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

The Problem: Language Models Make Cultural Mistakes in Moral Judgment

Language models are now deployed to make ethical decisions—from content moderation to loan applications—across the world. Yet these systems stumble on a fundamental issue: they're built and evaluated in English, then naively translated into other languages and cultures. A moral dilemma that matters deeply in one cultural context might not translate directly into another. Existing benchmarks don't capture this. Inference methods rely on rigid, English-first reasoning scaffolds. And training these systems demands expensive human experts or stronger models to supervise the learning. The authors show these gaps are real and fixable.

The Three-Part Solution

The benchmark: MCLASH

The authors built a multilingual moral decision-making benchmark that goes beyond translation. Rather than simply rendering English moral dilemmas into other languages, MCLASH adapts items to reflect culture-specific intuitions and social norms. This means a moral scenario that fits Japanese context is not the same as a translated English scenario—it reflects what actually troubles or guides people in that culture.

The inference method: MET

The core innovation is a two-step prompting approach grounded in moral theory. Instead of generic reasoning chains, the authors worked with philosophers and psychologists to extract theory-based "grounds"—principles and considerations from philosophy and psychology that justify moral decisions. The method works like this: given a moral scenario, the model first selects which grounds are relevant to that specific situation and culture. Then it reasons through those grounds in the user's native language, not English. This keeps reasoning culturally and linguistically authentic rather than forcing everything through an English-centric filter.

The training method: MET-D

Stronger models and human annotators are expensive. Instead, the authors use self-distillation: a smaller model learns by studying the reasoning patterns of the same model when given better prompts and grounds. No external supervision required. This training stage polishes the second reasoning step.

Results and Impact

The gains are substantial. On MCLASH and another existing benchmark (MMoralExceptQA), models improved by an average of 3.71 and 4.23 macro-F1 points respectively. For specific language-model pairs, the jump was dramatic—Qwen3-8B's Malay performance spiked by 12.94 points. The authors also found that native-language reasoning increased 62.13 points on average after training, confirming their method actually shifts models toward reasoning in the user's own language instead of translating to English internally.

Critically, the authors show that beneficial moral grounds are not universal: different cultures rely on different ethical principles to reach decisions. This validates the core insight—one-size-fits-all moral AI is inadequate.

The work matters because billions of people interact with language models in non-English contexts, often for high-stakes decisions. MET demonstrates that culture-aware, theoretically grounded multilingual moral reasoning is both possible and measurable. It's a step toward AI systems that respect cultural moral intuitions rather than imposing English-language ethics globally.