XHotpotQA Benchmark for Cross-Lingual Multi-Hop Question Answering
August 31, 2026
XHotpotQA provides 15,661 training and 7,405 validation instances to test reasoning across language boundaries in multi-hop QA. The benchmark uses evidence-dependency graphs where question, bridge, and answer elements reside in different languages to expose reasoning failures.
HOW THIS AFFECTS YOU
●
builderUse this to stress-test the robustness of multilingual RAG systems.
●
researcherThis benchmark allows you to evaluate if models fail due to language mismatch during reasoning.