Researchers have developed a new approach to training large language models for single-step retrosynthesis. They introduced Top-K prompting and used an ultra-large-scale dataset of verified reactions to train the C3LM model. The results show that this method achieves state-of-the-art performance on a benchmark test, capturing diverse and plausible reaction predictions. This is significant because it addresses the one-to-many nature of single-step retrosynthesis, which has been poorly captured by previous evaluation protocols. --- Why it matters: This matters to researchers in AI because it provides a more robust way to predict chemical reactions, which can be used for computer-aided synthesis planning. The ability to capture diverse and plausible reaction predictions is crucial for developing accurate and reliable synthesis planning systems. Source: https://arxiv.org/abs/2608.18940