RxnExBench Multimodal Benchmark for Chemical Reaction Extraction

RxnExBench is a multimodal benchmark containing 3,565 expert-annotated reactions and 2,361 molecular structures for evaluating vision-language models in chemistry. It utilizes a 15-role, 26-field schema to test structured data extraction from scientific literature figures. Evaluations of nine state-of-the-art VLMs indicate significant performance degradation as reasoning complexity increases.