Training Chemical Plausibility-Aware Large Language Models for Single-Step Retrosynthesis
Summary
This paper introduces an enhanced C3LM model for single-step retrosynthesis using Top-K prompting and a larger training dataset, demonstrating competitive performance with conventional models on the URSA-expert-2026 benchmark.
View Cached Full Text
Cached at: 08/20/26, 04:02 AM
Paper page - Training Chemical Plausibility-Aware Large Language Models for Single-Step Retrosynthesis
Source: https://huggingface.co/papers/2608.18940 This work extends the previously introduced ChemCensor framework for single-step retrosynthesis (https://arxiv.org/abs/2602.03554). It addresses the one-to-many nature of retrosynthetic prediction, where a target molecule may have several chemically valid synthetic routes, making single-answer and exact-match evaluation insufficient. The study introduces Top‑K prompting and training to generate multiple diverse reaction proposals rather than a single predicted disconnection.
A new C3LM version is trained on CREED-CCV-2+USPTO-XL, an expanded dataset of approximately 45.6 million verified reactions derived from expert-coded templates. Combined with the Top‑K strategy, this larger training set enables C3LM to compete with - and in some settings surpass - strong conventional, non-LLM retrosynthesis models on the URSA-expert-2026 benchmark.
Moreover, the study shows that the chemical spaces of reactions generated by the new C3LM and the best conventional models are complementary rather than intersecting, making the contribution of the new model more substantial if the model is harnessed into a synthetic planning engine.
Similar Articles
Simulator-Grounded Large Language Models for Industrial Causal Reasoning: Tool-Use, Structured Injection, and Plant-Portable Retrieval for Wastewater Treatment Decision Support
This paper compares three methods for grounding a frozen LLM (Qwen2.5-32B-Instruct) in a wastewater treatment simulator for causal question answering, including live tool-use, structured parameter injection, and a small decoupled retrieval model, achieving high accuracy and cross-plant transfer.
Guiding large language models to predict edit sequences for molecular synthesizability optimization
A Nature Machine Intelligence paper on guiding large language models to predict molecular edit sequences for optimizing the synthesizability of candidate drug molecules.
ChemVA: Advancing Large Language Models on Chemical Reaction Diagrams Understanding
ChemVA is a framework that addresses the visual and semantic bottlenecks in LLMs for understanding chemical reaction diagrams, achieving 92% structural recognition accuracy and consistent ~20 percentage point gains across nine LLMs.
Uncertainty-aware reinforcement learning for chemical language models
Proposes two complementary approaches to incorporate predictive uncertainty into reinforcement learning for chemical language models, improving robustness and increasing true hit rate by 0.25 in de novo molecular design.
Agentic generation of verifiable rules for deterministic, self-expanding reaction classification
This paper presents a multi-agent LLM pipeline that automatically generates and verifies reaction rules for chemical synthesis, expanding a standard taxonomy from 68 to 14,073 classes without human curation, achieving 97.7% classification accuracy on unseen reactions.