Haocheng Tang, Xingyu Dang, Junmei Wang

Researchers adapted DeepSeek-OCR-2 to recognize molecular structures in 2D chemical diagrams by framing the task as image-conditioned SMILES generation. The team developed a two-stage fine-tuning approach using LoRA followed by selective full-parameter tuning to stabilize training on a dataset combining synthetic PubChem renderings and real USPTO patent images. The resulting model, MolSeek-OCR, matches the best image-to-sequence baselines but underperforms state-of-the-art graph-based methods, and reinforcement learning and data curation failed to close the gap on strict sequence-level accuracy.