Papers

Large language models (LLMs) are rapidly transforming scientific discovery, showing promise in hypothesis generation, literature understanding, and symbolic reasoning. Yet, their capacity to conduct iterative, feedback-driven molecular design---a hallmark of real-world drug and materials discovery---remains underexplored. Existing benchmarks typically cast molecular tasks as one-shot question-answering or text-to-molecule translation, neglecting the iterative propose-evaluate-refine process central to scientific practice. We propose \textbf{ICIMBench}, an \textit{In-Context Iterative Molecular Design Benchmark} that evaluates LLMs in multi-turn molecular design episodes. In each task, the model receives a natural-language specification, generates candidate molecules in SMILES format, and iteratively refines them based on deterministic oracle feedback from RDKit. We introduce the \textbf{NumEval} metric---the number of evaluations required to satisfy the target---which captures both performance efficiency and robustness under realistic evaluation budgets. Experiments on frontier models (GPT-5, DeepSeek-V3.2, Intern-S1) show that while single-property design is largely solved (NumEval $=1$) by state-of-the-art LLMs like GPT-5, multi-property optimization remains a strong challenge, especially under coupled constraints such as lipophilicity and scaffold similarity. ICIMBench provides a principled framework for probing the in-context reasoning and adaptive optimization abilities of LLMs, paving the way toward autonomous, language-driven molecular discovery.