LLM-to-Flow Distillation adapts scientific generators to natural-language goals via DPO

ETH Zurich researchers and collaborators introduce LLM-to-Flow Distillation (LFD), a method that uses a frozen LLM to grade scientific generator outputs and distills those preferences into the generator through DPO. The authors say LFD needs no task-specific proxy rewards, works with any frontier LLM through API access, and outperforms reward-guided methods FlowGRPO and DiffusionNFT on held-out computational evaluations. The result is reported on drug-design generators, where the authors say the base model produced no usable hits in 38,500 samples before training. Rohan Paul, who reported the work on X, gives scores of 69.06 for LFD and 57.04 for the comparison method, figures that come from the paper rather than an independent check.