XHotpotQA: A Benchmark for Cross-Lingual Knowledge Composition in Multi-Hop Question Answering

XHotpotQA tests whether an AI system can connect facts when the question and supporting passages use different languages. It supplies both relevant and distracting passages, allowing researchers to examine evidence selection separately from answer construction. This exposes language boundaries inside a reasoning task that translating everything into one language would conceal. The paper describes the benchmark and its controlled setup.

The authors report larger language-associated performance gaps for answer-producing systems than for their adapted evidence selector. Within this test, finding relevant passages was less sensitive to language mismatch than using them to answer. The supplied passages make this a test of evidence integration, not unrestricted web search.