Modern computational tools can generate millions of candidate molecules, materials, and reaction pathways in hours, creating an overwhelming combinatorial design space for experimental science. In physical laboratories, however, syntheses and validation assays remain constrained by tight budgets, human labor, and equipment time. The central bottleneck in experimental discovery is deciding exactly which candidate to test next when every failed run carries a measurable cost in time and resources.
Traditional Bayesian optimization relies on surrogate models such as Gaussian processes to balance the exploration of uncertain candidate spaces with the exploitation of known leads. Yet these classical frameworks demand handcrafted, domain-specific chemical descriptors that rarely transfer across disciplines, meaning an optimization setup tuned for a specific metal-ligand catalytic system will fail when applied to crystal structures or drug design. Large language models contain broad cross-domain scientific knowledge, but their chronic overconfidence and tendency to hallucinate make them hazardous guides for high-stakes lab routines, where an uncalibrated recommendation risks wasting expensive reagents and derailing schedules.
Gaussian Processes Calibrate Language Representations
To bridge this structural gap between probabilistic rigor and generalized scientific context, a study published in Nature Machine Intelligence introduces GOLLuM, short for Gaussian process Optimized LLMs. The framework trains language models directly through Bayesian objectives, treating the model's uncalibrated predictions not as an obstacle, but as a direct learning signal derived from experimental outcomes under uncertainty.
The framework reshapes language model embeddings so that experiments with similar outcomes cluster together, turning uncertainty into a precise learning signal.
This training dynamic reorganizes the internal representation space of the language model. By aligning semantic embeddings with measured performance, the system maps out the underlying structure of unfamiliar chemical design spaces without requiring manual re-engineering of domain-specific descriptors from scratch.
Benchmark Performance Across Design Tasks
Starting from an initial set of only ten low-performing experiments, the framework was evaluated across 23 distinct tasks covering organic synthesis, materials science, process chemistry, and molecular design. GOLLuM ranked first on average against competing baseline methods, demonstrating generalizability across diverse physical science domains.
In experimental benchmarks, the uncertainty-calibrated model matched the performance of traditional Bayesian optimization while requiring over 40% fewer experiments. When applied to Buchwald–Hartwig reactions, the architecture discovered high-performing reaction conditions at a rate of 43%, nearly doubling the 24% to 25% success rates achieved by expert-crafted quantum-chemical descriptors and standard state-of-the-art language models.
Specializing foundation models for physical science does not depend on collecting larger pretraining datasets, but on guiding them with structured uncertainty estimates from experimental feedback. While the method reduces blind trial-and-error in early-stage synthesis, its real-world lab adoption still faces the risk of hallucinations in extreme edge cases of unexplored chemical spaces, leaving the transferability of linguistic priors to highly novel reaction mechanisms an ongoing challenge for experimental pipelines.