mlsanityData to 14 Oct 2026

Beyond Scattered Acceptance: Fast and Coherent Inference for DLMs via Longest Stable Prefixes

arXiv 2603.05454 ↗First posted 5 Mar 2026cs.CVcited by 0

On the map: Diffusion models → Discrete and text diffusion

Abstract (arXiv)

Diffusion Language Models (DLMs) promise highly parallel text generation, yet their practical inference speed is often bottlenecked by suboptimal decoding schedulers. Standard approaches rely on 'scattered acceptance'-committing high confidence tokens at disjoint positions throughout the sequence. This approach inadvertently fractures the Key-Value (KV) cache, destroys memory locality, and forces the model into costly, repeated repairs across unstable token boundaries. To resolve this, we present the Longest Stable Prefix (LSP) scheduler, a training-free and model-agnostic inference paradigm based on monolithic prefix absorption. In each denoising step, LSP evaluates token stability via a single forward pass, dynamically identifies a contiguous left-aligned block of stable predictions, and snaps its boundary to natural linguistic or structural delimiters before an atomic commitment. This prefix-first topology yields dual benefits: systemically, it converts fragmented KV cache updates into efficient, contiguous appends; algorithmically, it preserves bidirectional lookahead over a geometrically shrinking active suffix, drastically reducing token flip rates and denoiser calls. Extensive evaluations on LLaDA-8B and Dream-7B demonstrate that LSP accelerates inference by up to 3.4x across rigorous benchmarks including mathematical reasoning, code generation, multilingual (CJK) tasks, and creative writing while matching or slightly improving output quality. By fundamentally restructuring the commitment topology, LSP bridges the gap between the theoretical parallelism of DLMs and practical hardware efficiency.

Key facts. No citing papers in the map yet (data through 14 Oct 2026). Placed in Discrete and text diffusion, one of 14 directions in Diffusion models. Builds on 28 of its 30 references that are in the map. First posted on arXiv 5 Mar 2026.

Builds on

28 of 30 references.

Closest ancestors · 13

Same direction, newest first

  1. 2025Diffusion Language Models Know the Answer Before Decoding27 Aug
  2. 2025Dream 7B: Diffusion Large Language Models21 Aug
  3. 2025DPad: Efficient Diffusion Language Models with Suffix Dropout19 Aug
  4. 2025Time Is a Feature: Exploiting Temporal Dynamics in Diffusion Language Models12 Aug
  5. 2025Seed Diffusion: A Large-Scale Diffusion Language Model with High-Speed Inference4 Aug
  6. 2025FlashDLM: Accelerating Diffusion Language Model Inference via Efficient KV Caching and Guided Diffusion27 May
  7. 2025CtrlDiff: Boosting Large Diffusion Language Models with Dynamic Block Prediction and Controllable Generation20 May
  8. 2025Block Diffusion: Interpolating Between Autoregressive and Diffusion Language Models12 Mar
  9. 2025Large Language Diffusion Models14 Feb
  10. 2024Masked Diffusion Models are Secretly Time-Agnostic Masked Models and Exploit Inaccurate Categorical Sampling4 Sep
  11. 2024Simple and Effective Masked Diffusion Language Models11 Jun
  12. 2024Simplified and Generalized Masked Diffusion for Discrete Data6 Jun
1 more
  1. 2024Your Absorbing Discrete Diffusion Secretly Models the Conditional Distributions of Clean Data6 Jun

Other references · 15

  1. 2021Training Verifiers to Solve Math Word Problemscited by 10,109
  2. 2021Evaluating Large Language Models Trained on Codecited by 8,320
  3. 2015Deep Unsupervised Learning using Nonequilibrium Thermodynamicscited by 7,058
  4. 2021Program Synthesis with Large Language Modelscited by 3,637
  5. 2021Structured Denoising Diffusion Models in Discrete State-Spacescited by 1,719
  6. 2021Argmax Flows and Multinomial Diffusion: Learning Categorical Distributionscited by 574
  7. 2022A Continuous Time Framework for Discrete Denoising Modelscited by 451
  8. 2025Fast-dLLM: Training-free Acceleration of Diffusion LLM by Enabling KV Cache and Parallel Decodingcited by 420
  9. 2025Mercury: Ultra-Fast Language Models Based on Diffusioncited by 235
  10. 2025dLLM-Cache: Accelerating Diffusion Large Language Models with Adaptive Cachingcited by 169
  11. 2025Accelerating Diffusion LLMs via Adaptive Parallel Decodingcited by 112
  12. 2025Sparse-dLLM: Accelerating Diffusion LLMs with Dynamic Cache Evictioncited by 65
3 more
  1. 2025Accelerating Diffusion Large Language Models with SlowFast Sampling: The Three Golden Principlescited by 45
  2. 2025Adaptive Classifier-Free Guidance via Dynamic Low-Confidence Maskingcited by 21
  3. 2025Structuring The Future: Diffusion LLM Speculative Decoding via Calibrated Draft Graphscited by 21

Built on by

Not cited yet.

Similar, not cited

Close in wording, no citation either way

  1. 2026CAI-DLLM: Convergence Aware Inference for Diffusion Language Models
  2. 2026Polestar: Drift-Aware Cache Calibration and Token Commitment for Efficient Inference of Diffusion LLMs
  3. 2026DAWN: Dependency-Aware Fast Inference for Diffusion LLMs
  4. 2025Improving the Throughput of Diffusion-based Large Language Models via a Training-Free Confidence-Aware Calibration
  5. 2026Window-Diffusion: Accelerating Diffusion Language Model Inference with Windowed Token Pruning and Caching
  6. 2026AsyncLane: Decoupling Refinement from Advancement in Diffusion Language Model Decoding
  7. 2026Streaming-dLLM: Accelerating Diffusion LLMs via Suffix Pruning and Dynamic Decoding
  8. 2026Roll Out and Roll Back: Diffusion LLMs are Their Own Efficiency Teachers