BVSR: one equation to measure AI-scientist creativity, beating baselines across fields
james_y_zou · x · 2026-11-11
Jiaru Zou, James Zou et al. propose BVSR (Blind Variation, Selective Retention), a framework for making AI scientists propose genuinely creative ideas, quantified by a single equation:
c = (1 − p) · u · (1 − v) — originality (how unlikely the idea is to be proposed) × utility (does it work) × non-obviousness (how hard its success is to predict in advance).
By this definition, an idea that works incredibly well is still uncreative if every model would propose it. Standard AI-scientist pipelines give full task context, anchoring models on familiar approaches. BVSR borrows from classic human-creativity theory:
- Blind Variation: deliberately withhold task information during ideation to induce original, less obvious proposals
- Selective Retention: experiments are costly, so a selector trained on previously verified proposals screens which new candidates are worth testing
BVSR plugs into existing AI-scientist workflows. Reported results: gene-expression analysis 0.918 vs best baseline 0.853; anomaly detection 0.679 vs 0.491; scaling-law discovery 0.967 vs 0.940, including cross-field transfers (physics→biology, economics→ML).
Related event: BVSR Framework Quantifies and Boosts AI Scientist Creativity(2 posts)→
More from AGI Musings
- New AI Futures timeline suggests intelligence explosion by 2027 is highly likely — daniel_mac8 · 2026-11-11
- Artist's 'Collaborative Thinking Model': letting agents make creative judgments mid-process — round · 2026-11-11
- New Asterisk essay asks: can evolutionary humanism survive AI? — round · 2026-11-11
- Researcher warns misaligned superhuman coder could backdoor the intelligence explosion — eli_lifland · 2026-11-11
- Alignment debate: RL and search is a terrifying way to build AGI, argues Roko — jessi_cata · 2026-11-11
- Enterprise AI ROI comes down to two variables: measurable tasks and business self-awareness — ivan_bezdomny · 2026-11-11