Web research

BrowseComp

Find hard-to-locate facts by browsing many websites.

In plain words

Like a scavenger hunt on the internet: the answer is short and easy to check, but finding it may take hours of digging.

An AI grader checks whether the answer means the same thing as the reference answer.

Humans: Human trainers solved 29% and gave up on 71% after two hours each source

Research assistants are judged on persistence. This measures whether AI keeps searching until it finds the answer.

Questions were built backward from a known fact, so other valid answers may occasionally exist.
Measures finding facts, not writing long reports or handling vague requests.
The authors ask that example questions not be posted online.