Blind Variation, Selective Retention: Improving Creative Idea Generation in AI Scientists
Withholding task details improved AI-generated proposals; selection informed by past results cut the candidate pool in half while keeping most top performers.
The researchers tested the framework across five proposer models and 12 scientific tasks. Limiting information during generation increased originality and non-obviousness, yielding more highly creative and high-utility proposals than leading prompting baselines. A separate selection stage used previously tested proposals and their outcomes on the same task to choose candidates for costly evaluation. Integrated into existing AI-scientist workflows, the approach produced stronger discoveries across five areas, including gene expression analysis and drug response prediction.
HF Daily Papers' note
The researchers tested the framework across five proposer models and 12 scientific tasks. Limiting information during generation increased originality and non-obviousness, yielding more highly creative and high-utility proposals than leading prompting baselines. A separate selection stage used previously tested proposals and their outcomes on the same task to choose candidates for costly evaluation. Integrated into existing AI-scientist workflows, the approach produced stronger discoveries across five areas, including gene expression analysis and drug response prediction.
HF Daily Papers' note
score 5