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

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.

💡 Research Summary

The paper tackles the persistent problem that large language models (LLMs) often produce culturally misaligned outputs because their pre‑training data are heavily skewed toward high‑resource, Western‑centric sources and because they lack a structured representation of human values. Existing mitigation strategies—cultural prompting, few‑shot exemplars, ValuesRAG retrieval, and multi‑agent debate—either rely on loosely grounded cultural signals, treat values as independent features, or sacrifice interpretability when aggregating multiple agents.

To overcome these limitations, the authors introduce OG‑MAR (Ontology‑Guided Multi‑Agent Reasoning), a four‑stage framework that integrates (i) value summarization, (ii) ontology construction, (iii) context retrieval, and (iv) multi‑persona simulation with a final judgment step.

Stage 1 – Value Summarization. The World Values Survey (WVS) is used as a rich, globally representative source of individual value data. Raw responses are parsed into two parts—demographic attributes and value‑related answers—and then fed to a Summarization Agent (G_sum). For each of 76 pre‑defined taxonomy classes (derived from 12 top‑level domains), G_sum produces a concise, class‑specific synopsis s_i,j for respondent i. The collection of these synopses forms a structured value profile V_i, which serves as a compact, interpretable representation of an individual’s cultural stance.

Stage 2 – Ontology Construction. Domain experts design a set of Competency Questions (CQs) that probe relationships between pairs of parent classes in the taxonomy. An LLM, constrained to use only the predefined classes and relation phrases, answers each CQ, yielding natural‑language triples t = (c_a, p_a,b, c_b). To embed cultural diversity, the LLM is conditioned on value profiles sampled from 120 respondents (20 per region across six continents). Human curators then validate, edit, and prune the generated triples for cultural plausibility, consistency, and redundancy. The final ontology comprises 76 classes and 150 object‑property triples, forming a semantic network that explicitly encodes cross‑category value dependencies.

Stage 3 – Query Analysis & Context Retrieval. Given a user query q and target demographic d_q, the system first runs a Topic‑Selection Agent (G_topic) that encodes q and selects the top‑k relevant domains and the top‑p fine‑grained categories (F_q) based on embedding similarity. Ontology triples whose end‑points lie in F_q are scored by the maximum similarity of their nodes to q, and the top‑M triples are retrieved as the ontology context O_q. Simultaneously, a dense‑embedding retrieval over the demographic database returns the K most similar respondents R_q, together with their value profiles V_q. This dual retrieval yields a culturally and demographically grounded context for the next stage.

Stage 4 – Multi‑Value Persona Agent Simulation & Judgment. For each retrieved individual i ∈ R_q, a Value‑Persona Agent (G_persona) is instantiated with a conditioning string z_i that concatenates O_q, the individual’s filtered value summaries V_i,q (restricted to classes appearing in O_q), and the demographic description d_i. Each agent generates (ˆy_i, ρ_i): an answer and a natural‑language reasoning trace that respects the supplied ontology constraints. All agent outputs are collected into a set A. A Judgment Agent (G_judge) then performs three operations: (a) checks that each candidate answer does not violate any retrieved ontology triple, (b) weights answers by demographic proximity to d_q, and (c) aggregates the weighted answers using a meta‑adjudication mechanism (e.g., learned voting or confidence‑based blending). The final output is a culturally aligned answer together with a consolidated reasoning trace.

Experimental Evaluation. The authors evaluate OG‑MAR on six regional benchmark suites derived from major social‑survey datasets (covering Asia, Europe, Africa, North America, Latin America, and India). Four LLM backbones are tested: GPT‑3.5‑Turbo, LLaMA‑2‑13B, Claude‑2, and Gemini‑1.5. Baselines include ValuesRAG, cultural prompting, and the Debate‑only multi‑agent framework. Across all backbones and regions, OG‑MAR yields an average 7.2 % increase in cultural‑alignment scores (measured against ground‑truth WVS distributions) and a 5.4 % boost in robustness to prompt perturbations and injected noise. Human evaluators rate the transparency of OG‑MAR’s reasoning traces at 0.84 on a 0–1 scale, substantially higher than the baselines (≈0.62). Ablation studies confirm that both the ontology component and the demographic‑grounded persona agents contribute significantly to performance gains.

Contributions & Significance.

1. Introduces a structured cultural knowledge base that fuses empirical value distributions with an expert‑curated ontology, enabling consistent reasoning about value interdependencies.
2. Proposes a demographic similarity retrieval mechanism that personalizes the multi‑agent simulation, moving beyond generic cultural prompts.
3. Develops a judgment module that enforces ontology consistency while aggregating diverse agent opinions, thereby improving both alignment and interpretability.
4. Demonstrates that the framework is model‑agnostic and scalable across heterogeneous LLMs and cultural contexts.

Limitations & Future Work. Ontology construction still relies on expert‑written CQs and manual validation, which may limit rapid adaptation to new domains or emerging cultural concepts. The CQ design process is labor‑intensive and could benefit from semi‑automated generation techniques. Moreover, the retrieval and multi‑agent inference steps incur non‑trivial computational overhead; optimizing these components for real‑time deployment is an open challenge. Future research directions include (a) automating CQ generation via meta‑learning, (b) expanding the ontology to capture dynamic cultural shifts, and (c) integrating low‑resource language models to reduce inference cost while preserving alignment.

In summary, OG‑MAR presents a novel, ontology‑driven multi‑agent architecture that substantially improves the cultural alignment, robustness, and explainability of LLM outputs, marking a meaningful step toward responsible, globally inclusive AI systems.