ML-Bench&Guard: Policy-Grounded Multilingual Safety Benchmark Aligns with Regional Regulations
arXiv·medium signal
ML-Bench introduces a multilingual safety benchmark that moves beyond general risk taxonomies and machine-translated prompts to align with region-specific regulations and cultural nuances. The companion guardrail model is designed to enforce policy-grounded safety rather than predefined category lists. For teams deploying LLMs across regulatory jurisdictions, this provides the first benchmark that maps to actual legal requirements rather than abstract harm categories.