@r_de_santi: Generative models can’t discover what they can’t reach. We’re excited to introduce ActFlow: a continued pre-training sc…
Summary
The article introduces ActFlow, a continued pre-training scheme that expands the valid design space for flow and diffusion models, enabling out-of-distribution generative modeling and evolvable search spaces in scientific discovery.
View Cached Full Text
Cached at: 08/27/26, 05:33 AM
Generative models can’t discover what they can’t reach.
We’re excited to introduce ActFlow: a continued pre-training scheme that actively expands the valid design space reachable by flow and diffusion models. We call this generable set expansion — a new learning principle for out-of-distribution generative modeling, and a step toward evolvable search spaces for scientific discovery. (1/5)
Generative models should not only match their training data distribution — they should actively expand beyond it. ActFlow uses the model’s own generative representation to select uncertain frontier samples, queries a verifier, and updates the model accessing new OOD regions. (2/5)
Across small molecules, drug-like molecules, therapeutic peptides, and proteins, ActFlow significantly improves the coverage–validity of recent flow and discrete diffusion models. It reaches up to 6× more valid coverage than recursive SFT baselines and shows the highest diversity across all settings. (3/5)
We reframe OOD generative modeling as a reachability problem in a learned representation: which valid regions — even far-away ones — can be reached from what the model can already generate through a sequence of small, local jumps? We formalize this view and establish statistical guarantees for actively expanding a generative model beyond its initial coverage. (4/5)
Joint work with amazing collaborators and advisors: Bruce Lee, @cristianpjensen, @KimonProto6689, @sophia_tang, Cheng-Hao Liu, @pranamanam, @yisongyue, and @arkrause. (5/5)
Paper: https://arxiv.org/pdf/2606.08802 Website: https://riccardodesanti.com/projects/actflow/… Code: https://github.com/riccardodesanti/active_diffusion_models…
Similar Articles
Residual-Space Evolutionary Optimization via Flow-based Generative Models
Introduces a framework combining flow-based generative editing with evolutionary algorithms to perform optimization in residual space, enabling controllable data editing with non-differentiable objectives. Validated on MorphoMNIST and crystal data.
Flow-Direct: Feedback-Efficient and Reusable Guidance for Flow Models via Non-Parametric Guidance Field
Flow-Direct introduces a non-parametric guidance field for flow-based generative models that accumulates reward feedback persistently, improving feedback efficiency and enabling reuse of collected samples to guide generation for multiple objectives without additional reward evaluations.
@alec_helbling: A growing body of work uses continuous flow models for discrete problems like language generation. A simple recipe: rep…
The tweet discusses the growing use of continuous flow models for discrete tasks like language generation, proposing a method to represent vocabulary as one-hot vectors and learn flows in continuous space.
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.
Flow-Map GRPO: Reinforcement Learning for Few-Step Flow-Map Generators via Anchored Stochastic Composition
Proposes Flow-Map GRPO, an online RL post-training framework for deterministic few-step flow-map generators, introducing Anchored Stochastic Flow Map Composition (ASFMC) to enable stochastic optimization without altering original model parameterization. Experiments on FLUX-based MeanFlow and sCM show improvement across reward-based, perceptual, and task-level metrics.