What's this about?
The benchmark contains 423 manually authored questions in 13 languages that require obscure, multi-step evidence across web pages, videos, images, scans and maps.
What exactly is parametric knowledge?
It is information stored in a model's weights during training. The benchmark filters out easier questions so success depends on live investigation rather than remembered facts.
Why does this matter?
Many browsing tests reward English text lookup. This one better measures persistence, multilingual search and evidence handling across media, though its value will depend on keeping web answers and evaluation current.