PaperlayerView exact revision on arXiv

Source-linked AI summary

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

Ayoung Lee, Ryan Kwon, Yunxiang Zhang, Yuxuan Liu, Peter Railton, Lu Wang

arXiv:2607.11736v1cs.CL

TL;DR

Existing multilingual moral decision-making research leaves cultural adaptation and culturally situated reasoning under-addressed. MET combines a culturally adapted benchmark with theory-grounded prompting and self-distillation; MET-D improves average F1 across MCLASH and MMoralExceptQA and increases native-language usage.

  • Problem

    Existing multilingual moral benchmarks directly translate scenarios without cultural adaptation, while culturally aligned decision-making remains overlooked despite users’ diverse linguistic and cultural contexts.

    Source: §1 · p. 1

  • Method

    The paper introduces MCLASH, a culturally adapted multilingual benchmark, and MET’s two-step theory-grounded prompting, extended by self-distillation in MET-D without external supervision.

    Source: §1 · p. 3, §3 · p. 5, §4.2 · p. 5

  • Results

    3.71 points in average F1 on MCLASH and 4.23 points on MMoralExceptQA; MET-D also increases native-language usage by 62.13 percentage points.

    Source: §1 · p. 3, §7 · p. 10

  • Takeaways & Limitations

    The most beneficial grounds align with each culture’s characteristics, while cross-lingual transfer follows linguistic typology rather than cultural similarity.

    Source: §1 · p. 3

    Ground selection remains dependent on the untrained base model because MET-D training targets theory-guided reasoning rather than selection.

    Source: Limitations · p. 11

Abstract

from arXiv · show

Language models are increasingly used for moral decision-making across diverse linguistic and cultural contexts, yet existing work overlooks multilinguality on three aspects: 1) multilingual evaluation benchmarks use direct translation, failing to adapt culture-specific items; 2) inference-time methods for moral reasoning rely on static, English-centric scaffolds and lack grounding in moral theory; 3) training methods for moral decision-making typically require expensive supervision from stronger models or human annotators. We address these gaps with three contributions. First, we introduce MCLASH, a multilingual moral decision-making benchmark to capture culturally situated moral intuitions and social norms across languages. Second, we propose MET (Multilingual Ethics with Theory-grounded reasoning), a two-step prompting method built on expert-curated, theory-based grounds drawn from psychology and philosophy: the model first selects situation- and culture-specific grounds, then reasons over them in the native language of the user. Third, we introduce MET-D (MET-Distillation), which enhances the second step through a self-distillation training stage that requires no external supervision. MET-D improves macro-F1 over the base model on all three models of different sizes and families (Qwen3-4B, Qwen3-8B, Gemma3-4B), by an average of 3.71 points on MCLASH and 4.23 on MMoralExceptQA, with a peak MCLASH gain of 12.94 points for Malay on Qwen3-8B. We further reveal that MET-D increases native-language reasoning by 62.13 points on average, and that beneficial grounds differ systematically across cultures. Together, these contributions open the path for culture-aligned, theory-grounded multilingual moral reasoning.

1 Introduction

The introduction frames culturally aligned multilingual moral reasoning as an overlooked need and presents MCLASH, MET, and MET-D as complementary responses.

Motivation

  • Existing multilingual benchmarks directly translate English dilemmas without adapting culture-specific items, while moral-reasoning methods often lack theory grounding, cultural adaptation, and multilingual support.The introduction notes that direct translation can import an inappropriate Western cultural context into scenarios translated into Korean or Hindi.Source: §1 · pp. 1–2

MCLASH

  • MCLASH culturally adapts CLASH into Chinese, Hindi, Korean, Malay, and Spanish, yielding 1,852 situations and 9,260 character descriptions across six languages.Qwen3-32B performed the adaptation, which native speakers then inspected.Source: §1 · pp. 2–3

MET

  • MET first selects situation- and culture-specific grounds from an expert-curated, psychology- and philosophy-based set, then generates a reasoning chain in the user's native language.The grounds are organized into six high-level dimensions; selections can differ across languages for the same situation.Source: §1 · pp. 2–3

MET-D

  • Because moral decisions lack verifiable ground truth, MET-D uses rejection-sampling self-distillation on synthetic CLASH-structured dilemmas, with automatically verifiable correct decisions as a proxy for substantive engagement with grounds.The introduction says prompting alone can yield shallow reasoning that engages with only a few selected grounds.Source: §1 · pp. 2–3

Results

  • MET-D improves average F1 by 3.71 points over the base model across three models; its largest gain is 12.94 points for Malay on Qwen3-8B.The models are Qwen3-4B, Qwen3-8B, and Gemma3-4B.Source: §1 · p. 3
  • MET-D gains average 4.23 points on MMoralExceptQA and 3.53 points on MoCa for Qwen3-4B, and increases native-language reasoning by 62.13 percentage points on Qwen3-4B.The most beneficial grounds align with cultural characteristics, while cross-lingual transfer follows linguistic typology rather than cultural similarity.Source: §1 · p. 3

2 Related Work

Prior work leaves gaps in culturally adapted multilingual moral decision-making datasets, context-sensitive theory-grounded prompting, and multilingual training for moral dilemmas. MCLASH and MET address the first two gaps, while MET-D introduces self-supervision for training.

2.1 Moral Decision-Making Datasets

2.2 Inference-Time Methods for Moral Decision-Making

  • Prior prompting applies static frameworks despite context-dependent moral frameworks; MET instead grounds reasoning in theory and adapts per instance and language for native-language, culture-aligned reasoning (AlKhamissi et al., 2024; Jin et al., 2022; Chakraborty et al., 2025; Fletcher, 1966; Dancy, 2004; Chiu et al., 2025).MoReBench rubrics are coarse, atheoretical, English-only, and labor-intensive, while prior theory-based scaffolds use a narrow, static, English-only set.Source: §2.2 · p. 4

2.3 Multilingual Training for Moral Decision-Making

3 Construction of MCLASH

MCLASH adapts CLASH situations and character descriptions for five languages and their corresponding cultures through model-assisted adaptation and native-speaker review, rather than direct translation.

3 Construction of MCLASH

  • The authors adapt CLASH into Chinese, Hindi, Korean, Malay, and Spanish using Qwen3-32B, selected for adaptation quality (Yang et al., 2025).They use language as a proxy for culture, considering all regions where each language is spoken.Source: §3 · p. 4
  • Native speakers review and modify narratives with a checklist, discard culturally irrelevant situations when necessary, and inspect adapted character descriptions in an additional round.The process also adapts character descriptions based on the revised situations.Source: §3 · pp. 4–5
  • Rather than directly translating instances, the process preserves cultural nuances, such as adapting “Amazon” to Korean companies “Naver” and “Kakao.”The authors describe direct translation as a limitation of previous work.Source: §3 · p. 5
  • MCLASH comprises 1,852 situations and 9,260 character descriptions.These instances were adapted from CLASH, which has 345 high-stakes situations paired with 3,795 character descriptions.Source: §3 · pp. 4–5

4 Method

The method pairs expert-curated theoretical grounds with two-step prompting in MET, then adds MET-D self-distillation to improve how models use selected grounds in reasoning. Its synthetic training data uses character descriptions to derive labels automatically, avoiding human annotation.

4.1 Theoretical Grounds

  • The expert-curated set contains 37 grounds across six dimensions, translated into all six languages, with descriptions supporting selection and a structure that allows extensions.The dimensions cover normative authority, value systems, ethical theories, cognitive reasoning strategies, conflict handling strategies, and moral uncertainty.Source: §4.1 · p. 5

4.2 Guiding Reasoning using Grounds

  • MET first selects situation-specific grounds to form an instance-specific framework, then uses those grounds with the situation to generate reasoning and an answer.The grounds are presented with their dimensions and descriptions during selection.Source: §4.2 · p. 5
  • MET-D adds self-distillation to MET using a CLASH-structured synthetic dataset whose character descriptions provide automatic ternary ground-truth labels, avoiding human annotation.Figure 3 depicts MET-D as adding this training stage before the two-step prompting procedure.Source: §4.2 · p. 5, Figure 3 · p. 6
  • 500 scenarios and 2,500 character descriptions per language are generated, covering six high-stakes topics equally across scenarios.For each dilemma, three values are sampled per side to generate scenarios, followed by five character descriptions per scenario with yes, no, or ambiguous labels.Source: §4.2 · pp. 5–6
  • Rejection sampling retains reasoning chains only when the model’s answer matches ground truth, then balances yes/no/ambiguous labels and training instances across languages.The target model is trained on the filtered chains; MET is then reapplied with the base model selecting grounds and the trained model reasoning.Source: §4.2 · p. 6

5 Main Results

Across three models and three benchmarks, the results show that training makes MET-D more reliable than prompting alone, with gains across resource levels and dataset types. Evaluation uses prompts and reasoning in each test instance’s language.

5 Main Results

  • Evaluation spans Qwen3-4B, Qwen3-8B, and Gemma3-4B on MCLASH, MMoralExceptQA, and MoCa, with prompts and reasoning in each test instance’s language.The evaluation includes an English-only benchmark with yes, no, and ambiguous labels; reasoning chains are enforced in the native language.Source: §5 · p. 6, Dataset Base MET MET-D Base MET MET-D… · p. 7

Dataset Base MET MET-D

  • Macro F1 over Base improves with MET-D on Qwen3-4B (+4.23), Qwen3-8B (+3.02), and Gemma3-4B (+0.63), except on MoCa for Gemma3-4B.MET prompting alone is inconsistent: it improves Qwen3-4B (+0.81) but lowers Qwen3-8B (-0.46) and Gemma3-4B (-3.46).Source: Table 1 · p. 7, Dataset Base MET MET-D Base MET MET-D… · p. 7
  • On MCLASH, MET-D gains are larger for low-resource than high-resource languages: +4.49 vs. +3.50 on Qwen3-4B and +5.25 vs. +2.36 on Gemma3-4B.The larger low-resource gain holds across all three evaluated models.Source: Dataset Base MET MET-D Base MET MET-D… · p. 7
  • MET-D generalizes to clear-cut MMoralExceptQA and ambiguous MoCa, except on MoCa for Gemma3-4B, where ambiguous F1 drops.Inspection suggests MET-D makes Gemma3-4B more inclined to give clear-cut yes/no answers.Source: §5 · p. 6, Dataset Base MET MET-D Base MET MET-D… · p. 7
  • Ablations show that reasoning-focused training does not transfer positively to ground selection, while theory-grounded prompting still outperforms direct answering after training.The default therefore uses the base model for ground selection and the trained model for theory-guided reasoning.Source: Dataset Base MET MET-D Base MET MET-D… · p. 7, Table 2 · p. 7

6 Analyses

The analyses test whether ground selection is meaningful, whether training increases use of selected grounds, and how effects vary across languages. They also examine native-language reasoning and cross-lingual transfer.

6.1 Selection Captures Relevant Grounds

  • Ground selection outperforms random selection across high- and low-resource languages, with a larger low-resource gain (+2.03); Cohen’s κ is 0.468 within models and 0.259 across models.The authors interpret this agreement as evidence that selection reflects systematic, non-random patterns despite task ambiguity.Source: §6.1 · p. 8, Table 3 · p. 8

6.2 Training Increases Utilization of the Selected Grounds

  • Training raises selected-ground utilization from 52.10% to 61.19% in English, 31.87% to 37.05% in Chinese, and 42.00% to 51.75% in Hindi.Utilization measures whether reasoning mentions and applies grounds, not whether it reconciles them; reconciliation remains future work.Source: §6.2 · p. 8

6.3 Beneficial Grounds are Different per Language

  • Every ground dimension has positive average gains, but effects vary by language; Moral Uncertainty lowers F1 in Korean and Chinese.Hindi benefits most broadly, which the authors attribute as likely related to weaker base-model structural priors for low-resource Hindi moral reasoning.Source: §6.3 · p. 8, Figure 4 · p. 9
  • Beneficial grounds reflect language-specific cultural patterns: English gains from rationalist grounds, Spanish from Relation-Based Authority, and Korean and Chinese from Contractarianism.English gains most from Probabilistic Moral Uncertainty (+9.6), First-principles reasoning (+9.0), and Maximin Reasoning (+4.9), associated with Anglo-American Analytic Philosophy; Spanish is led by Relation-Based Authority (+11.8), consistent with Latin familismo.Source: §6.3 · pp. 8–9, Figure 4 · p. 9

6.4 Training Improves Native-Language Reasoning

  • Training plus enforcement yields a +2.02-point gain, while training alone raises native-language reasoning from <1% to ≈90% across languages.Training without enforcement also outperforms the non-enforced base by +1.23 F1 percentage points; prompting alone to enforce native-language reasoning hurts performance on average by −2.17. Hindi and Chinese show exceptions to the overall pattern.Source: Figure 5 · p. 9, §6.4 · p. 10

6.5 Cross-Lingual Transfer Reflects Typology

  • Linguistic similarity, rather than cultural similarity, drives transfer: Korean and Hindi each improve the other’s F1 by 2.1 points, while Chinese and Korean transfer weakly.Chinese-to-Korean transfer is +0.8 and Korean-to-Chinese is −0.4; SVO Latin-script languages show positive mutual transfer, while Chinese breaks that pattern. Korean gains +9.8 points from own-language training, and training on a different language can help high-resource English and Chinese targets.Source: §6.5 · p. 10, Figure 6 · p. 10

7 Conclusion

The paper presents a culturally adapted moral-evaluation dataset and a theory-grounded reasoning method, then uses self-distillation to improve models’ adherence to theory-relevant grounds. Its results show gains in F1 and native-language reasoning, while beneficial grounds vary with culture and cross-lingual transfer follows linguistic typology.

7 Conclusion

  • Average F1 improves by 3.71 points on MCLASH and 4.23 points on MMoralExceptQA, while native-language usage increases by 62.13 percentage points.These results accompany the proposed method and self-distillation training strategy.Source: §7 · p. 10
  • The paper curates a culturally adapted evaluation dataset and grounds moral reasoning in psychological and philosophical theories.It carefully curates evaluation dimensions and theory-relevant grounds.Source: §7 · p. 10
  • The proposed method selects theory-relevant grounds before using theory-guided reasoning, and self-distillation trains models to adhere to those grounds.Source: §7 · p. 10
  • Beneficial grounds align with each culture’s characteristics, while cross-lingual transfer follows linguistic typology rather than cultural similarity.Source: §7 · p. 10

Limitations … A.2 Details of Adaptation

The paper identifies unresolved limitations in ground selection and reconciling multiple grounds, and describes how culturally adapted MCLASH items were selected, revised, and quality-checked.

Limitations

  • Training targets theory-guided reasoning, leaving ground selection to the untrained base model, while reconciling multiple selected grounds remains open.The authors identify addressing both limitations as future work.Source: Limitations · p. 11

A.1 Selection of the model for adaptation

  • Qwen3-32B adapted culturally specific items more effectively than GPT-4o, which mainly produced direct translations with limited cultural consideration.The authors selected Qwen3-32B for multilingual CLASH adaptation and had human annotators review and refine its outputs.Source: §A.1 · p. 16, Figure 7 · p. 17

A.2 Details of Adaptation

  • Adaptation first revises CLASH situations, which human inspectors evaluate and revise using a checklist before character descriptions are adapted.The process uses separate prompts for adapting situations and character descriptions.Source: §A.2 · p. 16
  • For adapted character descriptions, Qwen3-32B needed only manual correction of minor language issues, not the extensive checklist used for situations.One issue was occasional text in a language other than the target language.Source: §A.2 · p. 17
  • Because multiple entity revisions can suit a culture, the authors state that a single numeric inter-annotator agreement is not well-defined.For example, the U.S. company Meta could map to Naver or Kakao in Korea.Source: §A.2 · p. 17
  • Adaptation quality was checked with an inspection checklist and a second author pass back-translating narratives to flag misadapted or unadapted entities.Authors discussed flagged cases with annotators; an example was an American company name left unchanged.Source: §A.2 · p. 17

A.3 Dataset Statistics

The section describes MCLASH dataset statistics after filtering out culturally irrelevant or inadequately adapted situations. Culturally adapted variants can have substantially higher token counts than English.

A.3 Dataset Statistics

  • The dataset statistics exclude situations judged culturally irrelevant or inadequately adapted during curation.Table 4 reports situation counts and average token counts for Situation, Action, and Character Descriptions by language.Source: §A.3 · p. 17
  • Culturally adapted variants have higher average token counts than English, with increases of up to six times; prior work reports differences of up to 5× (Ahia et al., 2023).The statistics compare token counts for Situation, Action, and Character Descriptions across languages.Source: §A.3 · p. 17

B Method … C.2 Full Numbers for the Native Language Reasoning Experiment

The appendices specify MET’s selectable grounds and prompts, document full benchmark results, and describe the training-data and native-language reasoning setups.

B.1 Details of Grounds and Prompts

  • The prompts specify ground selection and descriptions (Figures 14–15) and response generation from selected grounds (Figure 16).Source: §B.1 · p. 17
  • MET provides selectable grounds and descriptions across six dimensions for the target model.Source: Figure 13 · p. 23

B.2 Prompts Used for Generating the Training Dataset

  • The prompt for extracting values from the WVS questionnaire is shown in Figure 17.Source: §B.2 · p. 23

C.1 Full Results and Comparison with MoReBench Dimensions

  • Table 5 reports that MET outperforms MoRe and MET-D outperforms MET-D(s) on Qwen3-4B.MET-D(s) uses situation-only training data and lacks verifiable ground truth for rejection sampling.Source: §C.1 · p. 23
  • Tables 5–7 present full per-language results for MCLASH, MMoralExceptQA, and MoCa across Qwen3-4B, Qwen3-8B, and Gemma3-4B.Source: §C.1 · p. 23

C.2 Full Numbers for the Native Language Reasoning Experiment

  • Table 8 presents four native-language reasoning settings: the base model without enforcement, with prefill-token enforcement, trained plus prompted without enforcement, and full MET-D with training, prompting, and enforcement.Source: §C.2 · p. 23

C.3 Comparison of the Ability of Theory-Guided Reasoning of the Base Model versus the Trained Model

On the English 854-ama case, MET-D engaged with eight of ten selected grounds versus three for Base, but broader framework coverage did not produce explicit reconciliation of conflicting verdicts.

C.3 Comparison of the Ability of Theory-Guided Reasoning of the Base Model versus the Trained Model

  • The case asks whether Dr. Allworthy should share healthcare-policy political views with patients, using ten grounds selected identically for both models.The grounds include Institution-Based, Responsibility-Based, Agreement-Based, and Right-Based.Source: §C.3 · p. 30
  • Base substantively invoked 3/10 selected grounds, while MET-D engaged with eight, adding four frameworks.The added frameworks were Right-Based, Contractarianism, Stakeholder, Harm–Benefit, and Principle-Based.Source: §C.3 · pp. 30–31, Figure 18 · p. 31
  • Despite broader engagement, MET-D gives one-line verdicts per framework and concludes yes without explicitly weighing conflicting verdicts against one another.Its internal trace includes conflicting framework-level verdicts, including “probably not” and “yes.”Source: §C.3 · p. 31

Related papers

Loading 2607.11736v1…