NeurIPS 2026 Workshop

DiffuLM

Diffusion Language Models:
Foundations, Efficiency, and Reasoning

December 11–12, 2026 Sydney, Australia Full-day program
Call for Papers Invited Speakers

About

A dedicated venue for diffusion language models

Language models have been autoregressive by default. Diffusion language models take a different route: they corrupt sequences with noise and learn to reverse it, producing text through iterative refinement rather than one token at a time.

That shift changes what a language model can do: tokens are decoded in parallel, context flows in both directions, and generation becomes far easier to steer and constrain. Over 2025 and 2026 the idea also stopped being theoretical. Inception Labs shipped Mercury, the first commercial-scale diffusion language model; LLaDA and Dream arrived from academic labs; and SEDD, MDLM, FS-DFM, and LaViDa showed that diffusion can match or exceed autoregressive baselines while decoding substantially faster.

Progress has been quick, but the foundations remain fragmented across competing formulations, and the practical questions of efficiency and reasoning are still open. This full-day workshop brings researchers from academia and industry together to consolidate the theory, confront the systems challenges, and examine what diffusion genuinely offers for reasoning.

01

Theory and Foundations

Masked, uniform, and continuous formulations remain only loosely connected. What is the right unifying formalism, which scaling laws govern training, and which architectural choices actually matter?

02

Efficiency

Parallel decoding is the central promise, yet practical models still need many denoising steps. How far does that parallelism really extend, and what serving infrastructure does it demand?

03

Reasoning

Refining a whole solution trace may avoid the cascading errors of left-to-right decoding. Where does that help, where does the lack of sequential structure hurt, and can the diversity gains be measured?

Topics of Interest

We welcome contributions across the following areas, and related work beyond them.

Categorical and masked diffusion formulations
Training objectives and probabilistic inference
Scaling laws and training efficiency
Sampling, parallel decoding and fast inference
Hybrid autoregressive and diffusion modeling
Flow matching and continuous diffusion
Reasoning, planning and code generation
Controllable generation and post-training
Reinforcement learning and alignment
Multimodal diffusion language models
Structured and biological sequence generation
Serving systems and hardware-aware design

Timeline

Key Dates

All deadlines are Anywhere on Earth.

Program

Invited Speakers

Each invited talk is a 30-minute presentation followed by a 5-minute discussion.

Stefano Ermon

Stefano Ermon

Associate Professor, Stanford University

Score-based diffusion and probabilistic inference

Subham Sahoo

Subham Sahoo

MBZUAI Institute of Foundation Models

Foundations and scaling of diffusion language models

Arash Vahdat

Arash Vahdat

Research Director, NVIDIA Research

Scalable generative modeling and generative AI for science

Volodymyr Kuleshov

Volodymyr Kuleshov

Assistant Professor at Cornell, Co-Founder & Chief Scientist at Inception Labs

Foundations of discrete diffusion: masked, uniform, block diffusion, inference-time scaling, guidance

Itai Gat

Itai Gat

Research Scientist, Meta Superintelligence Labs

Discrete flow matching and accelerated sampling

Aditya Grover

Aditya Grover

Assistant Professor at UCLA, CTO at Inception Labs

Commercial-scale and multimodal diffusion language models

Program

Schedule

A full day inside the NeurIPS 08:00–17:00 window, with six invited talks, two poster sessions, four oral presentations of accepted papers, and a panel discussion.

Morning Session
08:15–08:25Opening Remarks
08:25–09:00Invited Talk 1
09:00–09:35Invited Talk 2
09:35–10:35Poster Session I + Coffee Break
10:35–11:10Invited Talk 3
11:10–11:30Oral Talks
Afternoon Session
11:30–13:00Lunch Break
13:00–13:35Invited Talk 4
13:35–14:10Invited Talk 5
14:10–15:10Poster Session II + Coffee Break
15:10–15:45Invited Talk 6
15:45–16:05Oral Talks
16:05–16:50Panel Discussion
16:50–17:00Closing Remarks + Awards

Submissions

Call for Papers

We invite work advancing diffusion language models and related non-autoregressive methods

Submissions may span theory, algorithms, and systems: categorical and masked diffusion; training objectives and their connections to probabilistic inference; sampling, parallel decoding, and efficient inference; reasoning, code generation, and planning; controllable generation and post-training; reinforcement learning and alignment; and hybrid autoregressive and diffusion approaches.

All submissions are non-archival and reviewed double-blind on OpenReview. Work already published at NeurIPS 2026 or another archival venue is not eligible.

Short Papers

Up to 8 pages of main text, with unlimited references and an optional appendix. More complete contributions with experiments or theory.

Every submission receives at least three reviews. We will award a Best Paper and a Best Student Paper prize. Organizers recuse themselves from submissions involving their own institution or active collaborators.

Submission link coming soon

Committee

Organizers

Listed alphabetically, with complementary expertise across the foundations, efficiency, reasoning, and applications of diffusion language models.

Irina Belousova

Irina Belousova

ML Engineering Manager, Apple

LinkedIn ↗
Amin Karimi Monsefi

Amin Karimi Monsefi

Ph.D. Candidate, Ohio State University

Website ↗
Pavlo Molchanov

Pavlo Molchanov

Research Director, NVIDIA Research

Website ↗
Jinjie Ni

Jinjie Ni

Senior Research Scientist, Google DeepMind

Website ↗
MS

Michael Qizhe Shieh

Assistant Professor, NUS

Website ↗

Program Committee

We are actively recruiting reviewers. If you work on diffusion language models or a neighbouring area of generative modeling, we would be glad to have you on the committee.

Sign up as a reviewer