Toward Culturally Aligned LLMs through Ontology-Guided Multi-Agent Reasoning

2025-12-28 · Source: cs.AI updates on arXiv.org · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Data Science & Analytics · Depth: Expert, extended

Summary

The Ontology-Guided Multi-Agent Reasoning (OG-MAR) framework addresses cultural misalignment in Large Language Models (LLMs) caused by biased pretraining data and unstructured value representations. OG-MAR constructs a global cultural ontology with 76 classes and 150 object properties by summarizing World Values Survey (WVS) data and eliciting relations via competency questions. During inference, it retrieves ontology-consistent relations and demographically similar profiles to instantiate multiple value-persona agents. A judgment agent then synthesizes these outputs, enforcing consistency and demographic proximity. Evaluated across six regional social-survey benchmarks (EVS, GSS, CGSS, ISD, LAPOP, Afrobarometer) and four LLM backbones (GPT-4o-mini, Gemini 2.5, Qwen 2.5, EXAONE 3.5), OG-MAR achieved average accuracies up to 0.6317, outperforming competitive baselines and yielding more transparent reasoning traces.

Key takeaway

For AI Scientists developing LLMs for culturally sensitive applications, traditional bias mitigation methods are insufficient. You should integrate structured cultural ontologies and multi-agent reasoning, like OG-MAR, to ensure robust and demographically aligned outputs. This approach, which leverages empirically grounded value data and explicit value relationships, significantly improves interpretability and performance across diverse regions. Consider adopting ontology-guided frameworks to move beyond implicit cultural assumptions and enhance transparency in your models.

Key insights

Ontology-guided multi-agent reasoning with demographic grounding improves LLM cultural alignment and interpretability.

Principles

• Cultural values are structurally interdependent.
• Demographic grounding enhances LLM alignment.
• Ontology-guided reasoning boosts interpretability.

Method

Summarize WVS values, build a CQ-guided cultural ontology, retrieve ontology triples and similar profiles, simulate multi-value persona agents, and use a constrained judgment agent for final adjudication.

In practice

• Utilize World Values Survey for value data.
• Employ competency questions for ontology building.
• Instantiate K=5 persona agents for diverse views.

Topics

• Large Language Models
• Cultural Alignment
• Multi-Agent Systems
• Ontology Engineering
• World Values Survey
• Bias Mitigation

Best for: AI Scientist, Research Scientist, AI Ethicist

Related on AIssential

• Korean Culture into LLM Alignment: Toward Cultural Coherence — LLM cultural alignment requires constructive, locale-specific definitions of coherence, moving beyond mere output suppression.
• Enhancing Value Alignment of LLMs with Multi-agent system and Combinatorial Fusion — Multi-agent fusion with combinatorial analysis enhances LLM value alignment by leveraging diverse moral perspectives.
• AlignCultura: Towards Culturally Aligned Large Language Models? — Cultural alignment in LLMs requires joint HHH optimization grounded in diverse cultural taxonomies to avoid biased outputs.
• CLASH: Evaluating Language Models on Judging High-Stakes Dilemmas from Multiple Perspectives — LLMs struggle with nuanced human value judgments, including ambivalence, discomfort, and dynamic value shifts in high-stakes dilemmas.
• AlignCultura: Towards Culturally Aligned Large Language Models? — Cultural alignment in LLMs requires systematic evaluation against UNESCO principles to prevent biased outputs.
• Korean Culture into LLM Alignment: Toward Cultural Coherence — LLM cultural alignment requires defining culturally coherent responses, not just suppressing harmful ones.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by cs.AI updates on arXiv.org.