Molecular AI models have been acing advanced chemistry benchmarks while flunking the basics.
Researchers have published MolBasic, a training framework that addresses a quiet embarrassment in molecular large language models: despite strong scores on property prediction and molecule generation, these models routinely fail simple structure-recognition tasks. The core problem is that current models learn from SMILES — a text notation for chemical structures — without ever reliably mapping that notation to an actual molecular graph. MolBasic fixes that by making bidirectional SMILES-to-graph conversion the central training task, forcing models to reconcile sequential text representations with the underlying topology before moving on. A progressive learning scheme, anchored by a standardized Chain-of-Thought protocol, then walks models from structural basics up to higher-order reasoning.
This matters because chemistry has a foundational principle: structure determines function. A drug discovery model that can predict binding affinity but cannot reliably identify a ring system is, in a meaningful sense, pattern-matching rather than reasoning. MolBasic's gains on downstream tasks — property prediction and objective optimization — suggest that structural grounding isn't just academically tidy; it translates to measurable performance.
The approach echoes a recurring lesson in AI: models trained to skip prerequisites tend to build on sand. Fixing the foundation first is rarely the flashy path, but the benchmark gaps MolBasic closes suggest the field may have been celebrating fluency before literacy.