LLM-to-Flow Distillation lands kinase hit for about $77

The pre-trained model produced zero valid hits across 38,500 samples on a kinase-focused campaign. After adapting it with LLM-to-Flow Distillation, the paper reports, a hit appeared inside a 512-sample batch at round six.

The setup from Riccardo De Santi and co-authors is deliberately plain. A domain-specific flow generator cannot reach a far-off chemical goal on its own, so a frozen large language model is placed above it as a judge. "A frozen LLM acts as an adaptive teacher for a domain-specific generator," the abstract states. The teacher writes a curriculum of intermediate sub-goals, scores candidate molecules each round, and those preferences are iteratively distilled back into the generator through Direct Preference Optimization, pushing it toward "poorly covered, out-of-distribution regions" without a hand-crafted reward model.

The teacher is Gemini. "All experiments use Gemini 3.1 Pro to construct the curriculum and Gemini 3.7 Flash for LLM judgments," the paper notes. Three campaigns were logged end-to-end: a cell-penetrating peptide run at roughly $8.83 in API cost over about 5.3 hours on an NVIDIA GH200, an anticancer-peptide run at about $37.4 over roughly 20 hours, and the kinase-focused molecular-design campaign at about USD 77 on a single GPU with 8 CPUs under a 48-hour time limit.

LFD is compared against "classifier Guidance, DPO, and recent state-of-the-art reward-guided fine-tuning methods, including Flow-GRPO and DiffusionNFT." On the peptide tasks, the reported hit rate moves from 8.81% to 69.06% for cell-penetrating peptides, and from 26.43% to 53.54% for anticancer peptides.

The paper is a preprint, the kinase result is a single hit, and no wet-lab validation is reported.