OpenDS Benchmark Reveals Top Models Hit Only 39% Accuracy in Real-Web Search

1 reports archived1 independent sourcesupdated 10/30/2026, 18:33:00

Synthesis & Latest Updates

1 Sources Cross-Validated

Researchers from HKUST(GZ) and other institutions introduced the OpenDS paradigm and OpenDS-World benchmark to evaluate end-to-end deep search capabilities of AI agents in real-world open web environments. Results show that even the strongest models achieve only 39% accuracy when facing multimodal, heterogeneous resources and interactive access requirements, highlighting significant gaps in autonomous planning and complex information integration.

LATEST/Researchers from HKUST(GZ) and other institutions introduced the OpenDS paradigm and OpenDS-World benchmark to evaluate end-to-end deep search capabilities of AI agents in real-world open web environments. Results show that even the strongest models achieve only 39% accuracy when facing multimodal, heterogeneous resources and interactive access requirements, highlighting significant gaps in autonomous planning and complex information integration.

Hourly heat curve

32 fully observed hours

Now

5.9

Peak

11.919:00

24h change

-43%

The line compares only sources observed throughout; gap hours are interpolated to keep the trend continuous.

Coverage timeline

Total 1 reports · Latest first

机器之心 (微信公众号)T2Oct 30, 18:33·78 pts
OpenDS Benchmark Reveals Top Models Hit Only 39% Accuracy in Real-Web Search
- Introduces the OpenDS paradigm focusing on complex search scenarios in real open webs without pre-given evidence locations or access methods.
- Constructs the OpenDS-World benchmark covering multimodal heterogeneous resources like video frames, podcasts, paper charts, and downloadable files.
- Empirical data shows top models achieve only 39% accuracy, indicating insufficient cross-source reasoning and interactive page operation capabilities in current agents.
View Related Reports & Analysis Original Source