Molecular LLMs Show Fragile Generalization to Structural Perturbations

A Molecular Perturbation framework reveals that LLMs trained on chemical sequences exhibit narrow local trust regions. Even single-edit structural changes via Graph Edit Distance (GED) cause significant performance drops, suggesting models fail to capture underlying topological constraints.