// ∿ New Architectures: SSM & Linear Attention
Diffusion Language Models (LLaDA)
Generates text by iteratively unmasking a fully masked sequence.
扩散语言模型 aka Masked Diffusion Language Model / MDLM / dLLM 2025 Research frontier core
IN PLAIN WORDS
Train by masking random fractions of tokens; generate by starting fully masked and filling tokens in parallel over steps. LLaDA showed an 8B diffusion LM can rival autoregressive models.
Used in: LLaDA and MDLM research; Inception’s commercial Mercury dLLMs, with Mercury 3 released 19 Oct 2026 (vendor-reported 1M context, 1,000+ tokens/s, no independent tests yet); E-MoE (arXiv 2609.37533) improves few-step decoding
TIMELINE · prerequisites → this → what builds on it
01Original paper
Large Language Diffusion Models
Shen Nie, Fengqi Zhu, Zebin You, et al. · 2025
arXiv ↗ ✓ Verified · arXiv API
02Where it sits
Learn the left column first; the right column builds on this one.
Then you can learn
- Nothing in the atlas builds on it yet
Related:Transformer & Self-Attention
Left: prerequisites · right: builds on this · dashed: related. Full map →
03Learning resources
Official docs, explainers, courses and reference code. All links checked.
Reference code
Official docs
Explainers
04Where the course teaches it
On learning paths:Generative Models & Frontier Research
For AI / data
- Verification: verified, checked 2026-10-22.
- JSON: /api/learn/algorithms.json (id:
diffusion-lm)