Research

'AI scientist' systems move from generating papers to verifying them

Automated research pipelines now propose hypotheses, design experiments and — crucially — flag which of their own results fail replication checks.

By Priya Sharma, Research Editor — CAMBRIDGE

CAMBRIDGE — The first generation of automated research systems flooded preprint servers with plausible papers. The second, researchers say, is learning to do something more valuable: telling you which of its own findings not to trust.

New pipelines described at this week's machine-learning conference pair hypothesis generation with autonomous replication — re-running their own experiments under perturbed conditions and attaching robustness scores before any human reads the result.

In materials science and drug discovery, where candidate spaces are vast and wet-lab time is precious, early adopters report the self-filtering systems cut failed experimental follow-ups by more than half.

Enable JavaScript to read the full story on Neural Daily News.