Latest Sessions
View All Sessions
S30 | DFlash: Block Diffusion for Flash Speculative Decoding
Jian Chen presents DFlash, a speculative decoding framework that replaces autoregressive drafting with a lightweight block diffusion model, so all draft tokens come out of a single forward pass. Conditioning the draft model on context features extracted from the target model keeps draft quality and acceptance rates high. DFlash reaches over 6x lossless acceleration across a range of models and tasks, up to 2.5x higher speedup than EAGLE-3.
S29 | Unlocking Lossless Speedups in LLMs via Discrete Diffusion
Uno is a diffusion-augmented LLM that defines an autoregressive model distribution and uses diffusion to draw several tokens in parallel from that same distribution, so it accelerates generation without sacrificing the quality of the underlying AR model and needs no separate draft model. Uno beats leading speculative-decoding methods in throughput at every batch size and reaches up to 3x speedups over the base AR model, and the 8B Uno outperforms the 26B DiffusionGemma and the proprietary Mercury 2 on agentic tool use, coding, and long-context reasoning.
S28 | Connecting Looped Models and Masked Diffusion Models
A joint session on three concurrent papers that reuse the same transformer weights several times within a single denoising step, adding recursive depth as a third scaling axis for masked diffusion beyond parameter count and denoising steps. Alba Carballo-Castro (EPFL) presents Recursive Scaling in Masked Diffusion Models, Andrea Miele (EPFL) presents Fixed-Point Masked Generative Modeling, and Sanghyun Lee presents Looped Diffusion Language Models.
Featured Videos
View All Videos
How did diffusion LLMs get so fast?
Techniques for accelerating diffusion LLMs, from self-distillation and curriculum learning to KV caching and block diffusion
But How Do Diffusion Language Models Actually Work?
Jia-Bin Huang explores several ideas for applying diffusion models to language modeling
Simple Diffusion Language Models
Quick introduction to Masked Diffusion Language Models (MDLM) by Alexander Rush
About the Reading Group
Diffusion LLMs are faster, more controllable successors to traditional LLMs and are rapidly gaining adoption. This reading group builds a community for exchanging and debating emerging ideas in this space. While our primary focus is discrete diffusion models for language, we also welcome work on other modalities and applications, such as molecular design, drug discovery, and beyond.
Meet the Organizers
Subham Sahoo
Holds a Ph.D. from Cornell Tech, where he specialized in Diffusion Language Models. He has made foundational contributions to the field, with his work deployed at scale by Google, NVIDIA, and ByteDance across language generation and drug discovery.
Justin Deschenaux
PhD student in Machine Learning at EPFL, advised by Prof. Caglar Gulcehre. Previously interned at Apple MLR. His research interests include diffusion language models, fast generative models, and generalization.
