Publication
Speak-to-Structure: Evaluating LLMs in Open-domain Natural Language-Driven Molecule Generation
In brief
S²-Bench evaluates one-to-many molecular design from natural-language requirements.
Cite this work View paper
Citation
Jiatong Li*, Junxian Li*, Yunqing Liu, Dongzhan Zhou, and Qing Li. (2026). Speak-to-Structure: Evaluating LLMs in Open-domain Natural Language-Driven Molecule Generation. KDD 2026.
About this work
Speak-to-Structure evaluates molecular generation from open-ended natural-language instructions. Its S2-Bench benchmark allows multiple valid molecules for a design brief and includes molecule editing, optimization, and customized generation.
The study evaluates 31 language models and introduces OpenMolIns for instruction tuning. A tuned Llama3.1-8B outperforms GPT-4o and Claude-3.5 on S2-Bench in the reported experiments.