Towards Reliable Chemical Retrosynthesis: A Self-Reflection Approach for Robust Retrosynthesis Planning

ACL ARR 2026 January Submission9268 Authors

Keywords: Chemical Retrosynthesis, Self-Reflection, Machine Learning in Chemistry, AI for Science, Large Language Models, GRPO, Self-evolving​

Abstract: Retrosynthetic analysis is a critical task in organic chemistry. However, the application of Large Language Models (LLMs) in retrosynthesis is often hindered by hallucinations and the high demand for internalizing vast chemical knowledge. To address these challenges, we propose ChemValiSynth, a framework designed for retrosynthetic prediction. It constructs high-quality training data via a self-reflective loop. Furthermore, we introduce a decoupled dual-model architecture that separates molecular analysis from strategy prediction, thereby enhancing the model's capacity to incorporate extensive chemical information. Empirical results on the USPTO-50K benchmark report a Top-1 accuracy of 58.7\%, validating the effectiveness of our proposed approach.

Paper Type: Long

Research Area: AI/LLM Agents

Research Area Keywords: tool use,reinforcement learning in agents

Contribution Types: Approaches to low-resource settings, Data resources, Data analysis

Languages Studied: English

Submission Number: 9268