Expanding Flow Maps
The paper proposes generative flows whose output size can grow during generation.
Tang and Chatterjee introduce EFlows, which move between distributions of increasing dimensionality by adding conditional noise as the state expands. EFMs distill that process into few-step maps built from an expand operator and a transport map. The framework covers new coordinates or tokens, and extends to discrete simplex settings for variable-size graphs and variable-length sequences. ArXiv · AI/CL/LG's note
Tang and Chatterjee introduce EFlows, which move between distributions of increasing dimensionality by adding conditional noise as the state expands. EFMs distill that process into few-step maps built from an expand operator and a transport map. The framework covers new coordinates or tokens, and extends to discrete simplex settings for variable-size graphs and variable-length sequences. ArXiv · AI/CL/LG's note
score 5