Science

ChemCoTBench

ChemCoTBench: evaluates LLM chemical reasoning via modular chemical operations on SMILES. Per-instance binary scoring of two released models (claude3.7, ether0) across Molecule Understanding, Editing, Optimization and Reaction Prediction, using the benchmark's own RDKit-based eval logic (functional-group checks, count match, scaffold / SMILES exact match, property-improvement thresholds, MCQ accuracy).