Definition
Plain language
A version of a web-research benchmark where every document is labeled as helpful, misleading, or irrelevant.
As stated in the literature
A variant of BrowseComp annotating each retrieval-index document as supporting evidence, gold, hard negative, or irrelevant, enabling controlled evidence-blocking experiments.
Why it matters: By labeling every document, it lets researchers deliberately remove or block evidence and test whether a search agent truly relies on what it finds.
For example, a single document in its index might be tagged as the gold answer, a misleading look-alike, or simply irrelevant to the question.
Heard on the show
“There's a companion benchmark called BrowseComp-Plus.”Episode 092 — When Search Agents Don't Really Search: The Memory Shortcut Hiding in Browsing Benchmarks