Reproduce “GTR-CoT: Graph Traversal as Visual Chain of Thought for Molecular Structure Recognition” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Jingchao Wang, Yifan He, Haote Yang, Jiang Wu, Lingli Ge, Xingjian Wei, Yinfan Wang, Linye Li and 5 more
GTR-VL is a Vision-Language Model that enhances Optical Chemical Structure Recognition by implementing a graph traversal "Chain of Thought" and consistently representing chemical abbreviations as "superatoms". The framework outperforms prior methods by approximately 9 percentage points on printed molecules with abbreviations and achieves a significant improvement from 9.53% to 75.44% SMILES exact match on hand-drawn molecular structures.
Key insights
- Interleaving atom and bond predictions via a "Graph Traversal as Visual Chain of Thought" mechanism provides mutual constraints and reduces complexity, improving recognition accuracy.
- Maintaining semantic consistency by faithfully recognizing visual abbreviations as "superatoms" is crucial for model generalization and performance on real-world chemical diagrams.
- Reinforcement learning with a graph-based reward function allows a VLM to effectively learn from weak supervision and achieve high accuracy on challenging hand-drawn molecular structures.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.