An automated pipeline of large language models has taught itself to classify chemical reactions — and write the classification rules — across nearly 666,000 US patent reactions.
Researchers built a multi-agent LLM framework that both labels reactions and generates the symbolic rules governing those labels, each rule verified against the full corpus before it sticks. Starting from a standard taxonomy of 68 reaction classes, the system expanded the ruleset to 14,073 classes with no human curation. A lightweight fingerprint classifier built on top of those rules then correctly classified 97.7% of reactions it had never seen before — matching a leading proprietary classifier while resolving chemistry at finer grain and extending to reaction types outside its training distribution.
Synthesis planning software has long depended on fixed, hand-coded rulesets that break down at the edges of known chemistry. That constraint mattered because chemistry is long-tailed: rare reaction types are numerous in aggregate, and static libraries miss them. A system that writes and verifies its own rules on demand sidesteps that bottleneck, and the proprietary-parity accuracy figure suggests it does so without sacrificing reliability.
The real test is whether the pipeline holds up outside patent literature, where reaction reporting is unusually structured — messier lab notebooks and preprints will be a harder target.