C3LM: Chemical Plausibility-Aware LLM for Retrosynthesis
August 18, 2026
The C3LM model uses a 45.6 million reaction dataset and Top-K prompting to improve single-step retrosynthesis accuracy. It integrates ChemCensor-based rewards to achieve state-of-the-art performance on the OOD URSA-expert-2026 benchmark.
HOW THIS AFFECTS YOU
●
researcherYou can leverage Top-K prompting to better capture the one-to-many nature of chemical synthesis.
●
healthThis advances automated drug discovery through more plausible chemical reaction predictions.