TL;DR: A language model learned to doubt itself and found better experiments faster

A peer-reviewed method published 28 August 2026 in Nature Machine Intelligence trains a language model through a Bayesian objective so it can help choose the next experiment, and the study reports first-place average ranking across 23 tasks after starting from only ten low-performing experiments.

The work comes from Bojana Ranković, Ryan-Rhys Griffiths and Philippe Schwaller, and it names the approach GOLLuM, short for Gaussian process Optimized LLMs, according to the Nature Machine Intelligence article titled Large language models as uncertainty-calibrated optimizers for experimental discovery. The central result is sample efficiency: the paper reports matching traditional Bayesian optimization with over 40% fewer experiments. The sharpest chemistry comparison concerns Buchwald–Hartwig reactions, where the reported discovery rate for high-performing reactions was 43% for GOLLuM versus 24–25% for expert quantum-chemical descriptors and state-of-the-art language models. The broader claim is that richer, uncertainty-guided training can specialize a foundation model without requiring more data.

A published result with a date, authors and open record

The study appeared online on 28 August 2026 in Nature Machine Intelligence, volume 8, issue 9, pages 1466–1477, according to the Nature Machine Intelligence article titled Large language models as uncertainty-calibrated optimizers for experimental discovery. The listed authors are Bojana Ranković, Ryan-Rhys Griffiths and Philippe Schwaller. Institutional information in the article record associates Ranković and Schwaller with the Institute of Chemical Sciences and Engineering at École Polytechnique Fédérale de Lausanne and NCCR Catalysis, and associates Griffiths with QuantumGreen in San Francisco.

The article history records receipt on 13 February 2026, acceptance on 3 July 2026 and first online publication on 28 August 2026. The paper is open access, with text-and-data-mining and version-of-record permissions valid from 28 August 2026. The Crossmark record lists funding from the Schweizerischer Nationalfonds zur Förderung der Wissenschaftlichen Forschung under grant 225147 and states that the authors declared no competing interests. A preprint version is available at arXiv, and code and data are archived through Zenodo under the GOLLuM title.

That publication trail matters because experimental optimization claims are easy to overstate when methods, baselines and task counts are vague. Here the task count, starting budget, efficiency comparison and chemistry benchmark are stated explicitly in the abstract and main text. Independent coverage in the ALLATRA Media report titled Teaching an AI to Doubt Itself Cut the Chemistry Experiments Needed, Study Reports confirms the same publication date, author list, method name and headline numbers. The PapersCode explainer titled GOLLuM LLM Framework Accelerates Scientific Discovery likewise reports the 23-task evaluation, the over-40% experiment saving and the 43% versus 24–25% Buchwald–Hartwig comparison.

The problem: which expensive candidate to test next

Experimental discovery, from reaction optimization to molecular design, poses the same costly question described in the paper: which candidate should be tested next under time and resource constraints. Artificial intelligence can propose millions of drug candidates, material structures or chemical reactions in hours, but laboratories cannot test them all. The bottleneck is selection under scarcity.

Bayesian optimization is the established statistical answer to that bottleneck. It uses a surrogate model, typically a Gaussian process, to predict experimental outcomes together with uncertainty, then balances exploration of uncertain options against exploitation of promising ones. That uncertainty estimate is central because it guides whether to confirm a likely winner or probe an unfamiliar region that could contain a better result.

The limitation, as described in the Nature Machine Intelligence article titled Large language models as uncertainty-calibrated optimizers for experimental discovery, is representation. Bayesian optimization often depends on domain-specific features. Optimizing a chemical reaction may require translating structures, catalytic mechanisms and reaction pathways into descriptors that encode chemical intuition. Those descriptors rarely transfer. Reaction fingerprints do not characterize crystal structures, catalyst optimization looks different from drug design, and descriptors tuned for one metal-ligand system may not move effectively to another system. Each new domain can therefore start nearly from scratch.

The other half of the problem: knowledgeable but overconfident models

Large language models bring complementary strengths and a well-documented weakness. Their strength is generality through prior knowledge absorbed from broad scientific and technical text. Their weakness is calibration. They can hallucinate, producing fluent outputs that sound confident while being factually incorrect.

In everyday use that weakness causes frustration. In experimental science the paper warns that the consequences are more concrete: hallucinated experimental suggestions waste resources, delay timelines and compromise safety protocols in automated laboratory settings. For that reason, the authors describe direct use of language models as experimental optimizers as impractical without rigorous control.

Prior efforts to combine Bayesian optimization with language models gained traction, but the paper characterizes most existing approaches as treating the two parts separately. A language model might propose candidates while a separate optimizer scores them, or embeddings might be used without changing how the language model represents uncertainty. GOLLuM instead trains the language model through the same probabilistic objective that makes Bayesian optimization reliable.

What GOLLuM does differently

GOLLuM stands for Gaussian process Optimized LLMs. It pairs a language model with a Gaussian process, a statistical model that returns a prediction together with a confidence measure, and trains the two jointly rather than separately, according to the ALLATRA Media report titled Teaching an AI to Doubt Itself Cut the Chemistry Experiments Needed, Study Reports. The joint training uses the Gaussian process marginal likelihood. Gradients from that objective teach the language model to encode experimental outcomes rather than textual similarity.

The practical effect is a reshaped internal representation. Experiments with similar outcomes cluster together, producing distinct regions of high and low performance in the model’s latent space. The paper describes that transformation as turning overconfidence from a fundamental flaw into a precise learning signal. In other words, the penalty for confident but wrong predictions is not merely a warning; it is the training gradient that reorganizes the embeddings.

Guidance comes from natural language. Experimental procedures, reaction descriptions and molecular representations can be supplied as text, while the Gaussian process supplies calibrated uncertainty over the learned representation. That combination is intended to preserve the generality of language input while restoring the principled decision-making associated with Bayesian optimization.

The comparison reported in the ALLATRA Media article notes that all language-model methods in the evaluation used Google’s T5 model as the encoder. That detail clarifies that the reported advantage did not come from swapping in a larger proprietary backbone for one method while leaving competitors on a smaller encoder. The difference lay in how the shared type of encoder was trained.

The 23-task test and the ten-experiment start

The evaluation covers 23 tasks across organic synthesis, materials science, process chemistry and molecular design, according to the Nature Machine Intelligence article titled Large language models as uncertainty-calibrated optimizers for experimental discovery. That breadth is important because transfer across heterogeneous design spaces is the central claim. A method that only accelerates one reaction family would be useful, but it would not solve the representation-transfer problem described above.

The starting condition is deliberately harsh. GOLLuM begins from only ten low-performing experiments. Ten points is a small budget for any Gaussian-process-style surrogate, which typically needs enough observations to estimate both the outcome surface and its own uncertainty. Starting low also prevents the method from coasting on lucky initial hits. It must learn enough from failures to propose better candidates.

Under those conditions, the paper reports that GOLLuM ranked first on average among all competing methods. It also describes the result as state-of-the-art performance. The PapersCode explainer titled GOLLuM LLM Framework Accelerates Scientific Discovery repeats the same ranking claim and emphasizes descriptor-free transfer as the load-bearing result because the domains are heterogeneous. Matching traditional Bayesian optimization with over 40% fewer experiments is the headline efficiency figure shared across the primary paper and both secondary summaries.

Buchwald–Hartwig reactions: where the gain becomes concrete

Buchwald–Hartwig cross-coupling reactions provide the most detailed chemistry benchmark in the reported results. These are a widely studied class of carbon–nitrogen bond-forming reactions, and predicting reaction performance in that family has been a machine-learning benchmark since earlier work on yield prediction. The new study reports that GOLLuM nearly doubled the discovery of high-performing Buchwald–Hartwig reactions relative to expert quantum-chemical descriptors and state-of-the-art language models.

The numbers are 43% for GOLLuM versus 24–25% for the baselines. That gap is large enough to matter in a laboratory where each high-yield hit saves downstream purification, characterization and scale-up effort. The comparison also matters methodologically because expert quantum-chemical descriptors represent the domain-knowledge approach: carefully constructed features that encode electronic and steric effects. Beating that baseline without hand-built features supports the claim that the learned representation captures chemically relevant structure.

Interpretability evidence comes from the transformed representations. In Buchwald–Hartwig reactions, the paper reports that the learned space automatically clustered iodide-based reagents in high-yield regions while separating low-performing chlorides. That separation aligns with genuine chemical patterns in cross-coupling reactivity rather than arbitrary textual similarity between procedure descriptions. The model was not told to group halides by leaving-group ability; the outcome-driven training organized them that way.

Beyond one reaction family: solvents and materials

The paper extends the interpretability claim beyond organic synthesis. In analytic and process chemistry, the authors report that the model identifies optimal ether–alcohol solvent systems. In materials science, the transformed representations expose structure relevant to materials design tasks included in the 23-task suite. Those examples are presented as evidence that the same backbone, without domain-specific featurization, transfers across reaction optimization, materials science and molecular design.

That transfer claim is the point most worth reproducing, as noted in the PapersCode explainer titled GOLLuM LLM Framework Accelerates Scientific Discovery. Sample-efficiency gains on familiar tasks can sometimes mask transfer costs that appear only on held-out benchmarks. The reported evaluation already spans four areas, but laboratories considering adoption would still want to see the ten-experiment warm-start protocol repeated on their own non-chemistry tasks and to inspect whether predictive distributions are genuinely calibrated or whether ranking quality is driven by a softer proxy.

The broader framing in the paper is that specializing foundation models need not mean adding more data. Instead, GOLLuM points to richer, uncertainty-guided information as the specialization lever. That framing resonates with adjacent work on information-theoretic objectives, including a Fisher-information-guided continual-adaptation framework discussed alongside GOLLuM in recent coverage, although that adjacent result is separate from the 23-task chemistry evaluation.

Why calibration changes how labs can use language models

Calibration is the difference between a suggestion engine and a decision engine. A suggestion engine can list plausible reactions, solvents or molecules. A decision engine must also say how uncertain it is, so a laboratory can decide whether the next slot on an instrument should go to a safe bet or a risky probe. Gaussian processes provide that uncertainty natively. Standard language models do not.

By training the language model jointly with the Gaussian process, GOLLuM makes uncertainty part of representation learning rather than an afterthought bolted onto frozen embeddings. The marginal-likelihood objective penalizes both poor predictions and misplaced confidence, which reshapes the embedding geometry toward outcome similarity. The reported consequence is that natural-language descriptions become usable inputs for principled optimization without first converting every new domain into bespoke descriptors.

Safety and cost implications follow directly from the efficiency numbers. Over 40% fewer experiments to match traditional Bayesian optimization means fewer reagents consumed, fewer instrument hours booked and fewer failed runs to dispose of or troubleshoot. The nearly doubled Buchwald–Hartwig hit rate means a higher fraction of attempted reactions clears the high-performance threshold. Neither figure eliminates the need for human oversight or laboratory safety protocols, but both reduce the number of times those safeguards are stressed by low-value trials.

Limits, checks and what remains to reproduce

The reported results are retrospective benchmarks and prospective selections within defined task suites, not proof that the same savings will appear in every laboratory. The 23-task average ranking summarizes performance across the included tasks; it does not guarantee first place on any single new task. The ten-experiment start demonstrates cold-start ability in the studied settings, but ten points remains an aggressive budget that could behave differently under noisier biological measurements or strongly constrained design spaces.

Three verification steps follow directly from the reported design. First, compute the same top-performing discovery metric on a held-out chemistry benchmark outside the 23-task suite. Second, inspect the Bayesian training objective to confirm that predictive distributions are calibrated rather than merely well-ranked. Third, rerun the ten-experiment warm-start protocol on a non-chemistry task to stress-test cold-start generalization. Those checks preserve the paper’s action status: an announced and evaluated method, not yet an established laboratory standard.

For United States science and technology readers, the near-term significance is architectural. If uncertainty-guided training reliably converts generalist language priors into sample-efficient optimizers, the same pattern could apply to self-driving laboratories, materials exploration and molecular optimization programs that already combine generative proposals with experimental feedback. The study’s publication on 28 August 2026, its open-access availability, its arXiv preprint and its Zenodo archive make the method inspectable. The next step is independent reproduction on new benchmarks before adoption.

Sources referenced