mlsanityData to 7 Oct 2026

Diffusion LLM and MoE inference

156 papers2 this weekMedium confidence in this grouping: mostly one topic, with blurry edgesMedium confidence in this grouping: mostly one topic, with blurry edges · About

Keywords: decoding, diffusion, parallel, dllms, moe

Key facts. Diffusion LLM and MoE inference: 156 papers on arXiv since 2024 (data through 7 Oct 2026). 76 of its 156 mapped papers (49%) cite “Fast-dLLM: Training-free Acceleration of Diffusion LLM by Enabling KV Cache and Parallel Decoding” (2025, arXiv 2505.22618). 39 of its 156 mapped papers (25%) cite “dKV-Cache: The Cache for Diffusion Language Models” (2025, arXiv 2505.15781). 11 new papers 24 Sep–7 Oct.

What’s new

24 Sep – 7 Oct 2026

11 new papers, 24 Sep – 7 Oct; 2 this week.

  1. This week · 2
  2. Enhancing Diffusion Language Models with Autoregressive Post-Training Weights
  3. SoloQ: Calibration-Free Quantization for Diffusion Language Models also fits: KV cache quantization
  4. Earlier · 9
  5. SpecFold: Folding Multi-Branch Redundancy for Faster Speculative Decoding in Diffusion Language Models also fits: Speculative decoding
  6. Large Language Continuous Diffusion Models
  7. ITC-MoE: Importance-guided Token-aware Compression for MoE Diffusion Language Models
  8. Clock Diffusion: Efficient Semi-Autoregressive Continuous Diffusion Language Models
  9. Rank-Aware Speculative Sampling for Diffusion Draft Trees also fits: Fast autoregressive visual generation
  10. Denoising Surface: Modeling and Predicting Inference Cost for Diffusion LLM Serving
  11. Does This Action Still Explain the Task? Reverse Scoring for Diffusion Language Model Agents also fits: Efficient reasoning and agent inference
  12. Rethinking Soft Tokens for Parallel Decoding in Diffusion Language Models
All 11
  1. Reliable Parallel Decoding in Masked Diffusion Language Models

Foundations

Most cited by papers here

  1. Fast-dLLM: Training-free Acceleration of Diffusion LLM by Enabling KV Cache and Parallel Decoding2025cited by 76 here
  2. dKV-Cache: The Cache for Diffusion Language Models2025cited by 39 here
  3. Accelerating Diffusion LLMs via Adaptive Parallel Decoding2025cited by 27 here
  4. Diffusion LLMs Can Do Faster-Than-AR Inference via Discrete Diffusion Forcing2025cited by 27 here
  5. dParallel: Learnable Parallel Decoding for dLLMs2025cited by 23 here

Lineage

The papers this direction builds on, oldest first.

Each paper on the line cites the one before.

  1. 2022Fast Inference from Transformers via Speculative Decodingcited by 46 here
  2. 2023Accelerating Large Language Model Decoding with Speculative Samplingcited by 34 here
  3. 2024Medusa: Simple LLM Inference Acceleration Framework with Multiple Decoding Headscited by 29 here
  4. 2024Break the Sequential Dependency of LLM Inference Using Lookahead Decodingcited by 10 here
  5. 2025Accelerating Diffusion LLMs via Adaptive Parallel Decodingcited by 27 here
  6. 2025Diffusion LLMs Can Do Faster-Than-AR Inference via Discrete Diffusion Forcingcited by 27 here
  7. 2025dParallel: Learnable Parallel Decoding for dLLMscited by 23 here
17 more it builds on, not in this chain
  1. 2023SpecInfer: Accelerating Generative Large Language Model Serving with Tree-based Speculative Inference and Verificationcited by 10 here
  2. 2023Efficient Memory Management for Large Language Model Serving with PagedAttentioncited by 33 here
  3. 2024DistServe: Disaggregating Prefill and Decoding for Goodput-optimized Large Language Model Servingcited by 9 here
  4. 2024EAGLE: Speculative Sampling Requires Rethinking Feature Uncertaintycited by 19 here
  5. 2024EAGLE-2: Faster Inference of Language Models with Dynamic Draft Treescited by 12 here
  6. 2025EAGLE-3: Scaling up Inference Acceleration of Large Language Models via Training-Time Testcited by 16 here
  7. 2025dKV-Cache: The Cache for Diffusion Language Modelscited by 39 here
  8. 2025LaViDa: A Large Diffusion Language Model for Multimodal Understandingcited by 11 here
  9. 2025Dimple: Discrete Diffusion Multimodal Large Language Model with Parallel Decodingcited by 17 here
  10. 2025Fast-dLLM: Training-free Acceleration of Diffusion LLM by Enabling KV Cache and Parallel Decodingcited by 76 here
  11. 2025Sparse-dLLM: Accelerating Diffusion LLMs with Dynamic Cache Evictioncited by 20 here
  12. 2025Learning to Parallel: Accelerating Diffusion Large Language Models via Learnable Parallel Decodingcited by 12 here
  13. 2025AdaBlock-dLLM: Semantic-Aware Diffusion LLM Inference via Adaptive Block Sizecited by 9 here
  14. 2025Free Draft-and-Verification: Toward Lossless Parallel Decoding for Diffusion Large Language Modelscited by 11 here
  15. 2025Self Speculative Decoding for Diffusion Large Language Modelscited by 13 here
  16. 2025ParallelBench: Understanding the Trade-offs of Parallel Decoding in Diffusion LLMscited by 13 here
  17. 2025Accelerating Diffusion LLM Inference via Local Determinism Propagationcited by 11 here

The 24 most cited of 242.

Browse

30 of 156, ranked 6–35

  1. 6Sparse-dLLM: Accelerating Diffusion LLMs with Dynamic Cache Eviction202520
  2. 7Dimple: Discrete Diffusion Multimodal Large Language Model with Parallel Decoding202517
  3. 8ParallelBench: Understanding the Trade-offs of Parallel Decoding in Diffusion LLMs202513
  4. 9Learning to Parallel: Accelerating Diffusion Large Language Models via Learnable Parallel Decoding202512
  5. 10Free Draft-and-Verification: Toward Lossless Parallel Decoding for Diffusion Large Language Models202511
  6. 11Accelerating Diffusion LLM Inference via Local Determinism Propagation202511
  7. 12AdaBlock-dLLM: Semantic-Aware Diffusion LLM Inference via Adaptive Block Size20259
  8. 13Attention Is All You Need for KV Cache in Diffusion LLMs20259
  9. 14d3LLM: Ultra-Fast Diffusion LLM using Pseudo-Trajectory Distillation20268
  10. 15Structuring The Future: Diffusion LLM Speculative Decoding via Calibrated Draft Graphs20258
All 30
  1. 16d2Cache: Accelerating Diffusion-Based LLMs via Dual Adaptive Caching20258
  2. 17From Bits to Rounds: Parallel Decoding with Exploration for Diffusion Language Models20258
  3. 18Utility-Driven Speculative Decoding for Mixture-of-Experts20257
  4. 19SP-MoE: Speculative Decoding and Prefetching for Accelerating MoE-based Model Inference20257
  5. 20ReFusion: A Diffusion Large Language Model with Parallel Autoregressive Decoding20257
  6. 21TiDAR: Think in Diffusion, Talk in Autoregression20256
  7. 22DAWN: Dependency-Aware Fast Inference for Diffusion LLMs20265
  8. 23MoE-Spec: Expert Budgeting for Efficient Speculative Decoding20265
  9. 24Why Diffusion Language Models Struggle with Truly Parallel (Non-Autoregressive) Decoding?20265
  10. 25Accelerating Mixture-of-Experts Inference by Hiding Offloading Latency with Speculative Decoding20255
  11. 26CreditDecoding: Accelerating Parallel Decoding in Diffusion Large Language Models with Trace Credit20255
  12. 27LoPA: Scaling dLLM Inference via Lookahead Parallel Decoding20255
  13. 28Taming Latency-Memory Trade-Off in MoE-Based LLM Serving via Fine-Grained Expert Offloading20254
  14. 29Diffusion Models are Secretly Exchangeable: Parallelizing DDPMs via Autospeculation20254
  15. 30SpecMoE: A Fast and Efficient Mixture-of-Experts Inference via Self-Assisted Speculative Decoding20263
  16. 31Less Experts, Faster Decoding: Cost-Aware Speculative Decoding for Mixture-of-Experts20263
  17. 32Taming Masked Diffusion Language Models via Consistency Trajectory Reinforcement Learning with Fewer Decoding Step20253
  18. 33Mask Tokens as Prophet: Fine-Grained Cache Eviction for Efficient dLLM Inference20253
  19. 34Duplex: A Device for Large Language Models with Mixture of Experts, Grouped Query Attention, and Continuous Batching20243
  20. 35Beyond Hard Masks: Progressive Token Evolution for Diffusion Language Models20262