Back to Basics: Improving Molecular Understanding in LLMs via SMILES-Graph Translation

The paper introduces MolBasic, a structure-first framework for molecular language models. It treats bidirectional SMILES-to-graph translation as the central task in a multi-level structure perception benchmark, then uses progressive learning and standardized chain-of-thought supervision to move from structural acquisition to higher-level molecular reasoning. The authors report improved basic structure understanding and robust gains on downstream tasks such as molecular property prediction and objective optimization. The abstract does not provide numerical results, dataset sizes, baseline details, or ablation results.