CorporateBench scales enterprise Q&A to 230K documents

CorporateBench evaluates five LLMs across temporal enterprise corpora exceeding 230,000 documents and reports worsening performance as input size approaches realistic scale, in the EMNLP Findings paper. Its four synthetic firms range from 12 to 10,000 employees and preserve cross-document consistency, making it a sharper stress test than a small, static RAG set; builders should measure retrieval and answer consistency against growing, time-aware corpora alongside the archive’s agent-search benchmark economics.