ChemEAGLE pairs AI judgment with program checks to read reaction graphics into databases
A multi-agent system from a new arXiv paper converts reaction drawings into machine-readable records, reporting higher accuracy than a tool-equipped generalist agent on published figures.
In plain words
Chemists publish reactions as pictures, and computers often misread them. ChemEAGLE combines software tools that supply chemical structures with a language model that makes judgments the programs then apply. The authors report it built a large photochemistry database from published figures with fewer errors than a generalist agent.
What was announced
The paper's authors argue that AI-driven chemical research depends on high-quality databases, yet many reactions exist only as graphics drawn for human readers. They describe specialist software as precise but rigid, and multimodal language models as flexible but prone to writing incorrect structures, with neither approach flagging its own errors. ChemEAGLE is their multi-agent answer: tools supply structures, while a language model extends their reach through judgments that programs apply.
On 424 published graphics, the authors report an F1 score of 82.2% for exactly correct reactions, compared with 47.7% for a tool-equipped generalist agent. Calibrated confidence limited human review to 35% of reactions at 96.6% accuracy. Without review, the system built a 60,103-reaction photochemistry database whose records trained models to rank reaction conditions and predict products. The figures come from the paper's abstract alone, so the full article may add conditions, caveats or corrections not reflected here.
What it does not show
- Evidence is limited to the paper's abstract; the full article's methods, caveats and any corrections are not covered here.
- The accuracy figures are the authors' own reported results; no independent confirmation is available in the supplied sources.
- The comparison baseline is a single tool-equipped generalist agent described in the abstract; other systems were not evaluated in the supplied evidence.
Sources
Source-support review completed 19 Nov 2026. This story was drafted with AI assistance and checked against retained source passages. Automated checking is not independent verification.
Coverage: arXiv: abstract.
Updates and corrections
No corrections so far.
First published 19 Nov 2026
New to the subject? AI from the Ground Up starts from the beginning.