LLM-to-Flow Distillation: Teaching Natural Language Goals to Scientific Generators
Summary
论文提出 LLM-to-Flow Distillation（LFD），利用冻结的 LLM 作为自适应教师构建课程并通过 DPO 将偏好蒸馏到科学生成器中，无需显式代理奖励即可用自然语言目标控制治疗性肽和小分子结构的生成设计。
View Cached Full Text
Cached at: 10/15/26, 10:58 AM
Paper page - LLM-to-Flow Distillation: Teaching Natural Language Goals to Scientific Generators
Source: https://huggingface.co/papers/2610.17246
Abstract
LLM-to-Flow Distillation aligns scientific generators with high-level discovery goals by using a frozen LLM to build adaptive curricula and distill preferences via DPO, enabling control over peptide and small-molecule design.
Generative discovery in science requires coupling high-level scientific reasoning with the ability to generate samples respecting the structure of a design space. Yet these capabilities remain largely separated: foundation LLMs encode broad chemical and biological knowledge, while domain-specific flow- anddiffusion-based generatorscapture the priors needed to produce valid sequences and geometries. We study how to align such scientific generators to high-level natural-language discovery goals without requiring an explicit proxy reward, which is often unavailable or poorly captured by learned predictors in experimental sciences. We introduceLLM-to-Flow Distillation(LFD), where a frozen LLM acts as an adaptive teacher for a domain-specific generator. Given a goal such as ‘generate peptides with promising anticancer activity’, the teacher constructs a curriculum of intermediate sub-goals aware of the flow’s current capabilities, and progressively directs them toward the original discovery goal. At each round, the LLM judges sampled candidates under the current objective. These judgments induce preferences that are iteratively distilled into the generator throughDPO. This procedure provides dense learning signals even when the initial model places negligible probability on designs satisfying the final goal, enabling adaptation toward poorly covered, out-of-distribution regions. Experiments ontherapeutic peptidesandsmall-molecule structure designdemonstrate that LFD enables a new form of generative control: aligning arbitrary domain-specific scientific generators to scientific discovery tasks expressed via high-level natural-language goals.
View arXiv pageView PDFProject pageGitHub13Add to collection
Models citing this paper0
No model linking this paper
Cite arxiv.org/abs/2610.17246 in a model README.md to link it from this page.
Datasets citing this paper0
No dataset linking this paper
Cite arxiv.org/abs/2610.17246 in a dataset README.md to link it from this page.
Spaces citing this paper0
No Space linking this paper
Cite arxiv.org/abs/2610.17246 in a Space README.md to link it from this page.
Collections including this paper0
No Collection including this paper
Add this paper to acollectionto link it from this page.
Similar Articles
@rohanpaul_ai: You can now type what you want in plain English and have a drug-design model chase it. Normally, to make a model genera…
Riccardo De Santi et al. introduce LLM-to-Flow Distillation (LFD), a method that lets an LLM grade a scientific generator's outputs against plain-English goals, removing the need for task-specific reward models and outperforming direct reward-function training in drug/molecule design.
FlowLM: Few-Step Language Modeling via Diffusion-to-Flow Adaptation
FlowLM introduces a flow matching language model derived from pre-trained diffusion models via efficient fine-tuning, enabling high-quality few-step text generation that rivals 2,000-step diffusion sampling with far fewer training epochs.
Distilled Reinforcement Learning for LLM Post-training
Introduces Distilled Reinforcement Learning, a method that uses a teacher model to provide fine-grained token-level gradient signals for LLM post-training, combining reinforcement learning with knowledge distillation.
PPDL: LLM-Based Flows as Probabilistic Programs
This paper introduces PPDL, a probabilistic language for programming LLM-based flows that enables developers to quantify and propagate uncertainty throughout applications, with experimental and case studies on theorem proving.
Training LLM Judges from Language Feedback via Position-Selective Self-Distillation
The paper proposes position-selective self-distillation for training LLM judges from natural language feedback, using per-position entropy shifts to mask memorization-prone tokens and improve out-of-distribution generalization over outcome-supervised RL methods like GRPO by 2–9 points on subjective tasks.