The Work
Bolton built et al. an AI-native research tool and investigated whether AI scientists can learn which experiments are worth pursuing, particularly in clinical drug development.
What Worked
Learning from historical domain data can improve performance, with several benchmarks revealing model failures that standard evaluations miss.
What Has Been Achieved
Models trained on historical clinical-development decisions beat frontier agents on a contamination-clean evaluation, with agentic scaffolding alone not helping. Across several evaluations and benchmarks, the recurring bottleneck for AI scientists was not generating plausible novel ideas but filtering and prioritising them.
What Was Learned
The limiting factor for AI scientists is filtering, not generation. Progress depends increasingly on reliable verifiers and real domain feedback: generating candidate ideas is cheap; deciding which are worth pursuing is not.
About
Will holds a PhD in AI for Healthcare from Imperial College London. He developed Steward.ai, an app to help healthcare practitioners prescribe antibiotics appropriately. Previously, he interned at GSK.ai, investigating the reliability of large language models in biomedical applications.
Lab & Advisors
Philip Torr
University of Oxford
Professor of Engineering Science; Five AI/Royal Academy of Engineering Research Chair in Computer Vision and Machine Learning
Philip Torr leads Torr Vision Group, which applies deep learning techniques to a wide ranging set of topics. He has been involved in numerous spin-outs as founder or advisor including: FiveAI, Onfido, Oxsight, Eigent, DreamTech, Visionary Machines, CamelAI, as well as working closely with big tech companies like Google, Meta, Apple, Microsoft, and Sony.
