Learning Mechanistic Reasoning for Chemical Reactions with Large Language Models

Area of Science:

• Computational Chemistry
• Artificial Intelligence in Chemistry
• Organic Reaction Mechanisms

Background:

• Current large language models (LLMs) struggle with accurate chemical reaction mechanism prediction, often exhibiting physical inconsistencies.
• Specialized models for mechanism inference have limited generalization across diverse chemical spaces.
• Understanding reaction mechanisms is crucial for advancing LLMs' chemical intelligence.

Purpose of the Study:

• To develop a large-scale dataset for training LLMs on reaction mechanisms.
• To establish a rigorous benchmark for evaluating hierarchical mechanism reasoning in LLMs.
• To enhance the chemical reasoning capabilities of LLMs through mechanism-aware training.

Main Methods:

• Construction of a novel, large-scale reasoning dataset focused on reaction mechanisms.
• Development of the FukuyamaBench benchmark using Fukuyama's Advanced Organic Reaction Mechanism book.
• Fine-tuning a large language model (Qwen3-30B-A3B) on the reaction mechanism dataset.

Main Results:

• The fine-tuned Qwen3-30B-A3B model achieved an 8.3% exact pathway match on FukuyamaBench Set A.
• This performance surpasses the specialized FlowER model, which achieved 5.1%.
• Demonstrates the effectiveness of mechanism-aware training for improving LLM chemical reasoning.

Conclusions:

• Mechanism-aware training significantly enhances the chemical reasoning abilities of large language models.
• The developed dataset and benchmark facilitate more robust evaluation of LLMs in chemical mechanism inference.
• Future work can leverage these resources to build more chemically intelligent AI systems.