Research map / Label-efficient and robust learning / Recognition and long-tail
MolSight: Optical Chemical Structure Recognition with SMILES Pretraining, Multi-Granularity Learning and Reinforcement Learning
Wenrui Zhang, Xinggang Wang, Bin Feng, Wenyu Liu
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add chemical bond head to SMILES decoder (ablation) | MolSight · USPTO | graph accuracy: 93.7 -> 94.2 | Improved |
| jointly train atom localization with SMILES and bond tasks (ablation) | MolSight · USPTO | stereo accuracy: 10.6 vs 80.9 (SMILES+edge) | Hurt |
| train coordinate head separately from other tasks (ablation) | MolSight · USPTO | stereo accuracy: 10.6 -> 29.4 | Improved |
| add two extra decoder layers for coordinate branch (L1 loss) (ablation) | MolSight coordinate head · USPTO | stereo accuracy: 29.4 -> 82.4 | Improved |
| use MLE (Laplace) loss instead of L1 loss for coordinates (ablation) | MolSight coordinate head · USPTO | stereo accuracy: 82.4 -> 83.7 | Improved |
| RL post-training with Tanimoto reward only (ablation) | MolSight SFT · Stereo-2k | stereo accuracy: 80.1 -> 81.8 | Improved |
| RL post-training with weighted reward (ablation) | MolSight SFT · Stereo-2k | stereo accuracy: 80.1 -> 87.1 | Improved |
| RL post-training with weighted reward (ablation) | MolSight SFT · CLEF-2012 (out of domain) | stereo accuracy: 71.0 -> 80.6 | Improved |
| RL post-training with weighted reward (ablation) | MolSight SFT · CLEF-2012 (out of domain) | Tanimoto: 90.0 -> 90.0 | No clear effect |
| use stereo reward alone vs weighted reward in RL (ablation) | MolSight RL · Stereo-2k | stereo accuracy: 87.1 -> 86.5 | Hurt |
| MolSight pretraining of image encoder instead of ImageNet pretraining (ablation) | EfficientViT · MoleculeNet BBBP | ROC-AUC: 63.6 -> 68.0 | Improved |
| MolSight vs MolScribe fine-tuned 30 epochs | MolSight · CLEF | exact match accuracy: 85.5 -> 84.9 | No clear effect |
Closest papers by meaning
- Atom-Level Optical Chemical Structure Recognition with Limited Supervision (CVPR 2024 · similarity 0.91)
- RFL: Simplifying Chemical Structure Recognition with Ring-Free Language (AAAI 2025 · similarity 0.89)
- TinyChemVL: Advancing Chemical Vision-Language Models via Efficient Visual Token Reduction and Complex Reaction Tasks (AAAI 2026 · similarity 0.88)
- MarkushGrapher: Joint Visual and Textual Recognition of Markush Structures (CVPR 2025 · similarity 0.88)
- MolVision: Molecular Property Prediction with Vision Language Models (NeurIPS 2025 · similarity 0.87)
- ChemVLM: Exploring the Power of Multimodal Large Language Models in Chemistry Area (AAAI 2025 · similarity 0.87)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
