MET: Advancing Multilingual Moral Reasoning in Language Models with Culture-Aware Theory

Short summary

Researchers introduce MET, a two-step prompting methodology for multilingual moral reasoning in language models that selects culture-specific moral grounds then reasons in the user's native language. They also release MCLASH, a benchmark for culturally situated moral intuitions, and MET-D, a self-distillation approach that improves reasoning without external supervision. MET-D achieved macro-F1 gains up to 12.94 points on Malay using Qwen3-8B.

• •MET uses two-step theory-grounded prompting: select culture-specific moral grounds, then reason in the user's native language
• •MCLASH is a new multilingual benchmark focusing on culture-specific moral intuitions rather than direct translations
• •MET-D self-distillation improved macro-F1 by up to 12.94 points on Malay with Qwen3-8B, eliminating need for external supervision

Generated with AI, which can make mistakes.