Evaluating Chemistry-Guided Filtering Heuristics Using LLM-Extracted Reaction SMILES

Published: 25 Mar 2026, Last Modified: 22 Apr 2026AI4X-AC 2026 PosterEveryoneRevisionsBibTeXCC BY 4.0
Submission Type: I want my submission to be considered for both oral and poster presentation.
Keywords: LLM, organic chemistry, reaction template extraction
TL;DR: Fine-tuned GPT improved USPTO reaction extraction (62.9%→92.0%). Rare-template heuristic showed lowest FNR (8%), best identifying true data quality issues over extraction errors.
Confirmation Of Submission Requirements: I submit an abstract. It uses the template provided on the submission page and is no longer than 2 pages.
PDF: pdf
Submission Number: 353
Loading