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

AI summary

Overview

Research area: Natural Language Processing, specifically cultural alignment of large language models, combining ontology engineering, retrieval-augmented generation, and multi-agent reasoning.

Technical level: Advanced. The paper assumes familiarity with LLM prompting paradigms, retrieval-augmented generation, multi-agent deliberation, knowledge ontologies, and survey-based cultural evaluation metrics.

Scope in one sentence: The paper introduces OG-MAR, a framework that builds a cultural ontology from World Values Survey data and uses it to ground a multi-persona agent pipeline, then evaluates cultural alignment on six regional social-survey benchmarks across four LLM backbones.

What This Paper Is About

Large language models are trained on web data that overrepresents high-resource, Western-centric viewpoints, so they tend to fall back on culture-default assumptions when answering questions about social norms and values. Existing fixes, such as role prompting, curated demonstrations, retrieval of survey snippets, or multi-agent debate, either lack demographic grounding or treat cultural values as isolated, unstructured signals, which limits consistency and interpretability. The authors propose OG-MAR to inject structured cultural relations and demographically similar respondent profiles into a multi-agent reasoning pipeline, and test whether that yields better alignment across regions.

Key Contributions

1. A CQ-guided cultural ontology over a fixed taxonomy. Domain experts write Competency Questions (CQs) that probe relations between sub-classes of two given top-classes. An LLM then describes sub-class-level relations under cultural conditioning, sampled from value profiles of 120 individuals (20 per region) across six world regions, and human experts validate, edit, and remove relations. The resulting taxonomy has 12 top-classes and 64 sub-classes (|C|=76), connected by 150 object-property relations.

2. A value summarization pipeline. Each WVS respondent record is split into demographic attributes and values-related responses, and a Summarization Agent produces a category-conditioned synopsis s_i(c) for each class under a "no-new-concepts" constraint, yielding a structured value profile V_i.

3. A multi-persona inference pipeline with constrained meta-adjudication. At inference, the framework retrieves ontology triples and demographically similar respondents, instantiates a Value-Persona Agent per retrieved individual, and synthesizes their answers with a Judgment Agent that follows an evidence-first protocol: score grounding and ontology compliance per option, consult vote signals only as a secondary cue, and break remaining ties by persona relevance to the target demographics.

4. Evaluation across six regional benchmarks and four backbones. The framework is tested on EVS, GSS, CGSS, ISD, LAPOP, and Afrobarometer against five baselines, plus ablations on retrieval depth, value inference, and persona simulation, along with a human evaluation by nine domain experts.

Main Findings

• Average accuracy gains across backbones. OG-MAR reaches average accuracies of 0.6308 on Gemini 2.5 Flash Lite, 0.5705 on Qwen 2.5, 0.6317 on EXAONE 3.5, and 0.6007 on GPT-4o-mini. For comparison, the strongest baseline (ValuesRAG) reaches 0.5731 on Gemini 2.5 Flash Lite, 0.5339 on Qwen 2.5, 0.5613 on EXAONE 3.5, and 0.5961 on GPT-4o-mini.

• Largest gains on culturally distant settings. The paper reports particularly large improvements on CGSS (China) and ISD (India), arguing that structured cultural relations and demographically grounded personas help most when the target distribution deviates from dominant pretraining priors. For example, on EXAONE 3.5, OG-MAR reaches 0.7810 on ISD and 0.7022 on LAPOP; on Gemini 2.5 Flash Lite, it reaches 0.7017 on CGSS and 0.7007 on ISD.

• Statistical significance on many cells. Asterisks in Table 1 denote significant improvements (paired t-test with Holm–Bonferroni correction, p < 0.05) over all baselines. On Gemini 2.5 Flash Lite, all six regional results are marked significant.

• Retrieval depth of K = 5 is optimal. Varying the number of retrieved individuals over K ∈ {1, 3, 5, 10}, all four models peak at K = 5, outperforming other sizes by +0.003 to +0.07. Gains from K = 3 to K = 5 range from 0.03 to 0.05 for the models other than Gemini 2.5 Flash Lite, while K = 10 degrades accuracy by 0.02 to 0.07 relative to K = 5.

• Persona simulation contributes, but is not the whole story. Compared with a single-judge variant that skips persona simulation, OG-MAR is higher by +0.002 on GPT-4o-mini, +0.03 on Gemini 2.5 Flash Lite, +0.04 on QWEN 2.5, and +0.07 on EXAONE 3.5. The authors note the single-judge variant remains competitive, suggesting gains also come from the shared ontology-grounded retrieval and value summarization pipeline. The single-judge variant is consistently better on GSS.

• Multi-persona beats single value inference in most cases. Against a Value Inference Variant that replaces K persona agents with one agent that infers a profile, GPT-4o-mini slightly benefits from the variant (about 0.01 on average, with +0.07 on GSS and +0.11 on CGSS). The other three models gain with OG-MAR, ranging from +0.03 to +0.10 on average across datasets.

• Human evaluation supports interpretability. Nine domain experts rated reasoning traces on three 5-point Likert tasks: Persona Fidelity, Judgment Logic, and Retrieval Validity. Grounding peaks on CGSS (China) at 4.02, slightly above GSS (U.S.) at 3.97. Synthesis Logic averages 3.72 and retrieval relevance averages 3.64.

• Token cost is highest, by design. OG-MAR incurs the highest token budget among compared methods. The authors frame this as the cost of a structured reasoning framework rather than a token-efficient prompting replacement, and point to its higher accuracy and lower average MAE as evidence the extra computation is productive.

Methodology in Plain English

The researchers start from the World Values Survey, a large cross-national survey with demographic attributes such as country, age, gender, and education. They fix a taxonomy of 12 top-level value topics and 64 sub-topics, then have a summarization model turn each respondent's raw answers into short, category-specific summaries, keeping only information relevant to each class and introducing no new concepts.

To capture how value categories relate to each other, domain experts write Competency Questions that ask about meaningful interactions between sub-classes of two given top-classes. An LLM answers each question while restricted to the predefined classes, producing ordered triples in the form (sub-class, relation phrase, sub-class). To reflect diverse perspectives, the model is conditioned on value profiles drawn from 120 individuals (20 per region) across six world regions. Human experts then review the candidate relations for cultural plausibility, edit them for clarity, and delete spurious ones, leaving 150 relations.

At inference, a query goes through three retrieval steps. A fine-tuned text encoder picks the top-k relevant top-classes; dense retrieval over sub-class embeddings scores ontology triples by the relevance of their endpoint nodes, restricted to triples whose endpoints fall under the selected top-classes; and a separate dense retrieval ranks survey respondents by demographic similarity to the target profile. The implementation retains the top-3 fine-grained categories per query, up to the top-3 triples per category, and K = 5 demographically similar individuals by default.

Each retrieved individual becomes a Value-Persona Agent conditioned on the retrieved ontology triples (typically 3–9), that person's value summaries restricted to the referenced sub-classes, and their demographic attributes. Each agent returns an answer plus a natural-language reasoning trace. A Final Judgment Agent then reviews the collected answers and traces in a single LLM call. It does not see the ontology or value profiles directly; grounding flows through the persona outputs. Its protocol scores evidence grounding and ontology compliance per answer option first, treats vote counts only as a secondary signal when leading options are close, and breaks remaining ties using persona relevance to the target demographics.

Models used are GPT-4o-mini and Gemini 2.5 via APIs, with Qwen 2.5 and EXAONE 3.5 as open-source backbones, all at temperature 0; GPT-o4-mini handles object-property construction and value profile generation. Retrieval uses E5-base embeddings, and topic classification fine-tunes DeBERTa-v2-xxlarge on WVS data for 3 epochs with batch size 4 and learning rate 5×10⁻⁶.

Evaluation uses binary accuracy after response-type-specific rules (two-choice binary, derived binary, ordinal binarized at the scale midpoint, and nominal exact match), plus MAE on ordinal items. Distribution-level metrics are EMD, Proportion Correlation, and TVD. Test data comes from six regional surveys (EVS, GSS, CGSS, ISD, LAPOP, Afrobarometer), sampled via clustering to 2,000 representative instances.

Why This Matters

Impact on research. The paper argues that most prior cultural-bias mitigation treats values as unstructured snippets, and that ontology engineering remains underused in this space. It offers a concrete pipeline for combining structured value relations with demographic grounding and multi-agent simulation, and it reports per-region breakdowns and ablations that let other researchers see when structure helps (CGSS, ISD) and when it does not (GSS, where simpler variants sometimes win).

Real-world applications:

• Regional AI assistants and chat systems deployed in markets where dominant pretraining priors mismatch local norms, such as China, India, or Latin America.
• Content moderation and policy review workflows that need to surface value considerations behind a decision rather than emit a single opaque label.
• Social-science and survey research tooling that simulates responses from demographically specified profiles as a pre-testing aid.
• Auditing pipelines for evaluating whether a model's outputs track a target population's reported values, using the reasoning traces as inspectable evidence.

Industry relevance. Companies building localized or region-specific LLM products face the alignment problem the paper targets. OG-MAR's emphasis on transparent reasoning traces and explicit evidence-first adjudication addresses a practical need for justification in regulated or high-stakes contexts, though the paper acknowledges the highest token cost of the compared methods, which matters for production economics.

Ethical caveats the authors raise. The impact statement warns that demographic retrieval and survey-grounded signals create risks of profiling, stereotype reinforcement, survey-artifact encoding, and bias amplification. The authors state OG-MAR is meant to surface value considerations for scrutiny and auditing, not to justify demographic generalizations, and recommend evaluating across subpopulations and applying strong privacy and data-governance safeguards, including protection against re-identification.

Future Directions

• Stronger verification of ontology triples. The authors plan more robust verification mechanisms for triple accuracy and consistency, since imperfect triples can propagate structured but inaccurate constraints into reasoning.
• Better cross-domain topic classification and uncertainty-aware retrieval. Regional surveys are not always aligned with the WVS-derived taxonomy, and topic selection can fail on domain-specific phrasing or culturally local constructs; demographic retrieval may also miss minority or underrepresented profiles.
• Cheaper multi-agent simulation. The paper proposes efficient routing or pruning techniques inside the simulation to cut computational cost without sacrificing alignment quality.
• Uncertainty communication, debiasing, and safe deployment protocols. The impact statement calls for future work on these, alongside studying performance across subpopulations.

The paper also notes an unaddressed question: what parts of the gains come from ontology-grounded retrieval and value summarization versus persona simulation, since the single-judge variant remains competitive in several settings.

Target Audience

Researchers and practitioners working on cultural alignment, cross-cultural NLP, and evaluation of LLM value judgments; applied scientists building retrieval-augmented or multi-agent LLM systems; and ontology engineers interested in LLM-assisted construction with human review. Social scientists who use cross-national surveys such as WVS, EVS, GSS, CGSS, ISD, LAPOP, or Afrobarometer will also find the demographic-grounded retrieval design relevant. Readers need comfort with LLM prompting, retrieval metrics, and survey-based evaluation to follow the quantitative sections in full.

Authors’ 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.