Learning Mechanistic Reasoning for Chemical Reactions with Large Language Models

2026-07-14

AI Summary

This study addresses the limitations of existing chemical large language models, which predominantly focus on coarse-grained reaction name prediction and lack the capacity for stepwise mechanistic reasoning, often yielding physically inconsistent or hallucinated outputs, while specialized small models suffer from poor generalization. To bridge this gap, the work presents the first application of large language models to fine-grained, multi-step organic reaction mechanism inference. The authors construct a large-scale structured mechanistic dataset, fine-tune the Qwen3-30B-A3B model, and introduce a hierarchical evaluation framework along with the FukuyamaBench benchmark. On FukuyamaBench Set A, the model achieves an 8.3% exact path match rate, significantly outperforming FlowER (5.1%), thereby demonstrating the efficacy of mechanism-aware training and advancing chemical AI from product prediction toward genuine mechanistic understanding.

Abstract

Reaction mechanisms consist of the step-by-step sequences of elementary reactions that explain chemical transformations. Learning the mechanism logic is therefore essential for enhancing the fundamental chemical intelligence of large language models (LLMs). The stepwise deduction of reaction mechanism aligns naturally with the reasoning paradigms of reasoning LLMs. However, current chemical LLMs primarily emphasize coarse-grained name reactions for product prediction and retrosynthesis, often leading to physical inconsistencies and hallucinations. In contrast, specialized small-scale generative models for mechanism inference typically suffer from restricted generalization capacity across diverse chemical spaces. To overcome these limitations, we built a novel, large-scale reasoning dataset of reaction mechanisms. Furthermore, we established the FukuyamaBench, a difficult benchmark derived from Fukuyama's Advanced Organic Reaction Mechanism book, to rigorously evaluate model performance on hierarchical mechanism reasoning. Our fine-tuned Qwen3-30B-A3B achieves 8.3% exact pathway match on FukuyamaBench Set~A, surpassing the specialized FlowER model (5.1%), demonstrating that mechanism-aware training substantially enhances chemical reasoning in language models.

Problem

• reaction mechanism
• large language models
• chemical reasoning
• generalization
• mechanistic reasoning

Innovation

• reaction mechanism reasoning
• large language models
• FukuyamaBench
• mechanism-aware training
• chemical intelligence