Benchmark RadarXIH-Bench
Evaluates instruction hierarchy compliance in multilingual LLMs using same-language and cross-language conflicts across six languages, four domains (rule-following, safety, task-execution, persona), and three hierarchy types (system-user, system-tool, user-tool).
- Released
- 2026-07-26
- Readiness
- Runnable
- Primary field
- General AI
Why it matters
Existing instruction hierarchy benchmarks are largely English-centric, leaving a gap in assessing multilingual safety and reliability. This benchmark provides a reusable protocol to measure how language choice affects model compliance, supporting safer deployment in multilingual contexts.
Motivation
Instruction hierarchy (IH) requires models to prioritize instructions by source, ensuring that higher-priority instructions override lower-priority ones.
Primary resources
Benchmark Radar records only publicly supported details and links back to primary sources for verification.