Researchers have introduced OpenDS-World, a benchmark that tests whether AI agents can research answers across the live web, using search, browser interactions and varied file formats. The benchmark contains 200 tasks, with human checks to confirm the evidence and that each task has a unique answer.
In the researchers’ evaluation of 12 models under one tool setup, the best reached 39.0% accuracy. Human evaluators got 21.1% across 1,000 attempts. The paper also reports persistent problems with coordinating tools and managing long searches; it does not identify the best model in the supplied abstract.
Why it matters
For researchers building web-research agents, OpenDS-World offers a live-web test; its results suggest these tasks remain difficult, but the abstract does not show how the scores translate to everyday use.
