LLM Learns Uncertainty, Changing How Scientists Explore Experiments

According to AI TIMES, an EPFL research team has unveiled "Gollum," which combines an LLM with Gaussian processes, and said on September 2 that the approach cut the number of trials needed to find optimal conditions for scientific experiments by more than 40%. The system’s average top ranking across 23 benchmarks was driven by a design that makes the model handle prediction uncertainty alongside its answers, rather than simply outputting a result. The researchers said the LLM reconstructs similarities and differences among conditions on its own based on past experimental results, then refines its exploration strategy. They added that existing approaches faced high barriers to adoption because each new field required redesigning descriptors and retuning models, while this method makes it possible to use experiment protocols written in natural language directly for optimization. By contrast, when an LLM selects conditions directly without such calibration, instability becomes clear, including proposing structures that do not exist or giving answers outside the valid range. The team said the technology could leverage the strengths of language models that contain broad scientific knowledge while controlling hallucination risk within experimental optimization.