Continuous Diffusion for Text

Overview July 17, 2026

Coordinators: Dhruvesh Patel , Benjamin Rozonoyer

diffusion continuous-diffusion foundations

Continuous Diffusion for Categorical Data — Dieleman et al.
SSD-LM: Semi-autoregressive Simplex-based Diffusion Language Model for Text Generation and Modular Control — Han et al.
CANDI: Hybrid Discrete-Continuous Diffusion Models — Pynadath et al.
The Diffusion Duality — Sahoo et al.
Continuously Augmented Discrete Diffusion Model for Categorical Generative Modeling — Zheng et al.
ELF: Embedded Language Flows — Hu et al.
Self-conditioned Flow Map Language Models via Fixed-point Flows — Yoo et al.

Proposed reading cluster from #diffusion-reading-group: seven papers tracing continuous diffusion for categorical data through simplex-based LMs, hybrid discrete–continuous models, and recent flow-map approaches for language.

Discussion points:

Where do continuous/simplex methods offer advantages over masked discrete diffusion?
How do hybrid models (CANDI, augmented discrete diffusion) compare in practice?
What does the diffusion duality imply for choosing a modeling framework?
Which of these lines look most promising to scale up for text generation?