BVSR limits task info during AI-scientist idea generation
Original title: Blind Variation, Selective Retention: Improving Creative Idea Generation in AI Scientists
Across five proposers and 12 tasks, limited information raised originality and later selection discarded half the candidates.
The paper applies Blind Variation and Selective Retention to AI scientists so proposals are generated with deliberately limited task information, then filtered using prior verified proposals and outcomes. The authors report more original and non-obvious ideas, and a larger share of highly creative, high-utility proposals than prompting baselines. Selection discarded half of candidates while keeping most top performers. Plugged into existing workflows, they report stronger results on gene expression, drug response, anomaly detection, scaling laws and brain activity prediction.
Unconfirmed: Gains versus state-of-the-art prompting baselines are stated by the authors and not independently confirmed here.
- Exploration is intentionally under-informed; retention uses verified proposals and their outcomes.
- Outcome-based selection discarded half the candidates while retaining most top proposals.
- Reported domains include gene expression, drug response, anomaly detection, scaling laws and brain activity.
From the original article
1 min read at huggingface.co
Scientific discovery requires ideas that are original, useful, and non-obvious. Yet AI scientists are typically given as much task information as possible during idea generation, which can anchor generation toward familiar approaches.
By Guowei Xu, Jiaru Zou, Mete Erdogan · huggingface.co
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.