ReactBench Benchmark Exposes MLLM Limitations in Topological Reasoning on Chemical Diagrams

A new benchmark called ReactBench reveals significant shortcomings in Multimodal Large Language Models' ability to reason about complex topological structures. These AI models struggle with diagrams featuring branching paths, converging flows, and cyclic dependencies, even failing at basic tasks like counting endpoints. Chemical reaction diagrams serve as the testing ground because they naturally encompass diverse structures from linear chains to cyclic graphs. The benchmark comprises 1,618 expert-annotated question-answer pairs organized across four hierarchical task dimensions. Existing evaluation methods have focused primarily on semantic comprehension rather than structural reasoning capabilities. An extensive evaluation across 17 MLLMs demonstrates these fundamental limitations in visual reasoning. The research was announced on arXiv with the identifier 2604.15994v1.