ScientistOne: Towards Human-Level Autonomous Research via Chain-of-Evidence

Researchers introduced Chain-of-Evidence, a framework that forces claims to stay linked to supporting evidence throughout autonomous research and paper writing.

They also built ScientistOne, which preserves those evidence links from literature review to solution discovery and manuscript generation.

A companion tool, CoE Audit, checks score validity, specification compliance, reference accuracy, and method-code consistency.

Across 75 papers and five tasks, ScientistOne removed hallucinated references, passed score verification perfectly, and matched or beat human experts.

The system also generalized well to medical imaging, recognition, 3D perception, and language modeling benchmarks.