LLM molecular tasks depend on representation, study finds

A new study on arXiv benchmarks the performance of 16 large language models across nine molecular representations for eight chemical tasks. The research found that model performance is heavily dependent on the molecular representation used, with explicit structured text formats like CML and MolJSON excelling in structural tasks, while IUPAC proved best for semantic tasks. Chemistry-specialized models showed strong performance with SMILES but struggled with structured formats, indicating a potential bias in their evaluation. AI

IMPACT Highlights the critical need for task-specific molecular representations in LLMs for chemistry applications.

RANK_REASON The cluster contains an academic paper detailing empirical research on LLM performance with different molecular representations. [lever_c_demoted from research: ic=1 ai=1.0]

Read on arXiv cs.AI →

• arXiv
• InChI
• Large Language Models
• MolJSON
• SMILES

• paper
• model release

AI-generated summary · Google Gemini · from 1 sources. How we write summaries →