papers by meaning

• CultureLLM: Incorporating Cultural Differences into Large Language Models (NeurIPS 2024 · similarity 0.90)
• Culture Affordance Atlas: Reconciling Object Diversity Through Functional Mapping (AAAI 2026 · similarity 0.88)
• Communication Accommodation Between Large Language Models and Users Across Cultures (Student Abstract) (AAAI 2025 · similarity 0.88)
• The PRISM Alignment Dataset: What Participatory, Representative and Individualised Human Feedback Reveals About the Subjective and Multicultural Alignment of Large Language Models (NeurIPS 2024 · similarity 0.87)
• BLEnD: A Benchmark for LLMs on Everyday Knowledge in Diverse Cultures and Languages (NeurIPS 2024 · similarity 0.86)
• Reasoning Shapes Alignment: Investigating Cultural Alignment in Large Reasoning Models with Cultural Norms (AAAI 2026 · similarity 0.86)