On September 10, 2026, researchers posted “Autonomous Chemical Mechanistic Discovery through Agentic Reasoning and Validation” (arXiv:2609.11147), introducing ARCHE, a closed-loop system that pairs a general-purpose reasoning model with a domain-specialized computational chemistry model and a structured registry of computational tools to investigate reaction mechanisms.
ARCHE works by interpreting a mechanistic question, generating and prioritizing competing hypotheses, coordinating the computational chemistry workflows needed to test them - such as transition-state searches and energetics calculations - and iteratively refining its conclusions against the calculated evidence rather than stopping at the first plausible answer.
The authors validated the system on three cases of increasing difficulty: reconstructing the stereocontrolling transition states of an already-published asymmetric catalytic reaction, proposing and validating a plausible radical mechanism for a recently discovered but still-unpublished alpha-iodoboronate C-I cleavage reaction, and identifying a chemically meaningful descriptor that governs selectivity in nickel-catalyzed migratory cross-coupling reactions.
The unpublished-reaction case is the notable one, because it means ARCHE was tested on a mechanism no training corpus could have memorized, rather than reproducing textbook chemistry. This is a research-lab validation on a small number of hand-picked cases, not a large benchmark, so the real test of the approach is whether independent chemistry groups can reproduce its hypotheses on reactions ARCHE has never encountered.