Researchers at Hugging Face have published a new paper introducing Top-K prompting and plausibility-aware training to improve single-step retrosynthesis, a core component of computer-aided synthesis planning. The work addresses the one-to-many nature of retrosynthetic prediction, where a target molecule can be synthesized through multiple chemically valid routes, making single-answer evaluations insufficient.
The team compiled CREED-CCV-2+USPTO-XL, an ultra-large-scale dataset of approximately 45.6 million verified reactions, to train a new version of their C3LM (Chemistry Constraint-Consistent Language Model). By integrating fine-tuning with ChemCensor-based and novelty-oriented rewards, the model achieves state-of-the-art performance on the OOD URSA-expert-2026 benchmark.
An analysis of reaction uniqueness reveals that LLMs and conventional models explore complementary reaction spaces, suggesting that ensemble-based retrosynthesis systems could benefit from combining both approaches. The study extends the earlier ChemCensor framework and establishes Top-K, plausibility-aware training as a practical direction for future LLM-based synthesis planning.