mlsanityData to 14 Oct 2026

LLM-to-Flow Distillation: Teaching Natural Language Goals to Scientific Generators

arXiv 2610.17246 ↗First posted 12 Oct 2026cs.LGcited by 0

On the map: LLM reasoning → Prompting and chain-of-thought reasoning also fits: Test-time compute and efficient thinking

Abstract (arXiv)

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- and diffusion-based generators capture 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 introduce LLM-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 through DPO. 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 on therapeutic peptides and small-molecule structure design demonstrate 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.

Key facts. No citing papers in the map yet (data through 14 Oct 2026). Placed in Prompting and chain-of-thought reasoning, one of 17 directions in LLM reasoning. Builds on 59 of its 95 references that are in the map. First posted on arXiv 12 Oct 2026.

Builds on

59 of 95 references.

References · 59

  1. 2023Judging LLM-as-a-Judge with MT-Bench and Chatbot Arenacited by 8,626
  2. 2021Diffusion Models Beat GANs on Image Synthesiscited by 8,255
  3. 2023Direct Preference Optimization: Your Language Model is Secretly a Reward Modelcited by 8,065
  4. 2015Deep Unsupervised Learning using Nonequilibrium Thermodynamicscited by 7,058
  5. 2022Flow Matching for Generative Modelingcited by 5,702
  6. 2019Generative Modeling by Estimating Gradients of the Data Distributioncited by 4,101
  7. 2023G-Eval: NLG Evaluation using GPT-4 with Better Human Alignmentcited by 2,276
  8. 2022Scaling Laws for Reward Model Overoptimizationcited by 1,095
  9. 2023Training Diffusion Models with Reinforcement Learningcited by 982
  10. 2023On-Policy Distillation of Language Models: Learning from Self-Generated Mistakescited by 939
  11. 2023Diffusion Model Alignment Using Direct Preference Optimizationcited by 790
  12. 2024Self-Rewarding Language Modelscited by 681
38 more
  1. 2025Flow-GRPO: Training Flow Matching Models via Online RLcited by 661
  2. 2025Accelerating scientific discovery with Co-Scientistcited by 401
  3. 2024Flow Matching Guide and Codecited by 301
  4. 2020Emergent Complexity and Zero-shot Transfer via Unsupervised Environment Designcited by 275
  5. 2017Molecular De Novo Design through Deep Reinforcement Learningcited by 267
  6. 2021Therapeutics Data Commons: Machine Learning Datasets and Tasks for Drug Discovery and Developmentcited by 264
  7. 2024Direct Language Model Alignment from Online AI Feedbackcited by 260
  8. 2017Automatic Goal Generation for Reinforcement Learning Agentscited by 259
  9. 2023Using Human Feedback to Fine-tune Diffusion Models without Any Reward Modelcited by 252
  10. 2020GEOM: Energy-annotated molecular conformations for property prediction and molecular generationcited by 243
  11. 2024Adjoint Matching: Fine-tuning Flow and Diffusion Generative Models with Memoryless Stochastic Optimal Controlcited by 233
  12. 2025DiffusionNFT: Online Diffusion Reinforcement with Forward Processcited by 230
  13. 2017Reverse Curriculum Generation for Reinforcement Learningcited by 198
  14. 2024Large Language Models to Enhance Bayesian Optimizationcited by 135
  15. 2023Silkie: Preference Distillation for Large Visual Language Modelscited by 130
  16. 2023Motif: Intrinsic Motivation from Artificial Intelligence Feedbackcited by 102
  17. 2024Understanding Reinforcement Learning-Based Fine-Tuning of Diffusion Models: A Tutorial and Reviewcited by 98
  18. 2025Inference-Time Alignment in Diffusion Models with Reward-Guided Generation: Tutorial and Reviewcited by 88
  19. 2024OMNI-EPIC: Open-endedness via Models of human Notions of Interestingness with Environments Programmed in Codecited by 74
  20. 2023Particle Guidance: non-I.I.D. Diverse Sampling with Diffusion Modelscited by 71
  21. 2024Feedback Efficient Online Fine-Tuning of Diffusion Modelscited by 53
  22. 2024FlowLLM: Flow Matching for Material Generation with Large Language Models as Base Distributionscited by 53
  23. 2025Online Reward-Weighted Fine-Tuning of Flow Matching with Wasserstein Regularizationcited by 43
  24. 2026DiffusionOPD: A Unified Perspective of On-Policy Distillation in Diffusion Modelscited by 40
  25. 2026Flow-OPD: On-Policy Distillation for Flow Matching Modelscited by 34
  26. 2023Augmenting Autotelic Agents with Large Language Modelscited by 30
  27. 2024CurricuLLM: Automatic Task Curricula Design for Learning Complex Robot Skills using Large Language Modelscited by 22
  28. 2025FlowMol3: Flow Matching for 3D De Novo Small-Molecule Generationcited by 21
  29. 2025Flow Density Control: Generative Optimization Beyond Entropy-Regularized Fine-Tuningcited by 17
  30. 2024Eurekaverse: Environment Curriculum Generation via Large Language Modelscited by 16
  31. 2025Multi-Objective-Guided Discrete Flow Matching for Controllable Biological Sequence Designcited by 15
  32. 2025Provable Maximum Entropy Manifold Exploration via Diffusion Modelscited by 14
  33. 2025LLM Meets Diffusion: A Hybrid Framework for Crystal Material Generationcited by 13
  34. 2024Regression with Large Language Models for Materials and Molecular Property Predictioncited by 11
  35. 2025Iterative Distillation for Reward-Guided Fine-Tuning of Diffusion Models in Biomolecular Designcited by 11
  36. 2024Generalists vs. Specialists: Evaluating LLMs on Highly-Constrained Biophysical Sequence Optimization Taskscited by 10
  37. 2025Composition and Alignment of Diffusion Models using Constrained Learningcited by 10
  38. 2026Efficient Tail-Aware Generative Optimization via Flow Model Fine-Tuningcited by 8

50 most cited of 59 All 59 (JSON)

Built on by

Not cited yet.

Similar, not cited

Close in wording, no citation either way

  1. 2026LLM sequential decision making under uncertainty in biochemical domains
  2. 2025Toward Scientific Reasoning in LLMs: Training from Expert Discussions via Reinforcement Learning
  3. 2025Training a Scientific Reasoning Model for Chemistry
  4. 2025OmniScience: A Domain-Specialized LLM for Scientific Reasoning and Discovery
  5. 2026Training Large Language Models for Small-Molecule Design with Synthetic Task Scaling
  6. 2026MOOSE-Star: Unlocking Tractable Training for Scientific Discovery by Breaking the Complexity Barrier
  7. 2024Crossing New Frontiers: Knowledge-Augmented Large Language Model Prompting for Zero-Shot Text-Based De Novo Molecule Design
  8. 2025MiST: Understanding the Role of Mid-Stage Scientific Training in Developing Chemical Reasoning Models