Chemistry language model scales to 45M reactions for synthesis planning
Researchers have trained C3LM, a chemistry-focused language model, on a 45.6 million reaction dataset to improve single-step retrosynthesis prediction. The work introduces Top-K prompting to capture the inherent one-to-many nature of chemical synthesis planning, moving beyond single-answer benchmarking. By combining fine-tuning with reward signals for chemical validity and novelty, the model achieves state-of-the-art results on the URSA-expert-2026 benchmark. The finding that LLMs and conventional models explore complementary reaction spaces suggests domain-specific language models can unlock new chemical discovery pathways, with implications for automated drug and materials synthesis workflows.
Modelwire context
ExplainerThe paper's real contribution isn't just scale (45.6M reactions) but the explicit modeling of synthesis ambiguity through Top-K prompting, which treats retrosynthesis as a ranking problem rather than a single-answer prediction task. This reframes how we evaluate chemistry models.
This connects directly to the PGFS++ work from earlier this month, which tackled synthesis-aware molecular optimization by replacing indirect predictions with direct lookups. Both papers share a core insight: chemistry models fail when they ignore the gap between what's theoretically possible and what's actually synthesizable. C3LM adds a second layer by acknowledging that even valid synthesis routes aren't unique, which PGFS++ treated as a constraint but didn't explicitly model. The Monroe foundation model paper also released today shows the broader trend of domain-specific pretraining for chemistry, but Monroe focuses on bioassay prediction rather than synthesis planning, so the two solve adjacent but distinct bottlenecks.
If C3LM's Top-K approach outperforms single-answer baselines on held-out USPTO reactions from 2025 onward (data the model wasn't trained on), that confirms the method generalizes beyond benchmark contamination. If instead performance drops sharply on recent data, the gains are likely eval artifacts rather than genuine chemical reasoning.
Coverage we drew on
This analysis is generated by Modelwire’s editorial layer from our archive and the summary above. It is not a substitute for the original reporting. How we write it.
MentionsC3LM · CREED-CCV-2+USPTO-XL · ChemCensor · URSA-expert-2026 · Top-K prompting
Modelwire Editorial
This synthesis and analysis was prepared by the Modelwire editorial team. We use advanced language models to read, ground, and connect the day’s most significant AI developments, providing original strategic context that helps practitioners and leaders stay ahead of the frontier.
Modelwire summarizes, we don’t republish. arXiv cs.LG originally reported this story as “Training Chemical Plausibility-Aware Large Language Models for Single-Step Retrosynthesis”. The full content lives on arxiv.org. If you’re a publisher and want a different summarization policy for your work, see our takedown page.