This flow model learns both what to generate and how large it should be
KevinKaichuang · x · 2026-07-28
Expanding Generative Flows formalize generation where output size is learnable
The paper proposes Expanding Generative Flows (EFlows) and Expanding Flow Maps (EFMs) for generating data whose dimensionality or sequence length can increase during generation.
What the method does
- Starts from an expanding interpolant that grows the state with conditional noise.
- Learns an expand operator to add coordinates or tokens.
- Learns a transport map to denoise and advance the expanded state.
- Recovers standard fixed-size flow models as a special case when expansion is the identity.
Why it matters
- Handles continuous and discrete settings.
- Extends to variable-size graph generation and variable-length sequence generation.
- Treats output size as a first-class controllable degree of freedom.
The paper positions this as a principled framework for tasks where the model should decide not just what to generate, but how big the output should be.
Related event: New Flow Models Enable Dynamic Dimension Expansion During Denoising(2 posts)→
More from Research
- Paper proposes an eight-part vocabulary for multi-agent research systems — Bardiya Akhbari · 2026-07-28
- A step-by-step hand derivation shows how residual connections power deep nets — ProfTomYeh · 2026-07-28
- AI agents should talk to each other, share context, and learn skills together — heyshrutimishra · 2026-07-28
- Graft argues code agents should inject repo context automatically, not wait for MCP calls — shhdwi · 2026-07-28
- Kimi is described with a linear-attention variant and attention residuals — burny_tech · 2026-07-28
- Nature Communications links NLP embeddings to flexible semantic retrieval in the brain — bttyeo · 2026-07-28