Diffusion language models are super cool — but what's next? 🤔

Non-autoregressive LMs have come a long way: they decode in parallel, generate in any order, can revise their own outputs, and they've been scaled up. The field has grown up fast. So I keep asking — What are the most urgent open problems? what's the next paradigm? Where is all of this heading?

I wanted a place where we could actually get in a room and hash that out together. That's what NonAR-LM @ COLM 2026 is: a platform for the whole community — industry and academia — to talk about the future of language generation beyond next-token prediction.

We've brought together researchers from OpenAI, Google DeepMind, Meta, NVIDIA, and ByteDance, with academics from Duke, Stanford, Princeton, UCLA, Harvard, Cornell, HKU, and Oxford. We'd genuinely love for you to be part of it.

⏳ Two weeks to go — submissions close June 23. If you work on language generation beyond next-token prediction — diffusion, flow matching, or any-order autoregression — we'd love to see your work. Up to 8 pages, non-archival, double-blind.

🔗 Website & CFP: https://pengzhangzhi.github.io/NonAR-LM/

📨 Submit on OpenReview: https://openreview.net/group?id=colmweb.org/COLM/2026/Workshop/NonAR-LM

🧑‍⚖️ We're also recruiting reviewers — a light load, perfect for students and early-career researchers who want a close look at the newest work. Sign up: https://forms.gle/K3wfLD6WjuHgXCT78

A huge thanks to the speakers & panelists making this one fun: @StefanoErmon · Shansan Gong · @adityagrover_ · @thjashin · @ArashVahdat · @MengdiWang10 · @ssahoo_ · @aaron_lou

and our wonderful co-organizers: @mariannearr · @Jaeyeon_Kim_0 · @siyan_zhao · @AlexanderTong7 · @ArnaudDoucet1 · @bodonoghue85 · @kb_syx

Help us spread it out, repost appreciated!