Publication
MolViBench: Evaluating LLMs on Molecular Vibe Coding
In brief
A benchmark for executable molecular coding across 12 real-world drug-discovery workflows.
Cite this work View paper
Citation
Jiatong Li*, Yuxuan Ren, Weida Wang, Changmeng Zheng, Xiao-yong Wei, Qing Li, and Yatao Bian. (2026). MolViBench: Evaluating LLMs on Molecular Vibe Coding. arXiv preprint arXiv:2605.02351.
About this work
MolViBench evaluates code generated by language models for molecular research. It contains 358 tasks across five cognitive levels and 12 drug-discovery workflows, ranging from individual API calls to virtual-screening pipelines.
The evaluation combines type-aware output comparison with AST-based API-semantic analysis to assess execution and chemical correctness. The study tests nine coding models and compares three molecular coding workflows.