LLM literature priors boost Bayesian optimization: reaction yields jump from 6% to 39%
bravo_abad · x · 2026-10-22
McCoy and colleagues used an untuned Gemini to read chemistry papers and rank choices like solvents and catalysts, feeding those rankings as priors into Bayesian optimization for reaction-condition search.
Key design: the LLM's suggestions get extra weight while the optimizer stays free to explore other conditions and learn from results — mimicking how a chemist leans on prior literature.
The literature mattered: on one reaction dataset, LLM-suggested conditions had an average recorded yield of 6% without supplied papers vs 39% with ten relevant papers. Retrospective box plots across Buchwald-Hartwig, Suzuki-Miyaura and Direct Arylation show paper-informed LLM priors roughly matching or beating plain BO, while misleading literature significantly slows convergence — prior quality is decisive.
More from Research
- Training-free 20B model Viveki generates real-time video heatmaps under 150ms — not-ekalabya · 2026-10-22
- IBM's LASAGNA study shows RAG technique gains shrink when combined: +47% predicted becomes +17% realized — ibm · 2026-10-22
- Dev hot take: mechanistic interpretability is 'astrology for AI people' — dejavucoder · 2026-10-22
- eLife 'Landmark' paper: RNNs reveal a mechanistic theory of planning in prefrontal cortex — ShahabBakht · 2026-10-22
- Mathematician Kontorovich: 'All code needs to be Lean formalized like yesterday!' — ludwigABAP · 2026-10-22
- Cohere's CulturalRiddles: 28 LLMs score 79.3% on cross-cultural riddles, but just 22% in some cultures — Cohere_Labs · 2026-10-22