Can LLMs Solve Molecule Puzzles? A Multimodal Benchmark for Molecular Structure Elucidation
Kehan Guo, B. Nan, Yujun Zhou, Taicheng Guo, Zhichun Guo +5 more
2/3/2026
No official code found. Best implementation: KehanGuo2/MolPuzzle (18 stars, MIT, last commit 2026-01-02).
Abstract
Large Language Models (LLMs) have shown significant problem-solving capabili-1 ties across predictive and generative tasks in chemistry. However, their proficiency 2 in multi-step chemical reasoning remains underexplored. We introduce a new 3 challenge: molecular structure elucidation, which involves deducing a molecule’s 4 structure from various types of spectral data. Solving such a molecular puzzle, 5 akin to solving crossword puzzles, poses reasoning challenges that require inte-6