A research agent finds several useful candidates. Another learns to rank unseen candidates accurately. Those achievements can diverge, even in the same experimental campaign.
Autonomous Science Bench, introduced October 13, evaluates budgeted campaigns grounded in published experiments. Success requires both a discovery gate for measured designs and a learning gate for ranking held-out designs. Its pilot reveals different weaknesses across agent configurations. Read the research paper.
Good search and good understanding need separate tests
An agent can exploit a promising region without building a transferable predictive model. It can also learn a useful model while spending its limited experiments too cautiously to discover enough strong candidates.
The benchmark's initial six campaigns and pilot trials define a limited scope. A lab-service runtime grounded in published measurements is not the same as a fully autonomous wet lab creating new experimental outcomes.
Compare discovery yield, predictive transfer and budget consumption independently. Those metrics reveal whether improvement should target experiment selection or inference from results.
The distinction gives autonomous-science claims a firmer foundation. A successful campaign should establish what was found and what was learned. Treating one as proof of the other obscures which part of the scientific process the agent actually performed well.
References
Questions, news tips or corrections? Contact [email protected].
