Definition
Plain language
A web-research benchmark built from very recent, obscure facts the model couldn't have memorized, so it actually has to search.
As stated in the literature
A browsing-agent benchmark constructed from recent (within ~90 days) long-tail events across multiple domains, designed to suppress intrinsic-knowledge shortcuts and measure genuine search ability.
Why it matters: By using fresh, niche facts it measures real search ability rather than rewarding a model for guessing from memory.
For example, it asks about an obscure event from the last few months that the model couldn't have memorized, forcing it to actually go search.
Heard on the show
“The paper we're working from is called "LiveBrowseComp: Are Search Agents Searching, or Just Verifying What They Already Know?”Episode 092 — When Search Agents Don't Really Search: The Memory Shortcut Hiding in Browsing Benchmarks