R-GroundBench introduces a diagnostic benchmark for pharmaceutical Markush structures using real patent data to evaluate molecular, textual, and chemical grounding. Current models show a significant performance drop from 90% accuracy on easy VQA tasks to 56–66% on harder grounding tasks.
HOW THIS AFFECTS YOU
●
researcherYou can use this to measure how well your models handle variable R-group placeholders in chemical representations.
●
healthThis provides a way to benchmark AI performance on critical pharmaceutical patent data structures.