🧬 Just saying "generate peptides with anticancer activity and low toxicity" in plain language is now enough to steer a molecule generator toward that goal. Title: LLM-to-Flow Distillation: Teaching Natural Language Goals to Scientific Generators URL: Flow and diffusion models are great at generating valid molecules and peptides, but aligning them to a natural-language goal usually needed an explicit proxy reward. This paper's LFD removes that requirement by turning an LLM into a teacher for a pretrained generator student. Here are 3 highlights. 🎯 No proxy reward needed to hit the goal The LLM watches the generator's current samples and sets an achievable sub-goal step by step, nudging the generator toward the target region via DPO, round after round. 📊 Absolute scores are unreliable, but relative judgments aren't Asking the LLM for a 1-100 score gives numbers far off from ground truth, but which-is-better pairwise judgments stay consistently reliable, which is why LFD uses a scalable batch-classification scheme instead of pairwise duels. 🏆 Consistently beats existing methods LFD hits 69.06% pLM4CPPs for cell-penetrating peptides (vs. 57.04% for DiffusionNFT), 53.54% for anticancer peptides (vs. 35.94% for Guidance), and the best Vina score of -8.48 for dopamine D2 receptor docking among all methods. This feels especially valuable for real experimental science, where a clean proxy reward usually isn't available. #LLM# #AIDrugDiscovery#