Large Language Model-Guided Scientific Discovery
Overview
The field studies how large language models can help turn scientific literature and data into hypotheses, mechanisms, algorithms, and new materials, making discovery more systematic and widening the range of ideas scientists can test. Common approaches combine literature retrieval or knowledge graphs with tool-using agents, symbolic regression, evolutionary search, Bayesian optimization, and feedback from simulations or experiments. Work spans research-idea generation, equation and mechanism discovery, experimental planning, and materials or molecular design, with evaluation ranging from scientific benchmarks and formal verification to closed-loop laboratory tests. A central challenge is ensuring that promising outputs are novel, physically or biologically grounded, reproducible, and generalizable: strong predictive fit or fluent explanations alone do not establish a valid discovery, so independent checks and human oversight remain important.
Latest papers
278 papers49 papers in the last four weeks, up 206% on the four weeks before. 0.3% of all new papers.