Hyper-BrowseComp

Unlike English-centric browsing benchmarks, Hyper-BrowseComp asks questions in many source languages and often requires models to work across multilingual text, audio, and video before they can settle on an answer. Current coverage includes the languages below, with more to be added.

• Chinese
• Egyptian Arabic
• German
• Indonesian
• Javanese
• Kazakh
• Korean
• Russian
• Spanish
• Telugu
• Thai
• Uzbek
• Vietnamese

We are strongly inspired by BrowseComp, but Hyper-BrowseComp pushes beyond trivia-like intersections of vague descriptions. Many tasks require locating specific information inside books, PDFs, papers, videos, official tables, maps, and other sources that are unlikely to be stored directly in model weights.

Each task still has a compact answer, but reaching it can require identifying entities, disambiguating records, calculating from data, or aggregating many small clues. The benchmark rewards careful search strategy, source selection, and final-answer verification.

Beyond plain text search, Hyper-BrowseComp can require models to interpret or aggregate information from images, audio, video, maps, and interactive websites. A solver may need to combine a video frame, a public record, a map position, and a current storefront before the answer becomes clear.

This makes the benchmark closer to real browsing work: the relevant evidence is often distributed across modalities, interfaces, and languages, and the model must decide which tool or source is worth consulting next.