AI Doubt Detector Cuts Experimental Tests by 40%

Researchers at the École Polytechnique Fédérale de Lausanne have developed a novel artificial intelligence framework that significantly reduces the time and cost required for experimental optimization in scientific research. The system, named GOLLuM, merges large language models with Gaussian processes to function as a probabilistic doubt detector for Bayesian optimization. The methodology, detailed in a recent publication in Nature Machine Intelligence, was authored by doctoral researcher Bojana Ranković and laboratory head Philippe Schwaller at the EPFL Laboratory of Artificial Chemical Intelligence. Traditional Bayesian optimization relies on expert-designed descriptors to predict promising experimental conditions, but this approach lacks portability across different scientific disciplines. Conversely, while large language models possess broad scientific knowledge and operate effectively on natural language, they are prone to hallucinations and lack reliable self-assessment of uncertainty. GOLLuM addresses these limitations by training the language model on the probabilistic uncertainty scores generated by a Gaussian process. As the system processes experimental feedback, it dynamically reorganizes its internal representation of the design space, clustering similar outcomes and separating distinct results. This continuous refinement allows the model to prioritize high-value experiments while discarding unproductive paths. The framework was evaluated across twenty-three benchmark tasks spanning organic synthesis, materials science, catalysis, and molecular property optimization. Operating with a standardized configuration rather than problem-specific tuning, GOLLuM achieved the highest average performance among tested methods. Within a strict budget of fifty experiments, over thirty-six percent of its selected conditions ranked in the top five percent of all possible outcomes, outperforming traditional descriptor-based optimization. Most notably, the system converged to optimal solutions using more than forty percent fewer experiments than conventional approaches. Independent testing of unmodified large language models revealed severe reliability issues, with failure rates between ten and eighty percent due to fabricated structures and out-of-bounds suggestions. By decoupling experimental proposal generation from direct prompt-based questioning, GOLLuM establishes a new operational paradigm for laboratory automation. The architecture enables researchers to initiate optimization campaigns immediately using plain English procedural descriptions, bypassing months of traditional descriptor engineering. This integration of probabilistic uncertainty quantification with large language models demonstrates a scalable pathway toward more efficient, interpretable, and autonomous scientific discovery.