Branching Flows: Discrete, Continuous, and Manifold Flow Matching with Splits and Deletions

2026-06-16 · Source: stat.ML updates on arXiv.org · Field: Technology & Digital — Artificial Intelligence & Machine Learning, AI for Scientific Discovery · Depth: Expert, extended

Summary

Branching Flows, introduced on November 12, 2025, is a novel generative modeling framework designed to overcome the limitations of existing diffusion and flow matching approaches in handling variable-length sequences. Unlike models that require a fixed number of elements, Branching Flows allows the number of elements in a sequence to evolve dynamically during generation. It achieves this by modeling element evolution over a forest of binary trees, where elements branch and die stochastically with learned rates. This framework is highly versatile, composing with any flow matching base process across discrete sets, continuous Euclidean spaces, smooth manifolds, and multimodal product spaces. Demonstrations include small molecule generation, antibody sequence generation, and protein backbone generation, showing it is a capable distribution learner with a stable objective and new capabilities. For instance, 10,000 samples were generated for QM9 and antibody analysis.

Key takeaway

For research scientists developing generative models for complex, variable-length data, Branching Flows provides a unified and robust solution. This framework eliminates the need for ad hoc length-handling mechanisms in diffusion and flow matching models, offering superior control over sequence length during generation. You should explore integrating Branching Flows to simplify model architectures and enhance performance in applications like protein design, small molecule synthesis, or advanced text generation where output length is inherently unknown.

Key insights

Branching Flows enables generative models to naturally handle variable-length sequences by stochastically evolving elements over binary trees.

Principles

• Generative models can control sequence length via stochastic branching and deletion.
• Augmenting state space with a tree structure enables hierarchical evolution.
• Flow matching extends to variable-length, multimodal data.

Method

Augment a base Markov generator with branching and deletion processes, where elements evolve along a forest of binary trees, splitting and deleting based on learned rates, trained via conditional generator matching.

In practice

• Generate variable-length small molecules, antibody sequences, or protein backbones.
• Compose with existing flow matching base processes for diverse data types.

Topics

• Branching Flows
• Generative Modeling
• Flow Matching
• Variable-Length Sequences
• Protein Design
• Small Molecule Generation

Best for: AI Scientist, Research Scientist

Related on AIssential

• Flow matching for generative modelling in bioinformatics and computational biology — Flow matching is a principled generative AI framework for learning complex data transformations in biology.
• A unified perspective on fine-tuning and sampling with diffusion and flow models — A unified framework clarifies fine-tuning and sampling in diffusion/flow models via exponential tilting.
• Can We Build Scene Graphs, Not Classify Them? FlowSG: Progressive Image-Conditioned Scene Graph Generation with Flow Matching — FlowSG progressively generates scene graphs by jointly denoising discrete semantics and continuous geometry via hybrid flow matching.
• Categorical Flow Maps — Categorical Flow Maps enable accelerated, few-step categorical data generation via continuous flow matching and self-distillation.
• Graphical-Probabilistic Modeling of Generative Flows in LLM-Native Software Systems — The article proposes Generation Networks, a graphical-probabilistic modeling framework, to bring principled design and analysis to stochastic LLM-native software.
• Latent Process Generator Matching — Latent Process Generator Matching generalizes generative models by learning from time-dependent, projected latent Markov processes.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by stat.ML updates on arXiv.org.