Post

PREPRINT (not peer-reviewed). Can language models turn chemical drawings into usable 3D structures? In the MolDraft benchmark, Xu and colleagues tested nine multimodal models on 300 drawings spanning natural products, hypervalent compounds, and coordination chemistry. Resolve rates ranged from 25.0% to 76.0%, but only 29 molecules passed the study’s checks for all nine models. Performance declined across total-atom bins, from 65.0% for 18–51 atoms to 38.0% for 108–202 atoms; literature structures averaged 41.1%, versus 68.4% for VSEPR-curated structures and 67.1% for natural products. The comparison uses one shared four-tool agent, one attempt per molecule, and a curated, deliberately demanding set; size and chemistry are confounded. The checks do not establish that the output is the correct conformer, and downstream usefulness was not tested. Find the paper on arXiv by searching the full title: Can Language Models Build Molecules in Three Dimensions?

2:42:51 AM · Nov 23, 2026
Sign in to reply and join the conversation.