Research map / Large language models / Reasoning and chain of thought

Reasoning Shapes Alignment: Investigating Cultural Alignment in Large Reasoning Models with Cultural Norms

AAAI 2026 · Large language models › Reasoning and chain of thought · approach: Search and self-correction in reasoning · 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 CNCA, which mines cultural norms from limited World Values Survey data using reasoning models and uses them either in-context or internalized via norm-enhanced CoT self-distillation with SFT/DPO to improve cultural alignment.

Problem

Large reasoning models need to reflect diverse cultural values, but cultural norms are not readily available and must be derived from survey data; it is unclear how to mine and use them to improve cultural alignment.

Approach

Three norm-mining methods (topic-only, topic plus top-1 questionnaire answers, topic plus ranked answers) produce low- and high-level cultural norms that are either inserted into the prompt or used to generate norm-conditioned CoT self-distillation data for SFT and DPO, so the model internalizes norms without explicit input.

Key results

• CNCA-TQ (TA) with norms on R1-Qwen-14B raises the alignment score from 65.55 to 68.24 (+2.69).
• Internalized CNCA-DPO on R1-Qwen-14B reaches 66.90 vs 65.55 vanilla and 66.20 for CoT-SFT on WVS.
• On out-of-distribution CDEval, CNCA-DPO achieves the best average (60.52) vs vanilla 58.90 and CoT-SFT 60.38.

Surprising findings

• Ranked-answer norm mining (TQ-RA) is least effective and can introduce noise, while top-1 answers work best.
• Vanilla R1-Qwen-14B underperforms its non-reasoning counterpart Qwen2.5-14B-Instruct without norms, though it benefits more once norms are added.

Limitations

• Evaluation relies on a single survey-based alignment metric averaged over 18 countries and relatively small gains (often under 1-3 points).
• Norm mining and alignment are tested only on distilled R1 models and two non-reasoning models, limiting generality.
• Uses majority WVS answers as ground truth, which may not capture within-country diversity.

Experimental setup

Datasets: World Values Survey CDEval

Benchmarks: World Values Survey CDEval

Metrics: cultural alignment score (normalized distance to majority WV average score over 18 countries p-value from t-test

Baselines: Vanilla In-Context Learning SFT CoT-SFT Meta-Llama-3.1-8B-Instruct Qwen2.5-14B-Instruct

Builds on: DeepSeek-R1 Constitutional AI SELF-ALIGN Direct Preference Optimization CoT prompting

Closest papers by meaning

• CulturePark: Boosting Cross-cultural Understanding in Large Language Models (NeurIPS 2024 · similarity 0.86)
• Toward Culturally Aligned LLMs through Ontology-Guided Multi-Agent Reasoning (ICML 2026 · similarity 0.86)
• Training-Free Cultural Alignment of Large Language Models via Persona Disagreement (NeurIPS 2026 · similarity 0.86)
• Global PIQA: Evaluating Commonsense Reasoning Across 100+ Languages and Cultures (NeurIPS 2026 · similarity 0.86)
• Controllable Safety Alignment: Inference-Time Adaptation to Diverse Safety Requirements (ICLR 2025 · similarity 0.85)
• ProgressGym: Alignment with a Millennium of Moral Progress (NeurIPS 2024 · similarity 0.85)