LLM-to-Flow Distillation: teaching scientific generators to follow natural-language goals via DPO
mmbronstein · x · 2026-10-15
A new arXiv paper by De Santi et al. (with Michael Bronstein and Andreas Krause) introduces LLM-to-Flow Distillation (LFD): a frozen LLM acts as an adaptive teacher for domain-specific flow/diffusion generators, building a curriculum of intermediate sub-goals toward a natural-language discovery goal (e.g., 'generate peptides with promising anticancer activity'). The LLM judges sampled candidates each round, and these preferences are distilled into the generator via DPO — providing dense learning signals even when the base generator assigns negligible probability to valid designs, without needing an explicit proxy reward. Paper, code, and website are public.
Related event: LFD: Distilling LLMs into Generators That Understand Natural Language Goals(2 posts)→
More from Research
- DuplexJev Cuts Autoregressive Decoding for Voice Agents, Answering 80 Speech Decisions in 0.1s — adventists-ai · 2026-10-15
- LLM-to-Flow Distillation Adapts Scientific Generators to Natural-Language Discovery Goals — rbhar90 · 2026-10-15
- ETH's Uncertainty-Aware World Model Makes Offline RL Work on Real Robots — breadli428 · 2026-10-15
- 167,356 simulated shop-days: paper shows eval world design writes the model behavior it scores — liminal_bardo · 2026-10-15
- Multilingual NLP roadmap: researcher maps open problems from data selection to LID across 350+ languages — StellaLisy · 2026-10-15
- Beyond RAG: Injecting Tensor Memory Cartridges Into Layer-3 of a Frozen LLM — Nearby_Indication474 · 2026-10-15