Stefano Ermon explains why diffusion-based language models are superior for production, offering lower cost per token and higher speed, and shares the origin story of diffusion models.
要点 · TL;DR
扩散语言模型在相似质量下比自回归模型快 5-10 倍且更便宜。 Diffusion LLMs are 5-10x faster and cheaper than autoregressive models at similar quality.
推理时缩放是关键；扩散模型比自回归模型缩放得更好。 Inference-time scaling is key; diffusion models scale better than autoregressive ones.
扩散语言模型在相似质量下，服务速度比自回归模型快5-10倍，成本更低。 Diffusion language models are 5-10x faster and cheaper to serve than autoregressive models at similar quality.
推理时缩放现在是关键焦点，扩散模型在推理时比自回归模型扩展得更好。 Inference-time scaling is now the key focus, and diffusion models scale better at inference than autoregressive models.
扩散模型每步生成多个词元，比逐词元生成的自回归模型更快。 Diffusion models generate multiple tokens per step, enabling faster generation than token-by-token autoregressive models.
扩散模型通过调整去噪步数控制质量，提供了新的推理时算力调节旋钮。 Diffusion models allow controlling quality by adjusting denoising steps, offering a new inference-time compute knob.
Mercury 2 在质量上与速度优化的前沿模型相当，但速度显著更快。 Mercury 2 matches speed-optimized frontier models in quality but is significantly faster.
反共识 · Contrarian takes
扩散模型可以用稳定的去噪目标训练，不像GAN训练那样不稳定。 Diffusion models can be trained with a stable denoising objective, unlike unstable GAN training.
扩散语言模型支持原地纠错，无需延长输出轨迹。 Diffusion language models enable in-place error correction without lengthening the output trace.
扩散模型提供更好的生成可控性，因为完整输出从一开始就可用。 Diffusion models offer better controllability for generation, as the full output is available from the start.
强化学习后训练受益于扩散模型更快的推理，用于生成候选解。 Reinforcement learning post-training benefits from diffusion models' faster inference for rollouts.
扩散语言模型尚未达到最佳前沿模型的质量，仅匹配速度优化层级。 Diffusion language models are not yet at the quality of the best frontier models, only matching speed-optimized tiers.
本期章节 · Chapters（共 22）
引言与背景Introduction and Background
扩散模型的灵感来源Inspiration for Diffusion Models
推理时计算与思维链
从图像扩散模型到文本Inference-time compute and chain of thought
扩展扩散语言模型From diffusion models for images to text
扩散与自回归模型：推理缩放与上下文Scaling diffusion language models
训练扩散语言模型 vs 自回归模型Diffusion vs autoregressive models: inference scaling and context
扩散模型的部署设置Training diffusion language models vs autoregressive models
推理时缩放与旋钮设计Serving setup for diffusion models
质量、速度与成本的权衡Inference-time scaling and knob design
定性差异与扩散模型优势Trade-offs between quality, speed, and cost
局限性：上下文长度与多模态Qualitative differences and diffusion model strengths
扩展扩散语言模型Limitations: context length and multimodality
扩散语言模型中的开放科学问题Scaling up diffusion language models
扩散模型用于可控生成与逆问题Open Science Questions in Diffusion Language Models
大型实验室中的快速智能交互与切换成本Diffusion Models for Controlled Generation and Inverse Problems
其他实验室与学术界的扩散语言模型工作Fast agentic interactions and switching costs in big labs
图像与文本扩散的交叉融合Other labs and academic work on diffusion language models
多模态扩散方法Cross-pollination between image and text diffusion