BabelSafe: A policy-grounded multilingual LLM safety benchmark

BabelSafe is a multilingual LLM safety benchmark covering 13 language environments, with risk categories and rules extracted from regulatory documents of various regions. The paper also proposes BabelGuard, a guardrail model based on a diffusion large language model, with a 1.5B version for fast safe/unsafe classification and a 7B version supporting customizable policy-conditioned safety checks. It performs strongly against 11 guardrail baselines on 6 multilingual safety benchmarks and BabelSafe.