An assessment of the limitations of large language models in <scp>CRW</scp> reactive group identification

Abstract Chemical incompatibility management in process safety depends on accurate chemical reactive group identification. When chemicals are not covered by existing databases, classification often relies on expert judgment and may be inconsistent. This study was conducted based on 68 commonly used petrochemical products and the reactive group list embedded in the Chemical Reactivity Worksheet (CRW) database. Using this reference framework, the accuracy of six large language models (LLM) in identifying chemical reactive groups was systematically evaluated. The tests used repeated trials and two evaluation criteria: correct number of reactive groups and correct category assignment. The conclusion is that performance is only moderate; the models frequently underestimate or misclassify reactive groups, and outputs are not stable enough for safety‐critical use.