Model Architecture & Training MethodsXiang Lisa Li et al. (Stanford)Stanford University

Li et al. present Diffusion-LM

Presented Diffusion-LM — non-autoregressive text generation that iteratively denoises Gaussian vectors into word vectors

Why it matters

The starting point of the diffusion language model lineage

Source — Primary sources first; secondary reports are labelled “Press”

Paperarxiv.orghttps://arxiv.org/abs/2205.14217

Added in Week 35, 2026 · See that week

Same week

Dao et al. present FlashAttentionPaperKusupati et al. present Matryoshka Representation LearningPaperKojima et al. present Large Language Models are Zero-Shot ReasonersPaperJung et al. present Maieutic PromptingPaper

Nearby on the same axis

Emergent Abilities of LLMsPaperGoogle Research announces PaLM (540B)OfficialChinchillaPaperModel SoupsPaper