---

Large language models as priors for low-data chemical and materials discovery

Kevin Maik Jablonka¹; Adrian Mirza^2
1Laboratory of Organic and Macromolecular Chemistry (IOMC); Jena Center for Soft Matter (JCSM); Center for Energy and Environmental Chemistry Jena (CEEC Jena), Friedrich Schiller University Jena, Germany; Helmholtz Institute for Polymers in Energy Applications Jena (HIPOLE Jena), Germany
²Helmholtz-Zentrum Berlin für Materialien und Energie GmbH, Germany; Helmholtz Institute for Polymers in Energy Applications Jena (HIPOLE Jena), Germany

Commentary on

A.M. Bran, T.A. Neukomm, D. Armstrong, Z. Jončev, and P. Schwaller, “Chemical reasoning in LLMs unlocks strategy-aware synthesis planning and reaction mechanism elucidation”, Matter, 9(5):102812 (2026), https://doi.org/10.1016/j.matt.2026.102812
M.C. Ramos, S.S. Michtavy, A.D. White, and M.D. Porosoff, “Bayesian Optimization of Catalysis with In-Context Learning”, ACS Cent. Sci., 12(5):599 (2026), https://doi.org/10.1021/acscentsci.5c02418

Read the Commentary (PDF) Share this commentary on Bluesky

---

Statement of Significance

The chemical sciences often operate in the low-data regime, yet practitioners typically hold prior beliefs about what should be feasible or expected. Such priors have seldom been incorporated into modeling approaches in the chemical sciences. Large language models (LLMs) can offer a route to do so: recent findings indicate that they carry priors shaped by their architecture and training data that can make them more data-efficient. On top of that, they accept inputs in natural language and thus enable the use of data that could not have been used for modeling before. We discuss two recent works that leverage this prior—using LLMs to score fuzzy criteria in synthesis planning, and as in-context surrogate models for Bayesian optimization—and argue that fully delivering on this promise will require better ways to evaluate fuzzy objectives and to reproduce results.

How to cite

Cite the commentary as:

K.M. Jablonka, A. Mirza, "Large language models as priors for low-data chemical and materials discovery",
KIM REVIEW, Volume 4, Article 04, 2026. DOI: 10.25950/6c2e5798