Report
Hierarchical continuous diffusion LMs jointly diffuse tokens and coarse semantic clusters
The H-CDLM announcement claims denoising hierarchical token representations brings better generations for less compute.
TLDR
A post introducing Hierarchical Continuous Diffusion LMs (H-CDLMs) says they jointly diffuse tokens and coarse semantic clusters to denoise hierarchical token representations. It claims the approach produces better generations with less compute.
Combined views
2.8K
2 Sources, first seen ago
