Technical Summary: MET: Theory-Grounded and Culture-Aware Multilingual Moral Reasoning

1. Problem Statement

The paper identifies three critical gaps in the current state of multilingual moral reasoning for Large Language Models (LLMs):

1. Evaluation Limitations: Existing multilingual moral benchmarks rely on direct translation of English scenarios. This approach fails to adapt culture-specific items (e.g., named entities, institutions, currencies, customs), thereby importing inappropriate Western cultural contexts into non-English settings (e.g., translating "Thanksgiving" or "Wendy's" directly into Korean or Hindi).
2. Inference-Time Methodology: Current reasoning scaffolds are often static, English-centric, and lack grounding in moral theory. They fail to adapt to specific situations or cultural contexts, and existing theory-grounded work is limited to narrow sets of theories without multilingual support. Furthermore, multilingual reasoning techniques often assume a language-invariant reasoning structure, which does not hold when moral frameworks vary by culture.
3. Training Challenges: Training models for moral decision-making typically requires expensive supervision from stronger models or human annotators. Moral decision-making is inherently subjective and lacks verifiable ground truth, making standard rejection sampling difficult.

2. Methodology

The authors propose a comprehensive framework addressing these gaps through a new benchmark, a two-step prompting method, and a self-supervised training strategy.

2.1 MCLASH: A Culturally Adapted Benchmark

To address the evaluation gap, the authors introduce MCLASH (Multilingual Cultural Adaptation of Situations for Moral Reasoning).

• Construction: Unlike previous works that use direct translation, MCLASH is constructed by culturally adapting the existing CLASH corpus (345 high-stakes situations) into five additional languages: Chinese, Hindi, Korean, Malay, and Spanish.
• Process: The adaptation utilizes Qwen3-32B to rewrite scenarios, replacing Western entities with culturally relevant equivalents (e.g., adapting "Amazon" to "Naver" or "Kakao" for Korea). This is followed by rigorous human inspection by native speakers using a detailed checklist to ensure cultural relevance, naturalness, and factual accuracy.
• Scale: The final dataset contains 1,852 situations and 9,260 character descriptions across six languages.

2.2 MET: Theory-Grounded Prompting

MET (Multilingual Ethics with Theory-grounded reasoning) is a two-step prompting method designed to align reasoning with specific cultural and situational contexts.

• Theoretical Grounds: The authors curated a global set of 37 theoretical grounds derived from psychology and philosophy literature, organized into six dimensions:
  1. Normative Authority
  2. Value Systems
  3. Ethical Theories
  4. Cognitive Reasoning Strategies
  5. Conflict Handling Strategies
  6. Moral Uncertainty
• Two-Step Process:
  1. Ground Selection: Given a situation, the model selects the most relevant grounds from the curated set to form an instance-specific framework.
  2. Theory-Guided Reasoning: The model generates a reasoning chain and a decision based on the selected grounds, outputting in the user's native language.

2.3 MET-D: Self-Distillation Training

To address the issue of shallow reasoning (where models select grounds but fail to engage with them substantively) and the lack of verifiable ground truth, the authors introduce MET-D (MET-Distillation).

• Synthetic Dataset Construction: Leveraging the structure of CLASH, where character descriptions define value priorities that determine the "correct" decision, the authors generate a synthetic training dataset. They extract cultural values from the World Values Survey (WVS), generate scenarios, and pair them with character descriptions that yield verifiable ground-truth labels (Yes/No/Ambiguous).
• Rejection Sampling Self-Distillation: The model is prompted to select grounds and generate reasoning chains. Only instances where the model's final decision matches the ground truth are retained. These filtered chains are used to train the model via supervised fine-tuning.
• Inference Strategy: During inference, the base model performs ground selection, while the trained model performs the theory-guided reasoning, ensuring the reasoning chain is in the native language.

3. Key Contributions

1. MCLASH Benchmark: A multilingual moral decision-making benchmark of 1,852 long-form scenarios constructed via cultural adaptation rather than direct translation, covering six languages.
2. Expert-Curated Grounds: A set of 37 theoretical grounds organized into six high-level dimensions, translated into all six target languages, providing a structured framework for moral reasoning.
3. MET and MET-D: A two-step prompting method (MET) and a self-distillation training strategy (MET-D) that enables culture-aligned, theory-grounded reasoning without requiring external human supervision or stronger teacher models.

4. Results

The method was evaluated on three models of varying sizes and families: Qwen3-4B, Qwen3-8B, and Gemma3-4B.

• Performance Gains:
  • On MCLASH, MET-D improved the macro-F1 score over the base model by an average of 3.71 points across all three models.
  • On MMoralExceptQA, MET-D achieved an average improvement of 4.23 points.
  • The largest single gain was 12.94 points for Malay on Qwen3-8B.
  • Improvements were consistent across high-resource (English, Spanish, Chinese) and low-resource (Korean, Hindi, Malay) languages, with low-resource languages often seeing larger gains.
• Native-Language Reasoning: MET-D significantly increased the proportion of reasoning chains generated in the native language by 62.13 percentage points on average (e.g., from <1% to ~90% in some languages), enhancing legibility for non-English speakers without sacrificing performance.
• Cultural Alignment: Analysis revealed that the most beneficial grounds vary systematically across cultures. For instance, English reasoning benefited most from "Probabilistic Moral Uncertainty" and "First-principles reasoning" (reflecting Anglo-American rationalism), while Korean and Chinese benefited from "Contractarianism" (reflecting Confucian relational obligations).
• Cross-Lingual Transfer: Transfer performance was found to follow linguistic typology (e.g., SOV languages like Korean and Hindi transferring well to each other) rather than cultural similarity.

5. Significance and Claims

The paper claims to open a path for culture-aligned, theory-grounded multilingual moral reasoning. By moving beyond direct translation and static English-centric scaffolds, the proposed method demonstrates that:

• Moral reasoning frameworks are not language-invariant; they must be adapted to cultural contexts.
• Self-supervised training using character-paired synthetic data can effectively teach models to substantively engage with moral theories without expensive human annotation.
• Enhancing the legibility of reasoning (via native-language chains) is achievable alongside performance improvements, addressing a key barrier to deploying moral AI in diverse linguistic settings.

The authors acknowledge limitations, noting that their training targets theory-guided reasoning rather than ground selection (which still relies on the base model) and that the reconciliation of multiple conflicting grounds remains an open challenge for future work.