Training Chemical Plausibility-Aware Large Language Models for Single-Step Retrosynthesis

Summary

Single-step retrosynthesis is central to computer-aided synthesis planning, but its one-to-many nature is poorly captured by single-answer benchmarks. To address this, the authors introduce Top-K prompting as a training and inference paradigm for diverse, plausible predictions. They compile CREED-CCV-2+USPTO-XL, an ultra-large-scale dataset of ~45.6 million verified reactions, to train C3LM (Chemistry Constraint-Consistent Language Model). Fine-tuning with ChemCensor-based and novelty-oriented rewards yields state-of-the-art performance on the OOD URSA-expert-2026 benchmark. Analysis of reaction uniqueness shows LLMs and conventional models explore complementary spaces, suggesting ensemble-based systems. This establishes Top-K plausibility-aware training as a practical direction for robust LLM-based synthesis planning.