This research introduces a hierarchical approach to molecular representation in multimodal large language models, recognizing that different molecular features matter for different tasks.

• Addresses the overlooked multi-level nature of molecular graphs in current LLM applications
• Demonstrates how varying levels of molecular structure representation improve task-specific performance
• Provides a framework for integrating complex graph data with language models
• Creates more effective models for predicting chemical reactions and properties

For the medical field, this approach significantly advances drug discovery and pharmaceutical research by enabling more accurate predictions of molecular behavior and interactions, potentially accelerating development timelines and reducing costs.

Exploring Hierarchical Molecular Graph Representation in Multimodal LLMs