Reaction
One-hot flow is claimed to rise from 4% to 50% on TinyGSM without self-conditioning
The researchers say their work explores design choices for continuous diffusion language models, including how to embed tokens.
TLDR
The researchers say their work raised one-hot flow performance on TinyGSM from 4% to 50% without self-conditioning. A separate commenter describes self-conditioning as a crutch and says DoubleCat/SemiCat seems to offer a principled way to avoid it.
Combined views
2.2K
1 Source, first seen ago
