mlsanityData to 7 Oct 2026

S2D2: Fast Decoding for Diffusion LLMs via Training-Free Self-Speculation

arXiv 2603.25702 ↗First posted 26 Mar 2026cs.CLcited by 5

On the map: Diffusion models → Discrete and text diffusion

Abstract (arXiv)

Block-diffusion language models offer a promising path toward faster-than-autoregressive generation by combining block-wise autoregressive decoding with within-block parallel denoising. However, in the few-step regime needed for practical acceleration, standard confidence-thresholded decoding is often brittle: aggressive thresholds hurt quality, while conservative thresholds require unnecessary denoising steps. Existing approaches that address this issue either require additional training or incur extra test-time compute. We present S2D2, a training-free self-speculative decoding framework for block-diffusion language models. Our key observation is that a block-diffusion model becomes autoregressive when the block size is reduced to one, allowing the same pretrained model to act as both drafter and verifier. S2D2 inserts a speculative verification step into standard block-diffusion decoding and uses lightweight routing policies to decide when verification is worth its cost. This yields a hybrid decoding trajectory in which diffusion proposes tokens in parallel, while the autoregressive mode acts as a local sequence-level critic. Across three mainstream block-diffusion families, S2D2 consistently improves the accuracy-speed tradeoff over strong confidence-thresholding baselines. On SDAR, we observe up to 4.7× speedup over autoregressive decoding, and up to 1.57× over a tuned dynamic decoding baseline while improving accuracy by up to 4.5 points. On LLaDA2.1-Mini, S2D2 remains complementary to built-in self-correction, including a conservative setting where it is 4.4× faster than the static baseline with slightly higher accuracy.

Key facts. Cited by 5 papers in the map (data through 7 Oct 2026). Placed in Discrete and text diffusion, one of 14 directions in Diffusion models. Builds on 42 of its 51 references that are in the map. First posted on arXiv 26 Mar 2026.

Builds on

42 of 51 references.

Closest ancestors · 16

Same direction, newest first

  1. 2026LLaDA2.1: Speeding Up Text Diffusion via Token Editing9 Feb
  2. 2025ReFusion: A Diffusion Large Language Model with Parallel Autoregressive Decoding15 Dec
  3. 2025LLaDA2.0: Scaling Up Diffusion Language Models to 100B10 Dec
  4. 2025Variational Masked Diffusion Models27 Oct
  5. 2025SDAR: A Synergistic Diffusion-AutoRegression Paradigm for Scalable Sequence Generation7 Oct
  6. 2025Dream 7B: Diffusion Large Language Models21 Aug
  7. 2025Diffusion LLMs Can Do Faster-Than-AR Inference via Discrete Diffusion Forcing8 Aug
  8. 2025DiffuCoder: Understanding and Improving Masked Diffusion Models for Code Generation25 Jun
  9. 2025Block Diffusion: Interpolating Between Autoregressive and Diffusion Language Models12 Mar
  10. 2025Large Language Diffusion Models14 Feb
  11. 2024Energy-Based Diffusion Language Models for Text Generation28 Oct
  12. 2024Scaling Diffusion Language Models via Adaptation from Autoregressive Models23 Oct
4 more
  1. 2024Simple and Effective Masked Diffusion Language Models11 Jun
  2. 2024Simplified and Generalized Masked Diffusion for Discrete Data6 Jun
  3. 2023Discrete Diffusion Modeling by Estimating the Ratios of the Data Distribution25 Oct
  4. 2021Autoregressive Diffusion Models5 Oct

Other references · 26

  1. 2017Attention Is All You Needcited by 59,932
  2. 2020Language Models are Few-Shot Learnerscited by 32,199
  3. 2023LLaMA: Open and Efficient Foundation Language Modelscited by 13,958
  4. 2022Chain-of-Thought Prompting Elicits Reasoning in Large Language Modelscited by 12,302
  5. 2021Training Verifiers to Solve Math Word Problemscited by 9,743
  6. 2021Evaluating Large Language Models Trained on Codecited by 8,053
  7. 2022Large Language Models are Zero-Shot Reasonerscited by 4,323
  8. 2021Program Synthesis with Large Language Modelscited by 3,509
  9. 2019XLNet: Generalized Autoregressive Pretraining for Language Understandingcited by 2,801
  10. 2021Structured Denoising Diffusion Models in Discrete State-Spacescited by 1,665
  11. 2022Fast Inference from Transformers via Speculative Decodingcited by 1,615
  12. 2022Diffusion Posterior Sampling for General Noisy Inverse Problemscited by 1,375
14 more
  1. 2023Instruction-Following Evaluation for Large Language Modelscited by 1,143
  2. 2022MaskGIT: Masked Generative Image Transformercited by 1,082
  3. 2022Diffusion-LM Improves Controllable Text Generationcited by 1,006
  4. 2023Accelerating Large Language Model Decoding with Speculative Samplingcited by 975
  5. 2025Fast-dLLM: Training-free Acceleration of Diffusion LLM by Enabling KV Cache and Parallel Decodingcited by 404
  6. 2025Mercury: Ultra-Fast Language Models Based on Diffusioncited by 225
  7. 2024Simple Guidance Mechanisms for Discrete Diffusion Modelscited by 148
  8. 2022Training and Inference on Any-Order Autoregressive Models the Right Waycited by 64
  9. 2022Show Me What and Tell Me How: Video Synthesis via Multimodal Conditioningcited by 37
  10. 2025Self Speculative Decoding for Diffusion Large Language Modelscited by 29
  11. 2025Direct Discriminative Optimization: Your Likelihood-Based Visual Generative Model is Secretly a GAN Discriminatorcited by 22
  12. 2025Reviving Any-Subset Autoregressive Models with Principled Parallel Sampling and Speculative Decodingcited by 14
  13. 2025ReDi: Rectified Discrete Flowcited by 13
  14. 2025Improving Classifier-Free Guidance in Masked Diffusion: Low-Dim Theoretical Insights with High-Dim Impactcited by 11

Built on by

Cited by year (2026 to 7 Oct)

Citing papers per year: 2026: 5.

5 papers.

  1. 2026SimSD: Simple Speculative Decoding in Diffusion Language Modelscited by 2
  2. 2026SLIM-RL: Risk-Budgeted Random-Masking RL for Diffusion LLMs Without Trajectory Slicingcited by 0
  3. 2026From Positionwise Confidence to Prefix Scheduling: Verifier Skipping in Speculative Decodingcited by 0
  4. 2026From Position Risks to Block Survival: Faster Generation for Diffusion Language Modelscited by 0
  5. 2026FastGuide: Accelerating Reward Guidance for Diffusion Large Language Modelscited by 0

Similar, not cited

Close in wording, no citation either way

  1. 2026D^2SD: Accelerating Speculative Decoding with Dual Diffusion Draft Models
  2. 2026Rethinking the Generation Order of Block Diffusion Language Models
  3. 2026Early-Bird Decoding: Accelerating Diffusion LLMs with Learnable Block Sizes and Parallel Sampling
  4. 2025Accelerating Diffusion LLMs via Adaptive Parallel Decoding
  5. 2026DepCap: Adaptive Block-Wise Parallel Decoding for Efficient Diffusion LM Inference
  6. 2026PSD: Pushing the Pareto Frontier of Diffusion LLMs via Parallel Speculative Decoding
  7. 2026Speculative Refinement: A Hybrid Autoregressive Diffusion Decoding Strategy and Its Behavior Across Benchmarks
  8. 2025From Bits to Rounds: Parallel Decoding with Exploration for Diffusion Language Models