Research map / Image generation / Fast diffusion sampling
Masked diffusion language model decoding
Accelerate and improve parallel decoding of masked diffusion language models.
147 accepted papers at ICML, NeurIPS, ICLR, CVPR and AAAI (2022–2026) · an approach in Fast diffusion sampling, keyword cluster “discrete · diffusion language · masked”
State History Tried Checklist Preprints Who Export
Where it stands
Not enough benchmark data Fewer than four papers report the same benchmark, so there is no record to follow.
4 papers in 2023–24 and 142 in 2025–26 (at ICML, NeurIPS, ICLR, CVPR and AAAI); its share of the topic grew, from 0.9% to 10.7%.
How this is decided
A benchmark is tracked when at least four papers of this approach report their own method on the same dataset and metric in at least two different years. Its record is the best number reported up to 2024 compared with the best reported in 2025–26: still improving means at least 1% better, flat means at most 0.2% better with at least two papers from 2025–26 trying. Numbers on a different scale from the rest (a fraction among percentages) and far outliers are left out. Results come from different papers with different setups, so they are signals, not head-to-head comparisons. The share of papers is counted at ICML, NeurIPS, ICLR, CVPR and AAAI.
How it got here
Most cited paper of each year
- 2022 · Step-unrolled Denoising Autoencoders for Text Generation (ICLR 2022 · 28 citations)
- 2023 · Likelihood-Based Diffusion Language Models (NeurIPS 2023 · 5 citations)
- 2024 · Diffusion Forcing: Next-token Prediction Meets Full-Sequence Diffusion (NeurIPS 2024 · 47 citations)
- 2025 · PepTune: De Novo Generation of Therapeutic Peptides with Multi-Objective-Guided Discrete Diffusion (ICML 2025 · 10 citations)
- 2026 · MDiff4STR: Mask Diffusion Model for Scene Text Recognition (AAAI 2026 · 1 citation)
Earliest papers here
- Step-unrolled Denoising Autoencoders for Text Generation (ICLR 2022 · 28 citations)
- Likelihood-Based Diffusion Language Models (NeurIPS 2023 · 5 citations)
- Tailoring Language Generation Models under Total Variation Distance (ICLR 2023 · 2 citations)
What has been tried
27 papers here report 300 measured changes (a component added, removed or swapped, compared with a baseline): helped 73%, no clear effect 5%, hurt 20%, mixed 3%. Papers mostly publish changes that helped, so "helped" is over-represented.
Ablations in recent papers
| What was changed | Measured on | Result | Effect | Paper |
|---|---|---|---|---|
| add Monte Carlo Guidance | Maze2D Single-task Average · reward | 129.67 -> 141.7 | Helped | Diffusion Forcing: Next-token Prediction Meets Full-Sequence Diffusion (NeurIPS 2024 · 47 citations) |
| Add attention dropout (25–75%) to AR training | Validation · validation loss | No clear effect | Diffusion Beats Autoregressive in Data-Constrained Settings (NeurIPS 2025 · 1 citation) | |
| Use CEDD* instead of CEDD | OpenWebText, L=128 · generative perplexity | 148.21 -> 143.86 | Helped | Efficient Perplexity Bound and Ratio Matching in Discrete Diffusion Language Models (ICLR 2025 · 1 citation) |
| Add unsupervised CFG without paired data | ARC-e, BoolQ, HellaSwag, Obqa, PIQA, RACE, SIQA, and LAMBADA · accuracy | 37.42 -> 39.02; 61.50 -> 62.17; 33.46 -> 34.10; 27.00 -> 34.20; 60.34 -> 60.39; 29.28 -> 30.81; 36.95 -> 37.41; 36.00 -> | Helped | Scaling up Masked Diffusion Models on Text (ICLR 2025 · 1 citation) |
| Add the shift operation during discrete diffusion adaptation | GSM8K-symbolic · accuracy | 34.5 -> 49.7 | Helped | Scaling Diffusion Language Models via Adaptation from Autoregressive Models (ICLR 2025 · 1 citation) |
| use carry-over autoregressive energy parameterization | OpenWebText test · PPL | 20.49 -> 17.58 | Helped | Energy-Based Diffusion Language Models for Text Generation (ICLR 2025 · 1 citation) |
| Increase inference budget from the default to 127 steps | Overall average score · average score | 35.50 -> 36.24 | Helped | Large Language Models to Diffusion Finetuning (ICML 2025 · 0 citations) |
| add curriculum learning | OWT zero-shot · PPL | 35.0 -> 33.7 | Helped | The Diffusion Duality (ICML 2025 · 0 citations) |
| add unconditional pre-training (model a to model c) | AudioCaps evaluation set · FAD | 1.35 -> 1.13 | Helped | IMPACT: Iterative Mask-based Parallel Decoding for Text-to-Audio Generation with Diffusion Modeling (ICML 2025 · 0 citations) |
| increase τ from 0.5 to 2.0 | CIFAR-10 · NFE | 42.44 -> 35.40 | Helped | Accelerated Diffusion Models via Speculative Sampling (ICML 2025 · 0 citations) |
| add uniform noise at p_u=0.1 or p_u=0.2 | OWT validation · PPL | 23.05 -> 23.67; 23.05 -> 24.38 | Hurt | Generalized Interpolating Discrete Diffusion (ICML 2025 · 0 citations) |
| Reduce block size from 32 to 2 | GSM8K · 0-shot pass@1 | 20.9 -> 50.1 | Helped | Encoder-Decoder Diffusion Language Models for Efficient Training and Inference (NeurIPS 2025 · 0 citations) |
What reviewers will expect
Datasets and baselines named in the 71 papers of this approach from 2025–2026 that have a full-text summary. Leaving out the common ones is a frequent review question.
Datasets
- GSM8K · 23 of 71 papers
- MBPP · 18 of 71 papers
- HumanEval · 16 of 71 papers
- OpenWebText · 14 of 71 papers
- MATH500 · 9 of 71 papers
- LM1B · 7 of 71 papers
- MATH · 7 of 71 papers
- LAMBADA · 4 of 71 papers
Baselines
- MDLM · 13 of 71 papers
- SEDD · 9 of 71 papers
- D3PM · 6 of 71 papers
- UDLM · 4 of 71 papers
- GPT2 · 4 of 71 papers
- AR · 4 of 71 papers
- SSD-LM · 4 of 71 papers
- Plaid · 3 of 71 papers
Recent preprints
Loading recent arXiv preprints…
Who publishes on it
Institutions on the most papers of this approach since 2023 (of 146 papers; from OpenAlex and the papers' affiliations).
- Tsinghua University · 4 papers
- Alibaba Group · 4 papers
- Korea Advanced Institute of Science and Technology · 4 papers
- Cornell University · 3 papers
- University College London · 3 papers
- ETH Zurich · 3 papers
- Stanford University · 3 papers
- Carnegie Mellon University · 3 papers
- The Chinese University of Hong Kong · 3 papers
- Nanyang Technological University · 3 papers
Use it
Download BibTeX of 23 key papersOpen the topic in the interactive map
Get alerts. Describe your idea in scime atlas and save it (free scime account): you get a weekly email when new preprints land close to it.
Related approaches
Same keyword cluster
- Discrete diffusion solvers (64 papers)
Elsewhere in the topic
- Training-free diffusion acceleration (257 papers)
- Diffusion theory and design space (252 papers)
- Diffusion models as perception tools (198 papers)
- Guidance for conditional diffusion (123 papers)
- Diffusion theory and learned samplers (121 papers)
- Fast diffusion ODE solvers (120 papers)
- Distributional control of diffusion (109 papers)
- Denoising diffusion probabilistic models (97 papers)
Benchmark numbers, ablations and baseline numbers are extracted by AI from each paper's full text and can be wrong; the paper is the authority. Data: results index of 2026-11-02, citation counts from OpenAlex (2026-10-09).