BabelSafe: A Policy-Grounded Multilingual Safety Benchmark for LLMs
BabelSafe introduces ML-BENCH, a multilingual safety benchmark built from 17 regional AI regulations across 14 countries and 14 languages, grounding safety in specific legal contexts. It also presents ML-GUARD, a diffusion Large Language Model guardrail that achieves state-of-the-art performance on this new benchmark. This work enables culturally and legally aligned safety evaluations beyond general risk taxonomies.