mlsanityData to 7 Oct 2026

Fail Fast, Win Big: Rethinking the Drafting Strategy in Speculative Decoding via Diffusion LLMs

arXiv 2512.20573 ↗First posted 23 Dec 2025cs.LG cs.AI cs.DCcited by 3

On the map: LLM reasoning → Test-time compute and efficient thinking

also in 1 more
Abstract (arXiv)

Diffusion Large Language Models (dLLMs) offer fast, parallel token generation, but their standalone use is plagued by an inherent efficiency-quality tradeoff. We show that, if carefully applied, the attributes of dLLMs can actually be a strength for drafters in speculative decoding with autoregressive (AR) verifiers. Our core insight is that dLLM's speed from parallel decoding drastically lowers the risk of costly rejections, providing a practical mechanism to effectively realize the (elusive) lengthy drafts that lead to large speedups with speculative decoding. We present FailFast, a dLLM-based speculative decoding framework that realizes this approach by dynamically adapting its speculation length. It "fails fast" by spending minimal compute in hard-to-speculate regions to shrink speculation latency and "wins big" by aggressively extending draft lengths in easier regions to reduce verification latency (in many cases, speculating and accepting 70 tokens at a time!). Without any fine-tuning, FailFast delivers lossless acceleration of AR LLMs and achieves up to 4.9× speedup over vanilla decoding, 1.7× over the best naive dLLM drafter, and 1.7× over EAGLE-3 across diverse models and workloads. We open-source FailFast at https://github.com/ruipeterpan/failfast.

Key facts. Cited by 3 papers in the map (data through 7 Oct 2026). Placed in Test-time compute and efficient thinking, one of 17 directions in LLM reasoning. Builds on 36 of its 59 references that are in the map. First posted on arXiv 23 Dec 2025.

Builds on

36 of 59 references.

Closest ancestors · 2

Same direction, newest first

  1. 2025Deep Think with Confidence21 Aug
  2. 2025SpecReason: Fast and Accurate Inference-Time Compute via Speculative Reasoning10 Apr

Other references · 34

  1. 2021Training Verifiers to Solve Math Word Problemscited by 9,743
  2. 2021Evaluating Large Language Models Trained on Codecited by 8,053
  3. 2021Measuring Mathematical Problem Solving With the MATH Datasetcited by 5,203
  4. 2024Qwen2.5 Technical Reportcited by 4,713
  5. 2024Transformers are SSMs: Generalized Models and Efficient Algorithms Through Structured State Space Dualitycited by 1,283
  6. 2025Large Language Diffusion Modelscited by 872
  7. 2024Medusa: Simple LLM Inference Acceleration Framework with Multiple Decoding Headscited by 821
  8. 2025Dream 7B: Diffusion Large Language Modelscited by 515
  9. 2025Fast-dLLM: Training-free Acceleration of Diffusion LLM by Enabling KV Cache and Parallel Decodingcited by 404
  10. 2025Block Diffusion: Interpolating Between Autoregressive and Diffusion Language Modelscited by 385
  11. 2025EAGLE-3: Scaling up Inference Acceleration of Large Language Models via Training-Time Testcited by 318
  12. 2024Break the Sequential Dependency of LLM Inference Using Lookahead Decodingcited by 295
22 more
  1. 2024Scaling Diffusion Language Models via Adaptation from Autoregressive Modelscited by 268
  2. 2024Unlocking Efficiency in Large Language Model Inference: A Comprehensive Survey of Speculative Decodingcited by 266
  3. 2025Mercury: Ultra-Fast Language Models Based on Diffusioncited by 225
  4. 2024An Empirical Study of Mamba-based Language Modelscited by 177
  5. 2025Seed Diffusion: A Large-Scale Diffusion Language Model with High-Speed Inferencecited by 173
  6. 2025dLLM-Cache: Accelerating Diffusion Large Language Models with Adaptive Cachingcited by 168
  7. 2025LLaDA2.0: Scaling Up Diffusion Language Models to 100Bcited by 154
  8. 2023REST: Retrieval-Based Speculative Decodingcited by 148
  9. 2025dKV-Cache: The Cache for Diffusion Language Modelscited by 140
  10. 2025Fast-dLLM v2: Efficient Block-Diffusion LLMcited by 128
  11. 2023Online Speculative Decodingcited by 109
  12. 2024MagicDec: Breaking the Latency-Throughput Tradeoff for Long Context Generation with Speculative Decodingcited by 87
  13. 2024SpecDec++: Boosting Speculative Decoding via Adaptive Candidate Lengthscited by 73
  14. 2024Turning Trash into Treasure: Accelerating Inference of Large Language Models with Token Recyclingcited by 48
  15. 2025ParallelBench: Understanding the Trade-offs of Parallel Decoding in Diffusion LLMscited by 42
  16. 2024TurboSpec: Closed-loop Speculation Control System for Optimizing LLM Serving Goodputcited by 36
  17. 2024SuffixDecoding: Extreme Speculative Decoding for Emerging AI Applicationscited by 34
  18. 2024Decoding Speculative Decodingcited by 33
  19. 2024Speculative Decoding via Early-exiting for Faster LLM Inference with Thompson Sampling Control Mechanismcited by 31
  20. 2025SpecDiff-2: Scaling Diffusion Drafter Alignment For Faster Speculative Decodingcited by 26
  21. 2025Structuring The Future: Diffusion LLM Speculative Decoding via Calibrated Draft Graphscited by 21
  22. 2025Not-a-Bandit: Provably No-Regret Drafter Selection in Speculative Decoding for LLMscited by 5

Built on by

Cited by year (2026 to 7 Oct)

Citing papers per year: 2026: 3.

3 papers.

  1. 2026Fast-dVLM: Efficient Block-Diffusion VLM via Direct Conversion from Autoregressive VLMcited by 8
  2. 2026TAPS: Target-Aware Prefix Tree Selection for Diffusion-Drafted Speculative Decodingcited by 7
  3. 2026Dynamic Delayed Tree Expansion For Improved Multi-Path Speculative Decodingcited by 0

Similar, not cited

Close in wording, no citation either way

  1. 2025DiffuSpec: Unlocking Diffusion Language Models for Speculative Decoding
  2. 2026LibraSpec: Dynamic Diffusion-Based Speculative Decoding via Marginal-Gain-Driven Optimization
  3. 2026DLoop: Looped Speculative Decoding
  4. 2025Speculative Verification: Exploiting Information Gain to Refine Speculative Decoding
  5. 2025Self Speculative Decoding for Diffusion Large Language Models
  6. 2026D^2SD: Accelerating Speculative Decoding with Dual Diffusion Draft Models
  7. 2026LongSpark: Efficient speculative decoding with a fixed-cost parallel drafter
  8. 2026DEdit: Iterative Draft Editing for Speculative Decoding