ReaXplore: A Physics-Grounded Agent for Chemical Reaction-Network Exploration

Reaction-network exploration is a sequential decision problem under expensive scientific feedback: which reaction hypothesis to test, which computational method to use, which intermediate to expand, and when the evidence is sufficient. We introduce an LLM-driven agent that makes these decisions under an explicit compute budget while operating on a persistent, physics-grounded reaction network. The language model is the adaptive planning and compute-allocation policy; the reaction network is the agent's persistent external state and scientific memory; a physics verifier converts calculations into trusted feedback (a step enters the network only if its transition state shows one dominant reaction-coordinate imaginary mode on the changing bonds and its intrinsic reaction coordinate connects two real endpoints); and declarative recipe and template stores provide auditable long-term adaptation across runs. We show how an LLM agent can adaptively explore chemical reaction networks while remaining grounded in a persistent, provenance-carrying scientific state whose contents are determined by computation rather than model assertion. The evidence spans a thirty-reaction benchmark on a machine-learned potential (28 admitted at the strictest intended-product level; both failures are a diagnosed continuum-solvation artifact), held-out chemical questions with predictions frozen before running, a multistep budgeted exploration, logged replanning events, one transferred recipe update, and a pilot evaluation in which network records lift a frontier model's record-grounded network-question accuracy from 68% to 97%.