MechParser: A Vision-Language Framework for Parsing Chemical Reaction Mechanism Diagrams

Reaction mechanisms are central to understanding chemical reactivity, guiding reaction prediction, synthesis design, and selectivity rationalization. In chemistry literature, they are universally depicted as curved electron-pushing arrows that capture stepwise electron flow and bond changes. Yet, this vast knowledge remains trapped in unstructured diagrams, forcing data-driven modeling to rely on small or template-expanded datasets. We introduce MechParser, a vision-language framework that converts mechanism diagrams into SM-SMARTS (Sequential Mechanism-Aware SMARTS), a representation encoding electron flow and atom state changes as a cumulative string sequence. At its core, MechParser-VL applies a left-right visual prompt design (arrow diagram alongside an atom-indexed reference) to separate dynamic arrow reasoning from static atom identification, and is trained via a four-stage curriculum on over 300K samples from a geometry-aware synthetic data pipeline and an expert-annotated real-world dataset. On both synthetic and real-world benchmarks, MechParser-VL substantially outperforms much larger proprietary and open-source VLMs despite using only a 4B backbone. By standardizing visual mechanisms into machine-readable data, MechParser lays the groundwork for constructing large-scale mechanism databases from the chemical literature, which in turn can support the training of advanced mechanism-centric models on authentic mechanistic data.