Cultural Palette: Pluralising Culture Alignment via Multi-agent Palette

TL;DR

The paper tackles monocultural biases in LLM cultural alignment by introducing Cultural Palette, a multi-agent framework that treats five continental cultures as primary colors and blends them via Cultural MoErges to produce country-specific outputs. It introduces the Pentachromatic Cultural Palette Dataset and a three-stage multi-agent workflow (Cultural Draft, Self-regulated Aggregation, Final Decision) guided by continent-level priors and dynamic parameter merging. Empirical results on 18 countries show consistent improvements in both semantic and real-world cultural alignment over baselines, with ablation studies underscoring the importance of continent-level priors and cross-cultural merging. The approach offers scalable, flexible cultural alignment without extensive country-specific data, with broader implications for equitable, context-aware AI systems across diverse populations.

Abstract

Large language models (LLMs) face challenges in aligning with diverse cultural values despite their remarkable performance in generation, which stems from inherent monocultural biases and difficulties in capturing nuanced cultural semantics. Existing methods struggle to adapt to unknown culture after fine-tuning. Inspired by cultural geography across five continents, we propose Cultural Palette, a multi-agent framework that redefines cultural alignment as an adaptive "color-blending" process for country-specific adaptation. Our approach harnesses cultural geography across five continents through three key steps: First, we synthesize the Pentachromatic Cultural Palette Dataset using GPT-4o, refining continental-level dialogues with Hofstede's cultural dimensions to establish foundational cultural representations. Second, five continent-level alignment agents form specialized cultural communities that generate region-specific draft responses. Third, a Meta Agent employs Cultural MoErges to dynamically blend these cultural "colors" through attention-gated parameter merging, akin to mixing pigments on a palette, resolving conflicts while preserving cultural nuances to produce the final culturally-aligned response. Extensive experiments across various countries demonstrate that \textit{Cultural Palette} surpasses existing baselines in cultural alignment.

Figures (10)

• Figure 1: Comparison between (A) country-specific alignment li2024culturellmli2024cultureparkfeng2024modular, (B) country-joint alignment li2024selfli2024culturellm and (C) Cultural Palette.
• Figure 2: Architecture of Cultural Palette consisting of continent-level alignment on Pentachromatic Cultural Palette Dataset (Sec. \ref{['sec:dataset']}), Cultural MoErges for Meta Agent (Sec. \ref{['subsec:cultural-orpo']}) and Multi-agent Cultural Palette (Sec. \ref{['subsec:multi-agent']}). Details of our ORPO alignment loss formulation are deferred to Appendix B for brevity.
• Figure 3: Comparative analysis of semantic representations on the PRISM and Pentachromatic Cultural Palette datasets using Multilingual embeddings (microsoft/Multilingual-MiniLM-L12-H384) with PCA-based dimensionality reduction. Data points are color-coded by continent, based on the country associated with each QA pair.
• Figure 4: Comparison of semantic-consistency alignment score between our Cultural Palette and other methods (Prompting, Joint) and merging strategies (Tie and Model Stock) on Llama3.1-8B-Instruct and Qwen2.5-7b-Instruct.
• Figure 5: Prompt for continent-aware responses