Language models rival trained baselines on molecular properties from a few examples

With 30 in-context examples, the most accurate of seven language models beat all eight classical baselines on three uncontaminated datasets.

Research Lab

Matthias Busch · Marius Tacke · Tamadur Albaraghtheh · Sviatlana V. Lamaka · Mikhail L. Zheludkevich · Roland C. Aydin · +2 more

Hamburg University of Technology · Helmholtz-Zentrum Hereon · German Research Center for Artificial Intelligence (DFKI) · Saarland University

Research Digest·2 November 2026·3 min read

Illustration: Language models rival trained baselines on molecular properties from a few examples

In one line

With thirty examples, the best LLM beats classical baselines on molecular property prediction, but its self-reported uncertainty ranks errors only as well as baselines.

Before you read this

An idea this paper builds on

large language modelA neural network trained on massive text corpora to generate coherent text and perform diverse language tasks.244 papers →

Read it as

Predicting molecular behavior is key in drug discovery, but examples are scarce. (LLMs) read a molecule's structure as text and predict a property, also stating their confidence. Can that confidence be trusted?

The study tested several LLMs on multiple datasets, giving each a small set of examples (up to thirty) and asking for predictions with uncertainties. These were compared to many classical models. The researchers ensured the data was not memorized.

On clean data, the best LLM beat every classical model using just that many examples. But its stated uncertainty ranked errors only moderately, with a correlation of 0.46, similar to classical uncertainties. Selecting the most confident half helped rarely, so self-reported uncertainty is of limited value.

How it works

before and after

RSS

The Daily Edition

The front page in your inbox each weekday morning: the lead, the picks, and why each one earned its place.

Concepts

in-context learning self-reported uncertainty large language model