Diffusion Language Models.
A student-led seminar on language models that generate by iterative denoising instead of left-to-right sampling — the diffusion recipe from image generation, applied to discrete text. This page will hold the slide-by-slide study companion once the deck is available.
§0Status
The seminar hasn't been presented yet, so no slides are available. Once the deck lands, this page will get the same slide-by-slide walkthrough treatment as the other seminars — every slide embedded verbatim with a first-principles explanation underneath.
§1What the seminar covers
Standard language models generate autoregressively: one token at a time, left to right, each token conditioned on everything already produced. Diffusion language models do something very different. They start from pure noise (a sequence of random or masked tokens) and, over many denoising steps, gradually refine the whole sequence in parallel until it is a coherent piece of text.
The idea is imported from image diffusion (DDPM, Stable Diffusion), where it revolutionised generation. Adapting it to text is harder because text is discrete, and the recent research line — D3PM, SEDD, MDLM, and more — has explored several ways to handle that. The payoff is parallel decoding (much faster generation than left-to-right sampling), stronger controllability, and iterative refinement of the final output.