The authors constructed 200 tasks across eight domains, including science, history, culture and technology.
Web research benchmark exposes persistent limits in autonomous search agents
Across 200 tasks requiring evidence from multiple live Web sources, the strongest of 12 tested models answered 39.0% correctly.
Big Tech
Xiaomin Yu · Hongliang Chen · Zhe Huang · Shilin Yan · Shao-Rong Su · Xuqiao Ma · +19 more
HKUST(GZ) · FDU · UW · IC · NTU
Research Digest··2 min read
Yu and colleagues introduce OpenDS-World, a benchmark testing whether AI agents can follow complex evidence trails across the live Web rather than retrieve answers from a fixed collection.
§