LFD Distills LLM Knowledge Into Molecular Generators Via DPO, Beating FlowGRPO

mmbronstein · x · 2026-10-15

A new preprint introduces LLM-to-Flow Distillation (LFD), a method that adapts scientific generators to high-level natural-language discovery goals like "design molecules binding to the D2 dopamine receptor."

• How it works: the LLM builds an adaptive curriculum of sub-goals matched to what the generator can currently produce, then distills goal knowledge via iterative DPO
• No proxy rewards: LFD needs no task-specific reward engineering, works with any frontier LLM via API, and requires no architecture changes
• Results: outperforms state-of-the-art reward-guided methods like FlowGRPO and DiffusionNFT on held-out benchmarks, and generalizes across scientific generators (peptides, small molecules)

Related event: ETH, Oxford and Penn Propose LFD: Distilling Natural-Language Discovery Goals from Frozen LLMs into Scientific Generators(5 posts)→

Original post →

More from Research

• kalomaze argues ensembled synthetic data can cover data-scarce domains — kalomaze · 2026-10-15
• Axiom claims proof of generalized quasi-Riemann hypothesis for all Dedekind zeta functions — basedjensen · 2026-10-15
• TorchLean expands into a verified tensor programming framework with einsum and GPU execution in Lean — burny_tech · 2026-10-15
• StreamErase hits 26 FPS real-time video object erasing on a single GPU, 200x faster than prior methods — Zhuobai Dong · 2026-10-15
• Huawei's SwarmRSI: evolving agent swarms achieve recursive self-improvement in science — Qianru Meng · 2026-10-15
• Glance Forcing cuts video distillation trajectory cost from thousands of GPU-hours to one — Zhuobai Dong · 2026-10-15

Research channel →