BrowseComp

A benchmark of 1,266 hard-to-find, multi-hop web-browsing questions whose answers are difficult to locate but easy to verify, measuring an agent's ability to persistently search and synthesize information from the web.