Preprint reports chemistry-aware language model for retrosynthesis

A preprint describes C3LM, a chemistry constraint-consistent language model trained for single-step retrosynthesis on a dataset of about 45.6 million verified reactions. The authors say their Top-K prompting approach, combined with ChemCensor-based and novelty-oriented rewards, achieved state-of-the-art results on the OOD URSA-expert-2026 benchmark. They also report that the model and conventional systems explored complementary reaction spaces, supporting ensemble-based synthesis planning.