OG-MAR uses multi-agent reasoning to align LLMs with cultural values

A team of researchers (Wonduk Seo et al.) has introduced OG-MAR, an Ontology-Guided Multi-Agent Reasoning framework designed to make large language models more culturally aligned. The framework tackles the common issue of LLMs lacking structured value representations by leveraging the World Values Survey (WVS), which captures respondent-specific values across cultures. OG-MAR first summarizes these values and constructs a global ontology by eliciting relations over a fixed taxonomy using competency questions. At inference time, the system retrieves ontology-consistent relations and demographically similar profiles to instantiate multiple value-persona agents. These agents then engage in reasoning, and their outputs are synthesized by a judgment agent that enforces both ontology consistency and demographic proximity.

Experimental results on regional social-survey benchmarks across four LLM backbones (including GPT and open-source models) demonstrate that OG-MAR significantly improves cultural alignment and robustness compared to existing methods. The framework also produces more transparent reasoning traces, allowing users to understand how cultural values influenced the output. The paper has been accepted at ICML 2026 Regular Track, marking a step toward more culturally aware AI systems that can handle sensitive decision-making without bias.