Continuous Diffusion Language Models
Technique · Text & code
Survey of continuous diffusion LMs, which denoise token embeddings; 2026 resurgence driven by flow maps (LangFlow, Flow Map Language Models, RePlaid)
StatusResearch
Evidence
- Aug 30Argues the key advantage is distillability: flow-map distillation gives few-step sampling, while discrete models hit a conditional-independence limit
- Aug 30Notes no clear recipe yet, weak evaluation across the field, and that generative perplexity can be gamed; Plaid-1B was ~64x less training-efficient than autoregressive