BrowseComp-Plus

A deep-research benchmark that replaces live web search with a fixed corpus of about 100K documents, so agents and retrievers can be compared fairly.

Source: BrowseComp-Plus: A More Fair and Transparent Evaluation Benchmark of Deep-Research Agent.

It's built from OpenAI's BrowseComp, 1,266 hard-to-find fact questions answered by browsing the live web. Live search results change over time and are a black box, which makes runs hard to reproduce and compare. BrowseComp-Plus keeps 830 of the questions, each with human-verified evidence documents and mined hard negatives, giving a corpus of 100,195 documents.

Referenced in Context Language Models.