Extended Abstracts
Up to 4 pages of main text; suitable for work-in-progress, preliminary results, and position papers.
NLP/CORE Unranked
DiffuLM 2026
Diffusion Language Models: Foundations, Efficiency, and Reasoning
DiffuLM @ NeurIPS 2026 is the first workshop dedicated to diffusion language models, focusing on their foundations, efficiency, and reasoning capabilities. It aims to establish diffusion-based approaches as a rigorous and practical alternative to autoregressive language models through theoretical, algorithmic, empirical, and systems research.
Paper fit
Contribution paths
A strong submission should clearly identify its contribution and evaluate it appropriately.
Short Papers
Up to 8 pages of main text; more complete contributions with experiments or theoretical results.
Research areas in scope
01
Foundations
Categorical, masked, and discrete diffusionTraining objectives and probabilistic inferenceScaling laws and model architecturesHybrid autoregressive–diffusion methods
02
Efficiency
Sampling and parallel decodingAccelerated and adaptive inferenceLong-context generation and servingEvaluation and hardware-aware systems
03
Reasoning
Reasoning, planning, and code generationTest-time scaling and self-correctionControllable generation and post-trainingReinforcement learning and alignment
Policies worth checking twice
- Submissions are non-archival; work already published at NeurIPS 2026 or another archival venue is not eligible.
- Review is double-blind and conducted via OpenReview.
- Each submission receives at least three reviews.
- Explicit conflict of interest (COI) declarations are required on OpenReview.
- Organizers must recuse themselves from reviewing submissions from their own institution or active collaborators.
- Papers must be formatted using the NeurIPS 2026 style template.
Official sources
Compiled from the official call for papers. The organizers’ pages remain authoritative.
Last verified September 9, 2026