Abstract
AI-Scientists have arrived, but a world for them to grow in did not. Human scientists take theirs for granted: centuries of open problems and shared ideas, waiting in the literature. To an AI-Scientist that world is readable but not runnable, so we build Scientist World, where every paper becomes an executable environment. As a benchmark, frontier AI-Scientists already beat the paper's own method 76% of the time. As a dataset, it holds 6,400 hours of research traces: every idea, experiment, failure, and success. As a gym, it raises a small model +158% to match the frontier. At last, their world has arrived, grown from and growing with science itself.
What research is
It repeats, problem after problem — so an AI-Scientist needs a world big enough to grow in.
The human scientist's world
For a human, the literature is that world: every open problem, and the ideas of everyone who tried before.
How the world is built
Every paper, rebuilt into an environment an agent can actually run.
Where it stands
Older worlds reach only two of three: scale, staying current, or real results. Scientist World reaches the corner where all three meet.