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

Multilingual moral reasoning benchmarks and methods often overlook cultural adaptation, theory grounding, and affordable supervision. This paper introduces MCLASH, MET, and self-distilled MET-D, which improve average F1 by 3.71 points on MCLASH and 4.23 on MMoralExceptQA.

  • Problem

    Multilingual moral decision-making research overlooks culturally aligned evaluation and methods that support users’ linguistic and cultural contexts.

  • Method

    The paper introduces a culturally adapted benchmark, theory-grounded two-step prompting, and self-distillation requiring no external supervision.

  • Results

    3.71 average F1 points on MCLASH and 4.23 on MMoralExceptQA are achieved over the base model across evaluated models.

  • Takeaways & Limitations

    Beneficial moral grounds differ systematically across cultures and align with their cultural characteristics.

  • Takeaways & Limitations

    Ground selection still relies on the untrained base model, while reconciliation of multiple selected grounds remains unresolved.

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 paper introduces MCLASH, a culturally adapted multilingual moral decision-making benchmark, and MET, a theory-grounded prompting method with self-distillation through MET-D. MET-D improves performance across models and datasets while substantially increasing native-language reasoning.

  • Motivation: Existing multilingual moral benchmarks directly translate scenarios without adapting culture-specific entities, institutions, currencies, or customs, undermining culturally aligned moral judgment.Such deployment matters because language models increasingly support high-stakes domains including law and healthcare across diverse cultural backgrounds.
  • Benchmark: MCLASH contains 1,852 culturally adapted long-form moral scenarios across six languages, addressing the cultural distortions introduced by directly translating English benchmarks.The benchmark adapts CLASH (Lee et al., 2025) into Chinese, Hindi, Korean, Malay, and Spanish, with native-speaker inspection and 9,260 character descriptions.
  • Method: MET first selects situation- and culture-specific moral grounds, then uses them to generate a grounded reasoning chain in the user’s native language.Its grounds are expert-curated from psychology and philosophy literature, organized into six high-level dimensions, and translated into all six languages.
  • Method: MET-D applies rejection-sampling self-distillation to a synthetic CLASH-structured dataset, using automatically verifiable correct decisions as a proxy for substantive engagement with moral grounds.This avoids supervision from stronger models or extensive human involvement.
  • Results: 3.71 points average F1 improvement on MCLASH over the base model across Qwen3-4B, Qwen3-8B, and Gemma3-4B, with a 12.94-point Malay gain on Qwen3-8B.MET-D also improves average F1 by 4.23 points on MMoralExceptQA and 3.53 points on MoCa for Qwen3-4B.
  • Results: 62.13 points average improvement in native-language reasoning demonstrates that MET-D strengthens reasoning in the user’s language, while beneficial grounds differ systematically across cultures.For the same situation, English selects First-principles reasoning whereas Korean selects Contractarianism, reflecting different cultural values.

2 Related Work

Prior moral decision-making research often fails to evaluate actual decisions, account for multilingual cultural variation, or support context-sensitive, native-language reasoning. Existing training approaches also depend on costly external supervision and rarely address multilingual moral dilemmas where values conflict.

  • 2 Related Work: Prior work often probes model preferences instead of decisions, while multilingual studies overlook cultural differences or use short, directly translated scenarios.These limitations make actual decision-making difficult to assess and reduce narrative and cultural adequacy.
  • 2 Related Work: Structured prompting typically applies one moral framework to every dilemma, despite moral frameworks needing to vary with context.MoReBench introduces per-instance adaptation through evaluative rubrics, but the supplied passage does not specify its limitations beyond noting that rubrics can be adapted.
  • 2 Related Work: Non-English reasoning methods in conventional domains commonly translate or align reasoning toward English, assuming identical reasoning structures across languages.Reported techniques align non-English representations with English, optimize similarity between reasoning chains, or connect multilingual encoders to English-trained reasoners.
  • 2 Related Work: Existing moral-reasoning training uses handcrafted rewards or stronger models for external supervision, operates in English, and rarely trains multilingual reasoning through value-conflict dilemmas.Multilingual training work is largely evaluation-focused or targets cultural commonsense using cultural surveys rather than reasoning through conflicting moral values.

3 Construction of MCLASH

MCLASH adapts CLASH into five languages and corresponding cultures through model-assisted cultural adaptation and native-speaker review, rather than direct translation. The resulting benchmark contains 1,852 situations and 9,260 character descriptions.

  • Human validation: Native speakers review and modify adapted situations using a rigorous checklist, then adapt character descriptions and conduct an additional human inspection.The adaptations are intended to preserve culturally specific nuances rather than merely reproduce the English narratives.
  • Cultural scope: Language serves as a proxy for culture, covering all regions where each selected language is spoken.The paper notes that cultural boundaries are dynamic and can also be defined by religion or geography.
  • Cultural adaptation: The construction process avoids direct translation by adapting narratives and culturally specific references, such as replacing “Amazon” with Korean companies including “Naver” and “Kakao.”Qwen3-32B performs the initial adaptation, followed by native-speaker modification and human inspection.
  • Dataset construction: MCLASH contains 1,852 situations and 9,260 character descriptions after adapting CLASH into Chinese, Hindi, Korean, Malay, and Spanish.The original CLASH contains 345 situations paired with 3,795 character descriptions.

4 Method

The method curates multilingual, theory-based grounds and uses them in MET’s two-step selection-and-reasoning procedure. MET-D adds self-distillation with synthetically generated, rejection-filtered training data to improve the use of selected grounds.

  • Theoretical grounds: The curated resource contains 37 theoretical grounds across six dimensions, translated into all six languages and paired with descriptions for relevance-based selection.Its uniform structure allows new grounds to be added by specifying a dimension and description without changing the selection process.
  • MET: MET first selects situation- and language-specific grounds, then generates reasoning and an answer from those grounds and the situation.Ground selection forms an instance-specific framework before theory-guided reasoning.
  • MET-D: MET-D adds self-distillation to MET after models often leave selected grounds unaddressed, producing shallow reasoning.The training stage uses synthetic data and is followed by MET prompting to enhance utilization of selected grounds during theory-guided reasoning.
  • MET-D: MET-D constructs synthetic dilemmas by sampling values, generating scenarios and character descriptions, and assigning yes, no, or ambiguous ground-truth decisions.The value set is derived from the WVS questionnaire using Qwen3-32B and GPT-oss-120B, while scenarios are generated with Qwen3-32B.
  • MET-D: MET-D retains only target-model reasoning chains whose final answers match ground truth, balances decisions equally, equalizes training instances across languages, and retrains the model.The filtered chains are then used for training before the two-step prompting procedure is reapplied.

5 Main Results

MET-D consistently improves over Base across the three evaluated models, while ablations show that training primarily benefits theory-guided reasoning rather than ground selection.

  • Main Results: The evaluation spans Qwen3-4B, Qwen3-8B, and Gemma3-4B across MCLASH and two additional datasets, using Qwen3-4B as the default model unless otherwise specified.The default model was chosen for its multilingual capabilities and inspectable internal reasoning chains.
  • Main Results: MET-D improves over Base on all three models, yielding gains of +4.23, +3.02, and +0.63, with only Gemma3-4B on MoCa as an exception.Table 1 reports macro-F1 results averaged across languages and grouped by resource level.
  • Main Results: MET improves over Base by +0.81 on Qwen3-4B but degrades by -0.46 on Qwen3-8B and -3.46 on Gemma3-4B, whereas MET-D restores consistent gains.Prompting alone is therefore unreliable across model sizes and families, while adding training closes the gap.
  • Ablation Study: MET-D achieves the best ablation performance, indicating that training mainly improves theory-guided reasoning rather than ground selection.Replacing the base selector with the trained selector performs worse, while the trained model’s principal benefit lies in reasoning.

6 Analyses

The analyses show that MET-D selects stable, relevant grounds, improves their use and native-language reasoning, and reveals culture- and language-specific effects. They also identify a remaining limitation: increased ground engagement does not guarantee reconciliation.

  • §6.1: Model-based ground selection outperforms random selection across language-resource settings, with a larger low-resource gain of +2.03; Cohen’s κ is 0.468 within models and 0.259 across models.These results support selecting relevant grounds rather than random noise.
  • §6.2: Training increases 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 weighs and integrates them; reconciliation remains future work.
  • §6.3: Beneficial grounds vary across languages: every dimension has positive average gain, but Moral Uncertainty hurts Korean and Chinese, while English and Spanish favor culturally distinct grounds.English gains most from Probabilistic Moral Uncertainty (+9.6), First-principles reasoning (+9.0), and Maximin Reasoning (+4.9); Spanish is led by Relation-Based Authority (+11.8).
  • §6.4: Prompt-only native-language reasoning lowers average performance by −2.17, whereas training raises native-language reasoning from <1% to ≈90% and improves F1 by +1.23 without enforcement.Combining training with enforcement yields a further +2.02-point gain, breaking the legibility–performance trade-off.
  • §6.5: Cross-lingual transfer follows linguistic rather than cultural similarity: Korean and Hindi improve each other’s F1 by 2.1 points, while Chinese breaks the SVO Latin-script pattern.For some high-resource targets, including English and Chinese, target-language training hurts performance while different-language training helps, suggesting multilingual exposure can mitigate reinforced cultural biases.

7 Conclusion

The paper contributes a culturally adapted evaluation dataset and a theory-grounded method for multilingual moral reasoning. It combines ground selection, theory-guided reasoning, and self-distillation to improve adherence to selected grounds.

  • 7 Conclusion: The paper curates a culturally adapted evaluation dataset for moral reasoning.
  • 7 Conclusion: Its method grounds moral reasoning in psychological and philosophical theories through carefully curated dimensions and theory-relevant grounds.
  • 7 Conclusion: The proposed two-step process selects grounds before performing theory-guided reasoning.
  • 7 Conclusion: A self-distillation training strategy is designed to improve models’ adherence to the selected grounds during reasoning.

Limitations … A.2 Details of Adaptation

The approach trains theory-guided reasoning but leaves ground selection and reconciliation among multiple grounds unresolved. MCLASH adaptation combines model selection, human inspection, manual linguistic correction, and back-translation checks to handle culturally variable revisions.

  • Limitations: The main limitations are reliance on the untrained base model for ground selection and unresolved reconciliation of multiple selected grounds.The authors identify both issues as promising directions for future work.
  • 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 therefore selected Qwen3-32B for multilingual CLASH adaptation and had human annotators review and refine its outputs.
  • A.2 Details of Adaptation: Situations are adapted with a prompt, reviewed against a checklist, revised by inspectors, and then used to adapt character descriptions.The adaptation prompts and inspection checklists are provided in Figures 8 through 12.
  • A.2 Details of Adaptation: Character descriptions required only manual correction of occasional irrelevant-language outputs because Qwen3-32B otherwise adapted them well.Unlike situations, character descriptions did not receive an extensive inspection checklist.
  • A.2 Details of Adaptation: Because culturally appropriate entity revisions can have multiple valid answers, a single numeric inter-annotator agreement is not well-defined.The example maps Meta to a large Korean technology company such as Naver or Kakao.
  • A.2 Details of Adaptation: Adaptation quality is instead controlled through a rigorous inspection checklist and a second author-led back-translation pass to flag misadapted narratives.The supplied passage ends while describing the back-translation quality-control procedure.

A.3 Dataset Statistics

This section summarizes MCLASH’s descriptive statistics and curation, including language-specific situation and description lengths after culturally irrelevant or inadequately adapted datapoints were removed. Culturally adapted variants generally use more tokens than English, sometimes by up to six times.

  • A.3 Dataset Statistics: MCLASH statistics report the number of situations and average tokens in Situation, Action, and Character Descriptions for each language.These statistics are presented in Table 4.
  • A.3 Dataset Statistics: The curation pipeline discards situations that are culturally irrelevant or inadequately adapted by the model.This filtering follows the cultural adaptation stages described in Section 3.
  • A.3 Dataset Statistics: Up to 6× more tokens occur in some culturally adapted variants than in English, consistent with cross-language tokenization differences reported by Ahia et al. (2023).Prior work found equivalent content can differ by up to 5× in token count compared with English.

B Method … C Results

The method builds culturally adapted moral scenarios and theory-grounded prompts around six dimensions, while the supplied passages provide no substantive empirical findings for C Results.

  • B.1 Details of Grounds and Prompts: The response-generation instructions operate on the grounds selected for each situation and are presented separately from the ground-selection prompts.Figure 16 contains the instructions for generating responses based on selected grounds.
  • B.1 Details of Grounds and Prompts: The benchmark adapts each English dilemma’s situation, action, and supporting and opposing reasons to a target language and culture.The adaptation prompt treats the data as Situation, Action, and Reasons, with Reasons divided into supporting and opposing reasons.
  • B.1 Details of Grounds and Prompts: Character descriptions are culturally adapted from English descriptions that intertwine the original supporting and opposing reasons.This process is specified by a dedicated prompt for adapting character descriptions.
  • B.1 Details of Grounds and Prompts: Human inspectors revise generated situations, actions, and reasons using general and task-specific checklists for coherence, consistency, meaning, and naturalness.The checklist requires logical correctness, thematic consistency, equivalent meaning when culturally adapted, and fluent native-language expression.
  • B.1 Details of Grounds and Prompts: Its theory-grounded inventory spans normative authority, value systems, ethical theories, cognitive reasoning, conflict handling, and moral uncertainty.The listed resources include Moral Foundations Theory, Schwartz’s Value System, deontology, utilitarianism, stakeholder analysis, counterfactual reasoning, context-sensitive resolution, and probabilistic moral uncertainty.
  • B.1 Details of Grounds and Prompts: The framework organizes candidate reasoning grounds into six dimensions, with selectable grounds and descriptions presented for model prompting.The dimensions, grounds, and descriptions are shown in Figure 13, while Figures 14 and 15 provide selection prompts and descriptions.
  • B.2 Prompts Used for Generating the Training Dataset: The supplied passages identify a prompt for extracting values from the World Values Survey questionnaire but provide no empirical findings for C Results.The extraction prompt is shown in Figure 17; result tables, comparisons, and quantitative outcomes are not included in the supplied passages.
  • C Results: The supplied passages contain no substantive result passage for C Results.Accordingly, no benchmark comparison or quantitative result can be reported from the provided evidence.

C.1 Full Results and Comparison with MoReBench Dimensions · C.2 Full Numbers for the Native Language Reasoning Experiment

The paper reports full per-language benchmark results across three models and compares MET with MoReBench dimensions and a situation-only variant. It also provides a four-condition experiment isolating native-language reasoning enforcement and MET-D’s combined components.

  • C.1 Full Results and Comparison with MoReBench Dimensions: Tables 5–7 report full per-language results for MCLASH, MMoralExceptQA, and MoCa on Qwen3-4B, Qwen3-8B, and Gemma3-4B.Table 5 covers Qwen3-4B, Table 6 Qwen3-8B, and Table 7 Gemma3-4B.
  • C.1 Full Results and Comparison with MoReBench Dimensions: For Qwen3-4B, Table 5 compares MET with MoRe, which replaces the theoretical grounds with MoReBench dimensions.3This comparison tests the proposed theoretical grounds against MoReBench dimensions.
  • C.1 Full Results and Comparison with MoReBench Dimensions: Table 5 also evaluates MET-D(s), a situation-only variant trained with balanced 1:1:1 yes/no/ambiguous reasoning-chain labels.Because MET-D(s) lacks character descriptions, it has no verifiable ground truth.
  • C.2 Full Numbers for the Native Language Reasoning Experiment: Table 8 compares a Qwen3-4B base model without language enforcement against explicit native-language prefill-token enforcement.This isolates the effect of enforcing native-language reasoning in the base model.
  • C.2 Full Numbers for the Native Language Reasoning Experiment: Table 8 compares the trained model with prompting but no enforcement against MET-D using training, prompting, and enforcement.The comparison evaluates the full combination of MET-D components against prompting without enforcement.
  • C.2 Full Numbers for the Native Language Reasoning Experiment: The native-language reasoning experiment therefore reports four configurations: base without enforcement, base with prefill enforcement, trained prompting without enforcement, and full MET-D.All four configurations are presented in Table 8.

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

The comparison shows that MET-D invokes substantially more theory-guided grounds than the base model in the example, but its expanded reasoning does not reconcile conflicting framework verdicts. The trained model therefore produces more extensive, theory-linked reasoning while retaining an explicit integration limitation.

  • Base-model reasoning: The base model also identified patient autonomy and professional boundaries, but its reasoning was less substantively grounded and more tentative.It considered whether sharing political information could empower patients while remaining within professional boundaries, ultimately answering yes and no to the two questions.
  • Comparison of base and trained models: MET-D invoked 8/10 grounds versus 3/10 for the base model, but listed them without reconciliation.The base model substantively invoked 3 of 10 grounds, whereas MET-D invoked 8 of 10; the trained model’s engagement still did not translate into reconciliation.
  • MET-D reasoning: MET-D linked its verdicts to specific theories and contextual conditions, treating the action as acceptable when framed as policy insight rather than personal opinion.Its reasoning invoked normative authority, patient autonomy, professional responsibility, and boundary risks when evaluating the doctor scenario.
  • Limitation: MET-D issued separate one-line verdicts for each ethical framework and concluded yes without explicitly weighing their conflicting conclusions.Its final emotional-response answer remained context-dependent, acknowledging that boundary crossing could produce discomfort.
Loading 2607.11736v1…