Research map / Multi-agent systems and games / LLM agents

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

ICML 2026 · Multi-agent systems and games › LLM agents · approach: LLM multi-agent decision making · 0 citations (OpenAlex)

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Proposes OG-MAR, which builds a CQ-derived cultural ontology over a fixed WVS-based taxonomy and uses ontology-retrieved triples, demographically similar WVS personas, and a judgment agent to improve cultural alignment on six regional survey benchmarks.

Problem

LLMs show culturally skewed value judgments, and existing steering methods (role prompts, RAG, debate) lack demographic grounding and treat cultural values as unstructured independent signals, reducing consistency and interpretability.

Approach

Combine a fixed 12-top/64-sub-class cultural ontology (CQ-guided relations, human-reviewed) with demographic retrieval of similar WVS respondents to instantiate value-persona agents, then synthesize their answers with an evidence-first judgment agent that checks ontology consistency.

Key results

• OG-MAR averages 0.6007 on GPT-4o-mini (best among methods, vs 0.5961 for ValuesRAG), 0.6308 on Gemini 2.5 Flash Lite, 0.5705 on Qwen 2.5, and 0.6317 on EXAONE 3.5.
• Single-judge ablation drops average accuracy to 0.6022 on Gemini 2.5 and 0.5311 on Qwen 2.5, indicating persona simulation contributes.
• On the ISSP 2023 post-cutoff control, OG-MAR reaches 0.6879 on GPT-4o-mini, the highest across all baselines.

Surprising findings

• On GSS, the single-judge variant outperforms full OG-MAR across models, and the Value Inference variant beats OG-MAR on GSS.
• GPT-4o-mini gains only about 0.002 from multi-persona simulation over the single-judge variant, so its gains there are mostly from the shared retrieval pipeline.

Limitations

• Multi-agent simulation increases inference cost and latency relative to single-pass baselines.
• Imperfect or sparse ontology triples can propagate inaccurate constraints; demographic retrieval may miss minority profiles.
• Topic classifier Top-1 accuracy on regional datasets is low (0.34 to 0.54), limiting topic selection.

Experimental setup

Datasets: World Values Survey EVS GSS CGSS ISD LAPOP Afrobarometer ISSP 2023

Benchmarks: EVS GSS CGSS ISD LAPOP Afrobarometer ISSP 2023 (post-cutoff control)

Metrics: accuracy MAE EMD proportion correlation TVD human Likert ratings

Baselines: Zero-shot Role Assignment Self-consistency Debate ValuesRAG

Builds on: ValuesRAG self-consistency multi-agent debate Competency Questions ontology engineering METHONTOLOGY

Closest papers by meaning

• Reasoning Shapes Alignment: Investigating Cultural Alignment in Large Reasoning Models with Cultural Norms (AAAI 2026 · similarity 0.86)
• Training-Free Cultural Alignment of Large Language Models via Persona Disagreement (NeurIPS 2026 · similarity 0.86)
• ALI-Agent: Assessing LLMs' Alignment with Human Values via Agent-based Evaluation (NeurIPS 2024 · similarity 0.85)
• CulturePark: Boosting Cross-cultural Understanding in Large Language Models (NeurIPS 2024 · similarity 0.84)
• Language Model Alignment in Multilingual Trolley Problems (ICLR 2025 · similarity 0.84)
• Global PIQA: Evaluating Commonsense Reasoning Across 100+ Languages and Cultures (NeurIPS 2026 · similarity 0.84)