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

TL;DR

The paper addresses cultural misalignment in LLMs arising from skewed pretraining data and ungrounded value signals. It introduces OG-MAR, an ontology-guided multi-agent framework that builds a fixed WVS-based cultural taxonomy, derives CQ-driven ontology relations, retrieves ontology-consistent context and demographically similar profiles, and uses multi-value persona agents with a constrained judgment agent to produce ontology-consistent predictions. Across six regional benchmarks and four backbones, OG-MAR improves cultural alignment and robustness while yielding interpretable reasoning traces grounded in explicit value relations. The approach offers a scalable path toward culturally aligned LLMs with transparent, evidence-based reasoning that respects demographic context.

Abstract

Large Language Models (LLMs) increasingly support culturally sensitive decision making, yet often exhibit misalignment due to skewed pretraining data and the absence of structured value representations. Existing methods can steer outputs, but often lack demographic grounding and treat values as independent, unstructured signals, reducing consistency and interpretability. We propose OG-MAR, an Ontology-Guided Multi-Agent Reasoning framework. OG-MAR summarizes respondent-specific values from the World Values Survey (WVS) and constructs a global cultural ontology by eliciting relations over a fixed taxonomy via competency questions. At inference time, it retrieves ontology-consistent relations and demographically similar profiles to instantiate multiple value-persona agents, whose outputs are synthesized by a judgment agent that enforces ontology consistency and demographic proximity. Experiments on regional social-survey benchmarks across four LLM backbones show that OG-MAR improves cultural alignment and robustness over competitive baselines, while producing more transparent reasoning traces.

Figures (18)

• Figure 1: Overall architecture of the OG-MAR framework. The pipeline illustrates the overall architecture of OG-MAR. It begins with Data Preprocessing & Ontology Construction (left). During inference, for a given query and target demographics, it performs Ontology & Demographic Retrieval (center) to gather relevant context. This context is used to instantiate multiple Persona Agents (top right) whose outputs are synthesized by a Judgment Agent (bottom right) to produce the final, culturally aligned prediction.
• Figure 2: Visualization of the final ontology structure. The ontology comprises 76 classes and 150 pairs of object properties, forming a comprehensive semantic network.
• Figure 3: Performance comparison of four models across K $\in$ {1, 3, 5, 10} on average. Red vertical dashed lines indicate the best K and gray horizontal lines show the overall mean accuracy.
• Figure 4: Performance comparison between OG-MAR and the Value Inference Variant. Accuracy over four models on six regional datasets. Dashed lines show per-method average accuracy; red boxes report the average gap (Avg $\Delta$ = OG-MAR - Variant).
• Figure 5: Average human evaluation scores (5-point Likert scale) across three tasks: Persona Fidelity (Consistency, Grounding), Judgment Logic (Synthesis, Context), and Retrieval Validity (Relevance). Scores are averaged over nine expert raters.