Researchers at EPFL, Bojana Ranković, Ryan-Rhys Griffiths, and Philippe Schwaller, published a method in Nature Machine Intelligence this week that nearly doubles how often an AI system finds a high-yielding chemical reaction. The rate goes from 24% to 43% in 50 experimental iterations. The method is called GOLLuM.
The problem it addresses is specific. Pharmaceutical labs use Bayesian optimization to search for the best conditions in a reaction. Buchwald-Hartwig coupling is one of the most important reactions in drug synthesis. It links two molecules together. Getting the yield high enough to matter is slow. A chemist runs one condition, reads the result, picks the next one. Repeat. A machine can do this faster, but only if it knows when it is guessing and when it is confident.
Current LLMs fail at this. They sound confident. They hallucinate conditions that do not exist. They have no built-in sense of "I do not know." GOLLuM fixes this by training the language model using the same uncertainty signal that makes Bayesian optimization reliable. The model learns to encode experimental performance, not just text similarity. Its latent space reorganizes into regions of high and low performing conditions. The model knows where it is confident and where it is not.
The commercial angle is direct. A pharma company running a reaction optimization campaign today spends weeks on trial and error. If the AI finds the right condition in half the iterations, that is weeks of lab time saved per compound. Across a pipeline of 50 molecules, the savings compound. The method also transfers across domains. Materials synthesis. Process engineering. Molecular design. One framework, 19 benchmarks, ranked first on average.
The gap between this paper and a deployed system is still real. The experiments ran on a single H100 GPU with LoRA updates on fewer than 0.2% of model weights. That is efficient, but it is not yet running inside a lab's automation loop. The code is open source. That lowers the barrier for the next team to try it on their own reaction class.
The shift is not that LLMs can now do chemistry. It is that they can now do chemistry while knowing what they do not know. That is the difference between a suggestion and a decision.