Verification-first autonomous catalysis: large language models as infrastructure for mechanism, computation, and experiment

Abstract Large language models are shifting catalysis from prediction-first pipelines to workflow-level coordination of hypotheses, computation, and experiments. In this Perspective, we argue that autonomous catalysis requires closing both representation and workflow gaps through a verification-first paradigm, in which large language model agents orchestrate heterogeneous tools, ground claims in auditable evidence, and use uncertainty-aware validation. We further propose system-level metrics to benchmark reliable autonomous catalysis.