A laboratory choosing its next experiment has to weigh predicted performance against uncertainty and a limited budget. GOLLuM combines a language model with a Gaussian-process objective so that natural-language descriptions become a representation for that choice. Its evaluation tests whether the combined system finds strong candidates with fewer suggestions on existing experimental datasets.[1][2]
The result is a hybrid system rather than a chatbot acting as a scientist. Previous experimental outcomes train the representation, the probabilistic model estimates performance and uncertainty, and an acquisition rule selects the next candidate. The system’s value is measured by how quickly it finds high-performing experiments—not by how scientific its explanation sounds.
The reported GOLLuM evaluation
| Measure | Value | Context |
|---|---|---|
| Task portfolio | 23 optimisation tasks | Organic synthesis, process chemistry, materials and catalysis, and molecular properties.Source 1 |
| Starting information | Ten low-performing experiments | Each campaign began below the median to test recovery from a weak initial set.Source 1 |
| Campaign budget | Fifty experimental suggestions | Results were averaged across twenty random seeds per task.Source 1 |
| Top-five-percent coverage | 36.3% | Compared with 26.5% for BoChemian, 12.1% for LAPEFT and 29.7% for a tailored Gaussian-process baseline.Source 1 |
| Sample efficiency | Median 41% fewer iterations | GOLLuM matched the final performance of traditional Bayesian optimisation with fewer experiments.Source 1 |
| Buchwald–Hartwig result | 43% high-performing conditions found | Compared with 24–25% for expert quantum descriptors and other tested LLM representations.Source 1 |
| Largest domain gain | 90% relative improvement in process chemistry | Reported relative to traditional Bayesian optimisation in that domain.Source 1 |
Uncertainty becomes a training signal
Standard language-model embeddings group text by semantic resemblance. Experimental optimisation needs a different geometry: candidates with similar measured outcomes should become close even when their descriptions use different words. GOLLuM jointly trains the language representation through the Gaussian process’s marginal likelihood, reshaping the space around what performed well or badly.[1]
The mechanism supplies a prediction and an uncertainty estimate to an acquisition rule, which balances promising candidates against less explored regions. That does not guarantee that every suggestion adds useful information or that uncertainty remains calibrated on a new scientific domain. Those properties need evaluation on the intended data and design space.
The operating loop for an uncertainty-calibrated experiment agent
| Stage | Machine contribution | Human or laboratory control |
|---|---|---|
| Define | Encode candidates and the measurable objective in natural language | Confirm that the objective represents scientific and operational value |
| Initialize | Learn from the first ten observed experiments | Verify measurements, units, failed runs and missing valuesSource 1 |
| Predict | Estimate outcome and uncertainty across untested candidates | Check constraints, safety and whether candidates can actually be run |
| Select | Use an acquisition rule to choose the next informative experiment | Approve the batch, cost and exploration tolerance |
| Measure | Ingest the observed result and update the representation | Preserve raw data, deviations and experimental conditions |
| Stop | Assess whether further experiments justify their cost | Decide whether the optimum is sufficient, robust and scientifically understood |
Direct prompting failed for operational reasons
The paper reports that direct language-model suggestions could be invalid, duplicated, unparsable or delayed by timeouts. Those failures matter even when the underlying scientific idea is sensible. A campaign must return a candidate in the permitted schema, avoid repeats and respect the design space before the laboratory can evaluate its merit.[1]
A robust implementation should therefore keep deterministic constraints outside the model. Candidate ranges, forbidden combinations, inventory and duplicate detection belong in code. The model contributes a transferable representation and prior knowledge; the optimisation layer and validators govern what can be selected.
Fewer experiments can be worth more than a better final score
In wet-lab work, each measurement can consume reagents, machine time and researcher attention. Matching a baseline’s final result with 41% fewer iterations can therefore be more valuable than a small gain achieved at the same or larger budget. The saving remains benchmark-specific until reproduced in a real prospective campaign with actual delays, failures and costs.
- Choose an objective that can be measured consistently across the campaign.
- Encode hard safety and feasibility rules outside the language model.
- Start with a deliberately documented initial set, including unsuccessful experiments.
- Track prediction, uncertainty, acquisition score and reason for every selection.
- Count invalid, duplicate and failed proposals as operational outcomes.
- Run a prospective comparison against the laboratory’s existing selection process.
GOLLuM demonstrates a credible way to make foundation-model knowledge useful under experimental scarcity. The language model supplies a general representation; the Gaussian process supplies calibrated uncertainty; the laboratory supplies the ground truth. That combination can reduce the number of costly trials while keeping the decision rule inspectable.
Sources & further reading
- Large language models as uncertainty-calibrated optimizers for experimental discovery Nature Machine Intelligence · 28 Aug 2026 · Accessed 11 Oct 2026
- GOLLuM: Gaussian process optimized language models for experimental discovery Ranković, Griffiths and Schwaller / arXiv · 8 Apr 2025 · Accessed 11 Oct 2026