VisibilityOS
notableReviewed Sun Nov 01

XBCP research finds language mismatch weakens agentic retrieval and citation fidelity

Lu and colleagues introduce Cross-lingual BrowseComp-Plus (XBCP), keeping questions and answers in English while varying the languages of supporting evidence. They evaluate four deep-research agents and sparse/dense multilingual retrievers; the multilingual condition distributes documents across 12 languages. Authors report lower evidence recall, answer accuracy and citation reliability with language-mismatched evidence.

Context

Lu and colleagues introduce Cross-lingual BrowseComp-Plus (XBCP), keeping questions and answers in English while varying the languages of supporting evidence. They evaluate four deep-research agents and sparse/dense multilingual retrievers; the multilingual condition distributes documents across 12 languages. Authors report lower evidence recall, answer accuracy and citation reliability with language-mismatched evidence.

What changed

The June 13, 2026 original research preprint introduces a controlled cross-lingual deep-research benchmark, separating retriever failures from downstream agent evidence-integration failures, including oracle-retrieval experiments.

Why it matters

AI discovery of non-English sources can fail at multiple stages: retrieving relevant evidence and faithfully integrating that evidence into a cited answer. Visibility and citation metrics need language-aware, stage-specific evaluation.

Visibility OS interpretation

Visibility OS interpretation: Multilingual discoverability should be assessed at retrieval, evidence inclusion and citation fidelity stages separately. The findings concern controlled fixed-corpus agents and do not prove production ChatGPT, Gemini or Perplexity behavior.

What remains uncertain

How much degradation comes from translation artifacts versus cross-language reasoning? Do live search systems show similar citation-fidelity gaps? How do results change with native-language questions and non-translated corpora?

Evidence claims

introduces

In the XBCP controlled benchmark, four deep-research agents and multilingual retrievers were tested on language-mismatched supporting evidence; the authors report degraded evidence recall and citation reliability, with some answer-accuracy degradation persisting even when gold evidence is provided.