New research reveals duality between continuous and discrete flow matching

This paper introduces a novel duality between continuous and discrete flow matching, typically considered separate constructions. By projecting continuous convex-interpolant paths through an argmax function, the research demonstrates how to derive discrete convex-interpolant paths. The study highlights that the choice of source geometry, such as Gaussian, bounded-uniform, or centered negative-exponential, significantly influences the transition timing and vocabulary-size dependence in categorical generation. Initial experiments with language modeling suggest these source-design effects can impact learned transports and early generative quality. AI

IMPACT Introduces a new theoretical framework for categorical generation that could influence future generative model architectures.

RANK_REASON The item is an academic paper published on arXiv detailing a new theoretical approach in machine learning. [lever_c_demoted from research: ic=1 ai=1.0]

Read on arXiv stat.ML →

• alphaXiv
• arXiv
• CatalyzeX Code Finder for Papers
• Connected Papers
• CORE Recommender
• DagsHub
• Flow Duality and Source Geometry for Categorical Generation
• Gaussian
• Gotit.pub
• Hugging Face
• IArxiv Recommender
• Influence Flower
• Litmaps
• machine learning
• ScienceCast
• scite Smart Citations

• paper
• model release

AI-generated summary · Google Gemini · from 1 sources. How we write summaries →