AI
HyperBrowseComp: A Multilingual and Multimodal Stress Test for Web-Browsing Agents
Publication: Hugging Face Daily Papers
Author: Not available in collected metadata
Source published: 01/10/2026, 20:00:00 UTC
Briefing created: 05/10/2026, 05:19:15 UTC
Based on a feed excerpt; full article content has not been acquired. Source evidence retrieved 05/10/2026, 05:18:24 UTC.
Briefings
HyperBrowseComp is a benchmark introduced in October 2026 containing 423 manually authored questions across 13 languages. These questions require locating obscure evidence, following multi-step clue chains, or inspecting heterogeneous sources like videos and scanned documents. Models are evaluated using provider-native search and a shared external retrieval harness under a common agent protocol, with a human evaluation conducted on a sample subset. The difficulty stems from the need to discover and connect evidence on the open web rather than relying on parametric knowledge.
Key points
- The benchmark consists of 423 manually authored questions across 13 languages, designed to require locating obscure evidence, following multi-step clue chains, or inspecting heterogeneous sources such as videos, scanned documents, images, or maps.
- Models are evaluated using provider-native search and a shared external retrieval harness under a common agent protocol, with a human evaluation conducted on a sample of the questions to contextualize performance and effort.
Context and limitations
The benchmark covers 13 languages and 423 questions, but the evidence is untrusted and preprint status is noted. Specific model performance metrics and human evaluation results are not fully detailed in the provided text.
Generated by Bittrees-hosted qwen35-2b sha256:aaf42c8b7c3c on 05/10/2026, 05:19:15 UTC. Based on publisher-provided feed evidence; may not cover the full original article. Check the source for details.
View archived briefing on IPFS
bafybeibgz5yjubwg4m4wa42u2p5a2c73yhgrc3raclinpggz22dkkw5o2a