Michigan, Railton unveil MET: culture-aware LLM moral benchmark

When alignment researchers check whether a language model reasons morally in a non-English language, the standard move is to take an English benchmark and translate it. A COLM 2026 paper, MET: Theory-Grounded and Culture-Aware Multilingual Moral Reasoning, argues that is where the evaluation quietly breaks. Ayoung Lee, Ryan Kwon, Yunxiang Zhang, Yuxuan Liu, Peter Railton, and Lu Wang describe a benchmark and a method built around culture-specific moral intuitions instead of translated English ones.

The approach reported in the abstract has three pieces. MCLASH is a multilingual benchmark meant to capture culture-specific moral intuitions across languages. MET is a two-step prompting method that leans on 'expert-curated, theory-based grounds drawn from psychology and philosophy' so a model first picks the right grounds and then reasons in the native language. MET-D is a self-distillation training approach that avoids expensive external supervision.

The numbers reported are modest but pointed. Macro-F1 improvements average 3.71 points on MCLASH and 4.23 on MMoralExceptQA across three models, with a peak MCLASH gain of 12.94 points for Malay on Qwen3-8B, and native-language reasoning reportedly increased by 62.13 points on average. Take those as reported, not settled: this is a fresh conference paper and the gains are relative to specific baselines and models.

Why it matters if you are not building benchmarks yourself: safety and alignment evaluations are largely written and validated in English then translated, and if translation systematically misses the moral intuitions of the underlying culture, a chunk of what gets called multilingual alignment is measuring the wrong thing. What the abstract does not settle is whether the theory-based grounds generalise cleanly to cultures the authors did not sample, or how the culture-native rewrites hold up under expert review from inside those cultures.

The upside is for anyone shipping models outside English-first markets. A culture-native benchmark plus a supervision-free training recipe gives smaller teams a way to check their models on non-English moral reasoning without paying for bespoke annotation, and the Railton co-authorship gives the moral-philosophy-plus-AI-safety pitch a fresh reference point.