Marianne Arriola

We have an amazing lineup of invited speakers joining us in NYC this October, including

@elmelis @mariannearr @Clement_Bonet_ @sitanch @YongxinChen1 @cdomingoenrich @ArthurGretton @yjelid @k_neklyudov @ssahoo_ @SchiffYair @sherryyangML @SoojungYang2

🚨Coming to #ICML2026 🇰🇷

Last year, we introduced Block Diffusion LMs, adopted by dLLMs at Google, NVIDIA, Meta, etc.

🚀Today, we introduce Set Diffusion, a faster, more flexible alternative that combines AR and diffusion by interpolating generation orders.

🧵1/8

I'm at #ICML2026 this week, come talk to me about all things diffusion!

On Wed morning, my student @Guanghan__Wang is presenting d2, improved techniques for RL in diffusion LLMs.

On Thu morning, @mariannearr is presenting Set Diffusion, the next generation of block diffusion supporting any-order generation.

On Fri, I am moderating a panel on diffusion vs. autoregression at the SPIGM workshop with an exciting set of speakers.

Hope to see you in Seoul! Links to @Guanghan__Wang @mariannearr papers are below.

We took a 30B model and split it in two to write tokens in parallel instead of one at a time.

Introducing Nemotron-Labs-TwoTower: a diffusion language model from NVIDIA Research adapted from Nemotron-3-Nano-30B-A3B. Here’s how it works: one half holds the context, the other writes the tokens, with both reusing the pretrained model instead of training a new one from scratch.

We found it kept 98.7% of the original model’s quality at 2.42× faster generation.

New OSS diffusion LLM out of a major lab, this time @nvidia.

Shout-out to Cornell students @mariannearr and @SchiffYair whose work on block diffusion and encoder-decoder diffusion is at the core of this model along with MDLM.

Marianne's and Yair's papers are linked below.

DiffusionGemma is an open, experimental model that brings our text diffusion research to Gemma 4. It’s a racehorse 🏇achieving up to 4x faster inference by generating entire blocks of text simultaneously vs predicting token-by-token (word-by-word) output!

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: pengzhangzhi.github.io/NonAR-LM/

📨 Submit on OpenReview: openreview.net/group?id=colmw…

🧑‍⚖️ 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: forms.gle/K3wfLD6WjuHgXC…

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!

A lot of people think that diffusion LLMs == BERT.

But masked dLLMs are just one form of text diffusion, popularized by @ssahoo_ and his group at Cornell. There's much more to explore in text diffusion land!

Check out our conversation (link below)!

A breakthrough in real-time video generation.

As a research preview developed with @NVIDIA and shared at @NVIDIAGTC this week, we trained a new real-time video model running on Vera Rubin. HD videos generate instantly, with time-to-first-frame under 100ms. Unlocking an entirely new creative paradigm and bolstering the foundations of our General World Model, GWM-1.

Real-time generation opens a fundamentally different design space for video models and world simulation. We're investing in co-designing our models alongside advances in hardware to keep pushing this frontier.

1/ 🚨 New paper! DUEL: Exact Likelihood for Masked Diffusion via Deterministic Unmasking

We give masked diffusion models (MDMs) proper likelihood — and therefore proper perplexity — for the first time. Turns out MDMs are closer to autoregressive models than previously thought.

🚨 In our paper “Learn from Your Mistakes: Self-Correcting Masked Diffusion Models” we tackle one of the most pressing issues for masked diffusion models (MDMs) → How can we enable fast parallel generation + recovery from mistakes

Diffusion clicked for me when I read about score-based models, a line of work pioneered by @StefanoErmon (et al.) at Stanford.

So it was a full-circle moment to collab with him and @_inception_ai on a video about training & sampling techniques for making diffusion LLMs faster.

Breaking the 1000/tok sec barrier at the intelligence level of GPT5-mini, Flash, Haiku 🚀🚀🚀

Do causal video diffusers really need dense causal attention at every layer, every denoising step?

We looked inside and found: no. Causality is separable from denoising.

Here are two surprising observations that hold across architectures, training objectives, and scales.

🚨In our NeurIPS paper, we bring encoder-decoders back.. for diffusion language models!

⚡️Encoder-decoders make diffusion sampling fast: a small (fast) decoder denoises tokens progressively and a large (slower) encoder represents clean context.

Runway Gen-4.5 represents significant advancements in both pre-training data efficiency and post-training techniques for video models and serves as our new foundation model for world modeling.

Gen-4.5 scored 1,247 Elo points in the Artificial Analysis Text to Video leaderboard, setting a new standard and surpassing all other AI video models in the world.

We’re releasing UNI-D², a unified codebase for discrete diffusion language models 🤝🚀

Co-led with @vincentpaulinef and an amazing advisor team: @stefanAbauer, @AlexanderTong7 , @andrea_dittadi, @AMK6610, @KaplFer 🙌

🔗 GitHub: github.com/nkalyanv99/UNI…
📚 Docs: nkalyanv99.github.io/UNI-D2/

Reproduce and extend state-of-the-art baselines with one toolkit. Let’s move beyond autoregressive models and push discrete diffusion together 🧵👇