Elamur Research⌕
☰
home / papers / 2026 / Fast-Decoding Diffusion Language Models via Progress-Aware Confidence Schedules

Fast-Decoding Diffusion Language Models via Progress-Aware Confidence Schedules

Paper · 2026 · Amr Mohamed, Yang Zhang, Michalis Vazirgiannis et al.

Summary

Diffusion large language models (dLLMs) can generate text in parallel and attend in both directions, but they are slow at inference. Decoding runs through many denoising steps, and the step budget is usually fixed in advance and set conservatively. The authors treat decoding as a question of when to stop. Their method, SchED, needs no training and works with any model. At each step it averages top-2 logit margins over the whole answer span. It then compares this average with a smooth threshold that decreases as decoding progresses, using linear, cosine or exponential schedules. Generation stops as soon as the confidence clears the threshold. The method was tested on Dream and LLaDA, in both base and instruction-tuned versions, across ten benchmarks covering multiple-choice, math, long-form QA/summarization and translation. Instruction-tuned models ran about 4× faster with essentially no loss in quality. SchED also beat Prophet on a new quality-penalized speed metric. An entropy analysis suggests that instruction tuning makes predictions stabilize faster, which explains the larger gains for those models.

Key points

  • Speed and quality by model type:
    • Instruction-tuned models ran 3.8–4.0× faster on average while keeping 99.8–100% of baseline quality.
    • Base models kept 99.1–100% of baseline quality, with speedups of up to 2.34× under more aggressive schedules.
  • New metric, Quality–Penalized Speed (QPS, γ=4): it explicitly penalizes accuracy drops.
    • On Dream Instruct, the best SchED variant scored 4.30, versus 2.91 for Prophet.
    • On Dream Base, the best variant scored 2.03, versus 1.07 for Prophet.
  • Design choices that stabilize stopping: averaging confidence over the full answer span prevents early exits triggered by isolated confident tokens. The threshold schedule is controlled by τ_high, τ_low and k; example settings are (7.5, 2.5) and (7.5, 0).
  • Early exit can improve answers: stopping sooner sometimes avoids errors that extra refinement steps would introduce, such as numerical slips.
  • Availability: the code is at github.com/amr-mohamedd/SchED. The paper appeared in Findings of ACL 2026.

Our summary of the paper, in our own words — not its text. The work belongs to its authors; read it at the links above.

 Institutions

🎓École PolytechniquePalaiseau · France14 works · 898 citations

 Fields, topics and concepts

📚 Shelves 4The field of science this work belongs to (OpenAlex fields).Computer Science · 16,667 worksArtificial Intelligence · 16,194 worksArts and Humanities · 16,039 worksLanguage and Linguistics · 15,891 works

🧭 Topics 1Research subjects and languages the work is about.Language models & pretraining · 3,320 works

Cite

Cite as Amr Mohamed, Yang Zhang, Michalis Vazirgiannis and Guokan Shang. 2026. Fast-Decoding Diffusion Language Models via Progress-Aware Confidence Schedules. Findings of the Association for Computational Linguistics: ACL 2026.

The paper itself, not this page. BibTeX from the library's record of the work. DOI 10.18653/v1/2026.findings-acl.1782

@inproceedings{mohamed2026fast,
  title = {{Fast-Decoding Diffusion Language Models via Progress-Aware Confidence Schedules}},
  author = {Amr Mohamed, Yang Zhang, Michalis Vazirgiannis and Guokan Shang},
  booktitle = {Findings of the Association for Computational Linguistics: ACL 2026},
  year = {2026},
  doi = {10.18653/v1/2026.findings-acl.1782},
  url = {https://doi.org/10.18653/v1/2026.findings-acl.1782}
}

References

40 references, 14 of them with a card of ours. ↗ leads to the work, «find» searches for one we have no id for yet.

  1. Yushi Bai, Xin Lv, Jiajie Jin, Yidong Zhang, Hongbo Zhang, Beichen Zheng, Yikai Chen, Jian-Guang Qian, Zeyu Chen, Wen-Han Liu, and 1 others. 2023. LongBench: A bilingual, multitask benchmark for long context understanding. arXiv preprint arXiv:2308.14508.
  2. Yonatan Bisk, Rowan Zellers, Ronan Lebras, Chandra Bhagavatula, and Yejin Choi. 2020. PIQA: Reasoning about physical commonsense in natural language. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 34, pages 7432–7439.
  3. Ondřej Bojar, Christian Buck, Christian Federmann, Barry Haddow, Philipp Koehn, Johannes Leveling, Christof Monz, Pavel Pecina, Matt Post, Hervé Saint-Amand, and 1 others. 2014. Findings of the 2014 workshop on statistical machine translation. In Proceedings of the Ninth Workshop on Statistical Mac…
  4. Ondřej Bojar, Rajen Chatterjee, Christian Federmann, Yvette Graham, Barry Haddow, Matthias Huck, Antonio Jimeno Yepes, Philipp Koehn, Varvara Logacheva, Christof Monz, and 1 others. 2016. Findings of the 2016 conference on machine translation (wmt16). In Proceedings of the First Conference on Machi…
  5. Tom Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared D Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, and 1 others. 2020. Language models are few-shot learners. Advances in neural information processing systems, 33:1877–1901.
  6. Tianqi Chen, Shujian Zhang, and Mingyuan Zhou. 2025. DLM-one: Diffusion language models for one-step sequence generation. arXiv [cs.CL]. find
  7. Karl Cobbe, Vineet Kosaraju, Mohammad Bavarian, Mark Chen, Heewoo Jun, Lukasz Kaiser, Matthias Plappert, Jerry Tworek, Jacob Hilton, Reiichiro Nakano, and 1 others. 2021. Training verifiers to solve math word problems. arXiv preprint arXiv:2110.14168.
  8. Alexander R Fabbri, Greg Durrett, Daniel Farcas, and Viv Ng. 2019. Multi-News: A large-scale multi-document summarization dataset and abstractive hierarchical model. arXiv preprint arXiv:1906.01749.
  9. Leo Gao, Jonathan Tow, Baber Abbasi, Stella Biderman, Sid Black, Anthony DiPofi, Charles Foster, Laurence Golding, Jeffrey Hsu, Alain Le Noac’h, Haonan Li, Kyle McDonell, Niklas Muennighoff, Chris Ociepa, Jason Phang, Laria Reynolds, Hailey Schoelkopf, Aviya Skowron, Lintang Sutawika, and 5 others.… find
  10. Chen Gong, Hantian Li, Denny Zhou, and Quoc V Le. 2024a. Beyond autoregression: Discrete diffusion for complex reasoning and planning. arXiv preprint arXiv:2410.14157. — [arXiv:2410.14157](https://arxiv.org/abs/2410.14157) find
  11. Shansan Gong, Shivam Agarwal, Yizhe Zhang, Jiacheng Ye, Lin Zheng, Mukai Li, Chenxin An, Peilin Zhao, Wei Bi, Jiawei Han, Hao Peng, and Lingpeng Kong. 2024b. Scaling diffusion language models via adaptation from autoregressive models. arXiv [cs.CL]. find
  12. Dan Hendrycks, Collin Burns, Steven Basart, Andrew Critch, Jerry Li, Dawn Ippolito, Jared Kaplan, Jana Scheurer, Sandeep Datta, Cameron Foote, and 1 others. 2020. Measuring massive multitask language understanding. arXiv preprint arXiv:2009.03300.
  13. Jun-Jie Huang, Zixuan Li, Han-Ning Wu, Zixun Lan, Weizhen Qi, Hangbo Bao, Zewen Chi, Wen-Hong Li, Jun-Cheng Chen, Li-Yuan Wang, Wei-Nan Zhang, and Ting Liu. 2025. Llada-moe: A sparse moe diffusion language model. arXiv preprint arXiv:2509.24389. — [arXiv:2509.24389](https://arxiv.org/abs/2509.24389) find
  14. Xiang Lisa Li, Jack W Rae, Yiding Jiang, Yixiao Li, Hui-Ling Zhen, Grzegorz Górny, Piotr Nawrot, Tim GJ Rudner, Jean-Baptiste Lespiau, Jonathan Richard Schwarz, Arthur Mensch, Roman Ring, Lorenzo Noci, Trevor Cai, Sebastian Borgeaud, Jeff Stanway, Charlie Chen, David Ding, Daniel Autobiography, and… — [arXiv:2410.18514](https://arxiv.org/abs/2410.18514) find
  15. Chin-Yew Lin. 2004. ROUGE: A package for automatic evaluation of summaries. In Text Summarization Branches Out, pages 74–81, Barcelona, Spain. Association for Computational Linguistics.
  16. Clara Meister, Ayan Pal, Emma Strubell, Mike Rabbat, and Samar Singh. 2024. Fs-dfm: Fast and accurate long text generation with few-step diffusion language models. arXiv preprint arXiv:2509.20624. — [arXiv:2509.20624](https://arxiv.org/abs/2509.20624) find
  17. Shen Nie, Fengqi Zhu, Zebin You, Xiaolu Zhang, Jingyang Ou, Jun Hu, Jun Zhou, Yankai Lin, Ji-Rong Wen, and Chongxuan Li. 2025. Large language diffusion models. arXiv [cs.CL].
  18. Friedrich Or-El, Zheng-Hao Liu, Chen-Lin Zhang, Hong-Yi Yuan, and Wei-Qiang Zhang. 2025. Accelerating diffusion llm inference via local determinism propagation. arXiv preprint arXiv:2510.07081. — [arXiv:2510.07081](https://arxiv.org/abs/2510.07081) find
  19. Ayan Pal, Meet Shah, Emma Strubell, Berkin Gunel, Clara Meister, Ameet Pokle, Yilun Zhou, Mike Rabbat, and Samar Singh. 2025. Mercury: Ultra-fast language models based on diffusion. arXiv preprint arXiv:2506.17298. find
  20. Li Pengxiang, Zhou Yefan, Muhtar Dilxat, Yin Lu, Yan Shilin, Shen Li, Liang Yi, Vosoughi Soroush, and Liu Shiwei. 2025. Diffusion language models know the answer before decoding. arXiv [cs.CL]. find
  21. Mary Phuong, Chen Gong, Hantian Li, Denny Zhou, and Quoc V Le. 2025. d1: Scaling reasoning in diffusion large language models via reinforcement learning. arXiv preprint arXiv:2504.12216. — [arXiv:2504.12216](https://arxiv.org/abs/2504.12216) find
  22. David Rein, Stas Raichuk, Orion Selesnick, Armel Fotso, Jarno Textor, Daniel Lindner, Yafang Du, Tsvi August, Dan Gutman, Mor Geva, and 1 others. 2023. GPQA: A graduate-level google-proof Q&A benchmark. arXiv preprint arXiv:2311.12022.
  23. Subham Sekhar Sahoo, Justin Deschenaux, Aaron Gokaslan, Guanghan Wang, Justin Chiu, and Volodymyr Kuleshov. 2025. The diffusion duality. arXiv [cs.LG]. find
  24. Keisuke Sakaguchi, Ronan Le Bras, Chandra Bhagavatula, and Yejin Choi. 2020. WinoGrande: An adversarial Winograd schema challenge at scale. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 34, pages 8732–8740. find
  25. Nimit S Sohoni, Chien-Yu Lin, Aniruddha Nrusimha, and Christopher Ré. 2025. Spiffy: Multiplying diffusion llm acceleration via lossless speculative decoding. arXiv preprint arXiv:2509.18085. — [arXiv:2509.18085](https://arxiv.org/abs/2509.18085) find
  26. Yuxuan Song, Zheng Zhang, Cheng Luo, Pengyang Gao, Fan Xia, Hao Luo, Zheng Li, Yuehang Yang, Hongli Yu, Xingwei Qu, Yuwei Fu, Jing Su, Ge Zhang, Wenhao Huang, Mingxuan Wang, Lin Yan, Xiaoying Jia, Jingjing Liu, Wei-Ying Ma, and 3 others. 2025. Seed diffusion: A large-scale diffusion language model w…
  27. Yixuan Su, Cheng Zhu, Wangchunshu Zhou, Jing Chen, and Graham Neubig. 2025. Don’t settle too early: Self-reflective remasking for diffusion language models. arXiv preprint arXiv:2509.23653. — [arXiv:2509.23653](https://arxiv.org/abs/2509.23653) find
  28. Chuhan Wang, Hong-Ning Dai, Yi-Chen Zhang, Chun-Wei Tsung, Xin-Hao Li, Wei-Guo Zheng, and Ming-Li Song. 2025a. d 2 cache: Accelerating diffusion-based llms via dual adaptive caching. arXiv preprint arXiv:2509.23094. — [arXiv:2509.23094](https://arxiv.org/abs/2509.23094) find
  29. Chuhan Wang, Yi-Chen Zhang, Xin-Hao Li, Hong-Ning Dai, and Ming-Li Song. 2025b. Creditdecoding: Accelerating parallel decoding in diffusion large language models with trace credits. arXiv preprint arXiv:2510.06133. — [arXiv:2510.06133](https://arxiv.org/abs/2510.06133) find
  30. Ming-Hao Wang, Yun-Zhu Song, Jun-Jie Huang, Zixuan Li, Wen-Hong Li, Jia-Jun Zhang, and Cheng-Chung Hsu. 2025c. Ultrallada: Scaling the context length to 128k for diffusion large language models. arXiv preprint arXiv:2510.10481. — [arXiv:2510.10481](https://arxiv.org/abs/2510.10481) find
  31. Qingyan Wei, Yaojie Zhang, Zhiyuan Liu, Dongrui Liu, and Linfeng Zhang. 2025. Accelerating diffusion large language models with SlowFast sampling: The three golden principles. arXiv [cs.CL]. find
  32. Zhilin Yang, Peng Qi, Saizheng Zhang, Yoshua Bengio, William W Cohen, Ruslan Salakhutdinov, and Christopher D Manning. 2018. HotpotQA: A dataset for diverse, explainable multi-hop question answering. In Proceedings of the 2018 Conference on Empirical Methods in Natural Language Processing, pages 236…
  33. Jiacheng Ye, Zhihui Xie, Lin Zheng, Jiahui Gao, Zirui Wu, Xin Jiang, Zhenguo Li, and Lingpeng Kong. 2025. Dream 7B: Diffusion large language models. arXiv [cs.CL]. find
  34. Shukang Yin, Chaoyou Fu, Sirui Zhao, Ke Li, Xing Sun, Tong Xu, and Enhong Chen. 2024. A survey on multimodal large language models. National Science Review, 11(12):nwae403.
  35. Hong-Yi Yuan, Zheng-Hao Liu, Chen-Lin Zhang, and Wei-Qiang Zhang. 2025a. Self speculative decoding for diffusion large language models. arXiv preprint arXiv:2510.04147. — [arXiv:2510.04147](https://arxiv.org/abs/2510.04147) find
  36. Hongyi Yuan, Zheng-Hao Liu, Chen-Lin Zhang, Xu-Ran Wang, Yu-Hong-Vey Wong, and Jia-Wei Liu. 2025b. dkv-cache: The cache for diffusion language models. arXiv preprint arXiv:2505.15781. — [arXiv:2505.15781](https://arxiv.org/abs/2505.15781) find
  37. Rowan Zellers, Ari Holtzman, Yonatan Bisk, Ali Farhadi, and Yejin Choi. 2019. HellaSwag: Can a machine really finish your sentence? In Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics, pages 4791–4800. find
  38. Chen-Lin Zhang, Zheng-Hao Liu, Hong-Yi Yuan, and Wei-Qiang Zhang. 2025a. Finish first, perfect later: Test-time token-level cross-validation for diffusion large language models. arXiv preprint arXiv:2510.05090. — [arXiv:2510.05090](https://arxiv.org/abs/2510.05090) find
  39. Lingzhe Zhang, Liancheng Fang, Chiming Duan, Minghua He, Leyi Pan, Pei Xiao, Shiyu Huang, Yunpeng Zhai, Xuming Hu, Philip S Yu, and 1 others. 2025b. A survey on parallel text generation: From parallel decoding to diffusion language models. arXiv preprint arXiv:2508.08712. — [arXiv:2508.08712](https://arxiv.org/abs/2508.08712) find
  40. Hao Zou, Zae Myung Kim, and Dongyeop Kang. 2023. A survey of diffusion models in natural language processing. arXiv [cs.CL]. find

Updated

↑ back to top

This page in Markdown → https://md.research.elamur.ai/free/en/papers/mohamed-2026-fast-decoding-diffusion-language-models/