The paper presents a multi-agent LLM pipeline that classifies reactions and writes reaction rules across 665,901 reactions extracted from US patents. Each generated rule is tested against the corpus in a verification loop. The system expands a standard taxonomy from 68 to 14,073 classes without human curation. A lightweight fingerprint classifier reportedly classifies 97.7% of unseen reactions, matching a leading proprietary classifier while offering finer-grained chemical resolution. The authors also describe on-demand extension to chemistry outside the original training distribution, framing the result as a living, self-expanding reactivity database.
No heat snapshots are available in the last 24 hours.