LLM-to-Flow Distillation Aligns Scientific Generators With Natural Language Goals

ETH, Oxford and collaborators introduced LLM-to-Flow Distillation (LFD), which uses a frozen LLM as a teacher to distill high-level natural language discovery goals into domain-specific scientific generative models, aligning them with human intent without explicit proxy rewards.

2026-10-15 ~ 2026-10-15 · 3 related posts

• LLM-to-Flow Distillation: teaching scientific generators to follow natural-language goals via DPO — mmbronstein · 2026-10-15
• LLM-to-Flow Distillation Adapts Scientific Generators to Natural-Language Discovery Goals — rbhar90 · 2026-10-15
• LFD Distills Natural Language Goals from LLMs into Molecular Generators, Beating FlowGRPO — AllThingsApx · 2026-10-15