pith. sign in

arxiv: 2606.10537 · v1 · pith:RR5ODKUTnew · submitted 2026-06-09 · 💻 cs.CL

Prefilling-dLLM: Predictive Prefilling for Long-Context Inference in Diffusion Language Models

Jing Xiong , Qi Han , Shansan Gong , Yunta Hsieh , Chengyue Wu , Chaofan Tao , Chenyang Zhao , Ngai Wong This is my paper

Pith reviewed 2026-06-27 13:28 UTC · model grok-4.3

classification 💻 cs.CL
keywords diffusion language modelslong-context inferencesparse prefillingKV cachingdenoisingattention sparsitylost-in-the-middle
0
0 comments X

The pith

Diffusion language models can use cached chunk KV states and top-K selection to make long-context denoising quadratic only in decode length while outperforming dense attention.

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

Diffusion LLMs re-encode the full prefix at each denoising step, causing quadratic costs that hinder long contexts. Prefilling-dLLM partitions the prefix into chunks, caches their KV representations once, and selects top-K chunks with intra-chunk sparsity for each decode step. This sparse prefilling is shown to outperform dense attention on quality while cutting complexity. Periodic BOS tokens in chunks act as anchors to fix lost-in-the-middle issues. The method is training-free and delivers large speedups on long-context benchmarks.

Core claim

Prefilling-dLLM is a training-free prefill-decode disaggregation framework for dLLMs that partitions the prefix into N chunks, caches their KV representations once, and selects the top-K most relevant chunks with intra-chunk token sparsity for decoding. This shows that sparse prefilling can outperform dense attention while reducing per-step complexity from quadratic in the full sequence length to quadratic only in the decode length. Beginning-of-sequence tokens prepended to each chunk act as periodic attention anchors that eliminate the lost-in-the-middle phenomenon.

What carries the argument

Prefill-decode disaggregation via chunked KV caching and top-K selection with intra-chunk sparsity, augmented by periodic BOS anchors.

Load-bearing premise

The top-K chunk selection plus periodic BOS anchors preserve all information required for accurate denoising at every step without introducing new errors that dense attention would have avoided.

What would settle it

If the Prefilling-dLLM method produces lower accuracy than full dense attention on the same dLLM for tasks in InfiniteBench, that would falsify the claim that sparse prefilling outperforms dense attention.

Figures

Figures reproduced from arXiv: 2606.10537 by Chaofan Tao, Chengyue Wu, Chenyang Zhao, Jing Xiong, Ngai Wong, Qi Han, Shansan Gong, Yunta Hsieh.

Figure 1
Figure 1. Figure 1: Lost-in-the-Middle evaluation on Dream-7B [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Attention weight decay as a function of dis [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Overview of PREFILLING-DLLM. (I) Prefill: The prefix is partitioned into N chunks, each independently prefilled with intra-chunk attention to produce per-chunk KV caches; chunks are ranked by a predictive score combining self-information and pseudo-label logits; the top-K chunks are selected, and only the top-B query￾relevant tokens per chunk are retained in the KV cache. (II) Sparse Attention: During deco… view at source ↗
Figure 4
Figure 4. Figure 4: Throughput comparison (tokens/s) on Long [PITH_FULL_IMAGE:figures/full_fig_p006_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Attention kernel benchmark (bf16, GQA 32/8 [PITH_FULL_IMAGE:figures/full_fig_p007_5.png] view at source ↗
Figure 7
Figure 7. Figure 7: Attention sink analysis (8K context, YaRN [PITH_FULL_IMAGE:figures/full_fig_p008_7.png] view at source ↗
Figure 8
Figure 8. Figure 8: Attention landscape during denoising (layer [PITH_FULL_IMAGE:figures/full_fig_p008_8.png] view at source ↗
Figure 10
Figure 10. Figure 10: Latency breakdown of PREFILLING-DLLM into prefilling (chunk scoring + cache build) and de￾coding (diffusion generation). Decoding time remains constant (∼0.86s) regardless of context length, while prefilling scales with input size. D Prefilling Efficiency Analysis To isolate the contribution of each phase, we sep￾arately measure the latency of prefilling (chunk scoring + KV cache construction) and decodin… view at source ↗
read the original abstract

Diffusion large language models (dLLMs) re-encode the entire prefix at every denoising step, causing recomputation that scales quadratically with context length and becomes prohibitive for long-context scenarios. We propose Prefilling-dLLM, a training-free prefill-decode disaggregation framework for dLLMs that partitions the prefix into N chunks, caches their KV representations once, and selects the top-K most relevant chunks with intra-chunk token sparsity for decoding, showing that sparse prefilling can outperform dense attention while reducing per-step complexity from quadratic in the full sequence length to quadratic only in the decode length. On LongBench and InfiniteBench, Prefilling-dLLM achieves state-of-the-art quality among dLLM acceleration methods, and an attention kernel that parallelizes decoding over the non-contiguously cached chunk KV yields 9.1--28.0x speedup at 8K--32K contexts. We further show that beginning-of-sequence tokens prepended to each chunk act as periodic attention anchors that eliminate the lost-in-the-middle phenomenon. Code is available at https://github.com/menik1126/Prefilling-dLLM.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

3 major / 1 minor

Summary. This paper presents Prefilling-dLLM, a training-free prefill-decode disaggregation framework for diffusion language models. It partitions long prefixes into N chunks, caches their KV representations once, selects the top-K most relevant chunks using a relevance metric with intra-chunk sparsity, and prepends BOS tokens to each chunk as periodic anchors. The method claims to reduce per-step attention complexity from quadratic in full sequence length to quadratic only in decode length, while achieving SOTA quality among dLLM acceleration methods on LongBench and InfiniteBench and delivering 9.1-28x speedups at 8K-32K contexts via a parallel attention kernel over non-contiguous cached chunks. Code is released at the cited GitHub repository.

Significance. If the empirical claims are substantiated, the work would offer a practical, training-free route to scalable long-context inference in dLLMs by showing that relevance-based sparse prefilling can exceed dense attention quality. The open-sourced code is a clear strength that supports reproducibility and follow-on research.

major comments (3)
  1. [Abstract] Abstract: the central claim that sparse prefilling can outperform dense attention rests on the unverified assumption that top-K chunk selection (plus BOS anchors) never drops information required for accurate denoising; the manuscript must supply ablations that replace top-K with full dense selection or random selection while holding BOS and other factors fixed, to isolate whether the reported quality gains are attributable to the selection heuristic.
  2. [Experiments] Experiments section (LongBench/InfiniteBench tables): no error bars, standard deviations across runs, or sensitivity analysis for N and K are reported, so it is impossible to determine whether the SOTA quality and 9.1-28x speedups are statistically robust or sensitive to the heuristic top-K choices.
  3. [Method] Method (chunk selection paragraph): the per-step relevance metric for top-K selection is described only at a high level; without an explicit definition or pseudocode showing how it is computed from the current denoising state, it remains unclear whether the metric aligns with the true cross-token dependencies in the diffusion process at every step.
minor comments (1)
  1. [Abstract] Abstract: the speedup range 9.1-28.0x should be tied to the precise context lengths, model sizes, and values of N/K used in each measurement for immediate clarity.

Simulated Author's Rebuttal

3 responses · 0 unresolved

We thank the referee for the constructive feedback. We address each major comment below and indicate where revisions will be made to strengthen the manuscript.

read point-by-point responses

  1. Referee: [Abstract] Abstract: the central claim that sparse prefilling can outperform dense attention rests on the unverified assumption that top-K chunk selection (plus BOS anchors) never drops information required for accurate denoising; the manuscript must supply ablations that replace top-K with full dense selection or random selection while holding BOS and other factors fixed, to isolate whether the reported quality gains are attributable to the selection heuristic.

    Authors: We agree that additional ablations are required to isolate the effect of the top-K heuristic and substantiate the claim that sparse prefilling can outperform dense attention. In the revised manuscript we will add experiments that replace top-K selection with full dense selection and with random selection while holding BOS anchors and all other factors fixed. revision: yes

  2. Referee: [Experiments] Experiments section (LongBench/InfiniteBench tables): no error bars, standard deviations across runs, or sensitivity analysis for N and K are reported, so it is impossible to determine whether the SOTA quality and 9.1-28x speedups are statistically robust or sensitive to the heuristic top-K choices.

    Authors: We acknowledge that the current results lack error bars and sensitivity analysis. We will rerun all LongBench and InfiniteBench evaluations across multiple random seeds, report means with standard deviations, and add sensitivity plots for N and K in the revised experiments section. revision: yes

  3. Referee: [Method] Method (chunk selection paragraph): the per-step relevance metric for top-K selection is described only at a high level; without an explicit definition or pseudocode showing how it is computed from the current denoising state, it remains unclear whether the metric aligns with the true cross-token dependencies in the diffusion process at every step.

    Authors: We will revise the method section to include an explicit mathematical definition of the per-step relevance metric together with pseudocode that shows its computation from the current denoising state at each diffusion step. revision: yes

Circularity Check

0 steps flagged

No circularity; algorithmic framework is self-contained

full rationale

The paper describes a training-free algorithmic procedure: partition prefix into chunks, cache KV once, select top-K by relevance, prepend BOS anchors. The complexity reduction (quadratic in decode length only) follows directly from the stated partitioning and selection without any fitted parameters, self-referential equations, or predictions that equal their inputs by construction. Quality claims rest on benchmark measurements rather than derivations. No self-citation load-bearing steps, uniqueness theorems, or ansatzes appear in the abstract or described method. This matches the default case of an independent algorithmic contribution.

Axiom & Free-Parameter Ledger

2 free parameters · 2 axioms · 0 invented entities

The approach rests on standard transformer KV-cache validity and attention sparsity assumptions; no new physical entities or fitted constants are introduced beyond routine hyperparameters N and K.

free parameters (2)
  • N (number of prefix chunks)
    Hyperparameter controlling prefix partitioning; value chosen per experiment but not fitted to target metric.
  • K (top chunks selected)
    Hyperparameter controlling sparsity level; chosen to balance speed and quality.
axioms (2)
  • domain assumption KV representations computed once per chunk remain valid across all subsequent denoising steps
    Invoked when the method caches chunk KV states before decoding begins.
  • domain assumption Top-K relevance scoring selects chunks containing all necessary context for correct denoising
    Central to the claim that sparse prefilling can outperform dense attention.

pith-pipeline@v0.9.1-grok · 5769 in / 1487 out tokens · 19207 ms · 2026-06-27T13:28:01.930497+00:00 · methodology

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

73 extracted references · 15 canonical work pages · 7 internal anchors

  1. [1]

    18th USENIX Symposium on Operating Systems Design and Implementation (OSDI 24) , pages=

    \ DistServe \ : Disaggregating prefill and decoding for goodput-optimized large language model serving , author=. 18th USENIX Symposium on Operating Systems Design and Implementation (OSDI 24) , pages=

  2. [2]

    ACM Transactions on Storage , year=

    Mooncake: A kvcache-centric disaggregated architecture for llm serving , author=. ACM Transactions on Storage , year=

  3. [3]

    arXiv preprint arXiv:2510.08544 , year=

    SPAD: Specialized Prefill and Decode Hardware for Disaggregated LLM Inference , author=. arXiv preprint arXiv:2510.08544 , year=

    work page arXiv

  4. [4]

    arXiv preprint arXiv:2504.19867 , year=

    semi-pd: Towards efficient llm serving via phase-wise disaggregated computation and unified storage , author=. arXiv preprint arXiv:2504.19867 , year=

    work page arXiv

  5. [5]

    Advances in Neural Information Processing Systems , volume=

    dkv-cache: The cache for diffusion language models , author=. Advances in Neural Information Processing Systems , volume=

  6. [6]

    2025 , eprint=

    SeerAttention: Learning Intrinsic Sparse Attention in Your LLMs , author=. 2025 , eprint=

    2025

  7. [7]

    International Conference on Machine Learning (ICML) , year=

    c sparse attention accelerating any model inference , author=. International Conference on Machine Learning (ICML) , year=

  8. [8]

    International Conference on Learning Representations , volume=

    Sageattention: Accurate 8-bit attention for plug-and-play inference acceleration , author=. International Conference on Learning Representations , volume=

  9. [9]

    International Conference on Machine Learning (ICML) , year=

    Sageattention2: Efficient attention with thorough outlier smoothing and per-thread int4 quantization , author=. International Conference on Machine Learning (ICML) , year=

  10. [10]

    International Conference on Learning Representations , year=

    Efficient Streaming Language Models with Attention Sinks , author=. International Conference on Learning Representations , year=

  11. [11]

    2025 , eprint=

    When Attention Sink Emerges in Language Models: An Empirical View , author=. 2025 , eprint=

    2025

  12. [12]

    2025 , eprint=

    CTR-Sink: Attention Sink for Language Models in Click-Through Rate Prediction , author=. 2025 , eprint=

    2025

  13. [13]

    arXiv preprint arXiv:2509.13866 , year=

    Masked Diffusion Models as Energy Minimization , author=. arXiv preprint arXiv:2509.13866 , year=

    work page arXiv

  14. [14]

    LLaDA-V: Large Language Diffusion Models with Visual Instruction Tuning

    Llada-v: Large language diffusion models with visual instruction tuning , author=. arXiv preprint arXiv:2505.16933 , year=

    work page internal anchor Pith review Pith/arXiv arXiv

  15. [15]

    A Survey on Diffusion Language Models

    A survey on diffusion language models , author=. arXiv preprint arXiv:2508.10875 , year=

    work page internal anchor Pith review Pith/arXiv arXiv

  16. [16]

    2025 , eprint=

    Block Diffusion: Interpolating Between Autoregressive and Diffusion Language Models , author=. 2025 , eprint=

    2025

  17. [17]

    2025 , eprint=

    LLaDA2.0: Scaling Up Diffusion Language Models to 100B , author=. 2025 , eprint=

    2025

  18. [18]

    2025 , eprint=

    Scaling Diffusion Language Models via Adaptation from Autoregressive Models , author=. 2025 , eprint=

    2025

  19. [19]

    2025 , eprint=

    SlimInfer: Accelerating Long-Context LLM Inference via Dynamic Token Pruning , author=. 2025 , eprint=

    2025

  20. [20]

    2025 , eprint=

    InfLLM-V2: Dense-Sparse Switchable Attention for Seamless Short-to-Long Adaptation , author=. 2025 , eprint=

    2025

  21. [21]

    2024 , eprint=

    InfLLM: Training-Free Long-Context Extrapolation for LLMs with an Efficient Context Memory , author=. 2024 , eprint=

    2024

  22. [22]

    2025 , eprint=

    Fast-dLLM: Training-free Acceleration of Diffusion LLM by Enabling KV Cache and Parallel Decoding , author=. 2025 , eprint=

    2025

  23. [23]

    2025 , eprint=

    dLLM-Cache: Accelerating Diffusion Large Language Models with Adaptive Caching , author=. 2025 , eprint=

    2025

  24. [24]

    2025 , eprint=

    Dream 7B: Diffusion Large Language Models , author=. 2025 , eprint=

    2025

  25. [25]

    2025 , eprint=

    Large Language Diffusion Models , author=. 2025 , eprint=

    2025

  26. [26]

    2021 , eprint=

    Training Verifiers to Solve Math Word Problems , author=. 2021 , eprint=

    2021

  27. [27]

    2025 , eprint=

    Attention Sinks in Diffusion Language Models , author=. 2025 , eprint=

    2025

  28. [28]

    2025 , eprint=

    Sparse-dLLM: Accelerating Diffusion LLMs with Dynamic Cache Eviction , author=. 2025 , eprint=

    2025

  29. [29]

    2025 , eprint=

    Mask Tokens as Prophet: Fine-Grained Cache Eviction for Efficient dLLM Inference , author=. 2025 , eprint=

    2025

  30. [30]

    2025 , eprint=

    Attention Is All You Need for KV Cache in Diffusion LLMs , author=. 2025 , eprint=

    2025

  31. [31]

    2025 , eprint=

    d ^2 Cache: Accelerating Diffusion-Based LLMs via Dual Adaptive Caching , author=. 2025 , eprint=

    2025

  32. [32]

    2025 , eprint=

    Discrete Diffusion in Large Language and Multimodal Models: A Survey , author=. 2025 , eprint=

    2025

  33. [33]

    2025 , eprint=

    A Survey on Diffusion Language Models , author=. 2025 , eprint=

    2025

  34. [34]

    2025 , eprint=

    LongLLaDA: Unlocking Long Context Capabilities in Diffusion LLMs , author=. 2025 , eprint=

    2025

  35. [35]

    2025 , eprint=

    UltraLLaDA: Scaling the Context Length to 128K for Diffusion Large Language Models , author=. 2025 , eprint=

    2025

  36. [36]

    arXiv preprint arXiv:2112.05682 , year=

    Self-attention Does Not Need O(n^2) Memory , author=. arXiv preprint arXiv:2112.05682 , year=

    work page arXiv

  37. [37]

    Advances in Neural Information Processing Systems , volume=

    H2o: Heavy-hitter oracle for efficient generative inference of large language models , author=. Advances in Neural Information Processing Systems , volume=

  38. [38]

    Advances in Neural Information Processing Systems , volume=

    Snapkv: Llm knows what you are looking for before generation , author=. Advances in Neural Information Processing Systems , volume=

  39. [39]

    Advances in Neural Information Processing Systems , volume=

    Infllm: Training-free long-context extrapolation for llms with an efficient context memory , author=. Advances in Neural Information Processing Systems , volume=

  40. [40]

    Quest: Query-Aware Sparsity for Efficient Long-Context LLM Inference

    Quest: Query-aware sparsity for efficient long-context llm inference , author=. arXiv preprint arXiv:2406.10774 , year=

    work page internal anchor Pith review Pith/arXiv arXiv

  41. [41]

    Advances in Neural Information Processing Systems , volume=

    Minference 1.0: Accelerating pre-filling for long-context llms via dynamic sparse attention , author=. Advances in Neural Information Processing Systems , volume=

  42. [42]

    arXiv preprint arXiv:2502.20766 , year=

    Flexprefill: A context-aware sparse attention mechanism for efficient long-sequence inference , author=. arXiv preprint arXiv:2502.20766 , year=

    work page arXiv

  43. [43]

    arXiv preprint arXiv:2509.24014 , year=

    SparseD: Sparse Attention for Diffusion Language Models , author=. arXiv preprint arXiv:2509.24014 , year=

    work page arXiv

  44. [44]

    2025 , eprint=

    Diffusion LLMs Can Do Faster-Than-AR Inference via Discrete Diffusion Forcing , author=. 2025 , eprint=

    2025

  45. [45]

    2023 , eprint=

    LLaMA: Open and Efficient Foundation Language Models , author=. 2023 , eprint=

    2023

  46. [46]

    2025 , eprint=

    What are you sinking? A geometric approach on attention sink , author=. 2025 , eprint=

    2025

  47. [47]

    2023 , eprint=

    Qwen Technical Report , author=. 2023 , eprint=

    2023

  48. [48]

    2025 , url=

    Ruyi Xu and Guangxuan Xiao and Haofeng Huang and Junxian Guo and Song Han , booktitle=. 2025 , url=

    2025

  49. [49]

    2025 , eprint=

    Native Sparse Attention: Hardware-Aligned and Natively Trainable Sparse Attention , author=. 2025 , eprint=

    2025

  50. [50]

    2025 , eprint=

    LLaDA 1.5: Variance-Reduced Preference Optimization for Large Language Diffusion Models , author=. 2025 , eprint=

    2025

  51. [51]

    arXiv preprint arXiv:2509.26328 , year=

    Fast-dllm v2: Efficient block-diffusion llm , author=. arXiv preprint arXiv:2509.26328 , year=

    work page arXiv

  52. [52]

    Advances in neural information processing systems , volume=

    Structured denoising diffusion models in discrete state-spaces , author=. Advances in neural information processing systems , volume=

  53. [53]

    Proceedings of the 61st annual meeting of the association for computational linguistics (volume 1: Long papers) , pages=

    Diffusionbert: Improving generative masked language models with diffusion models , author=. Proceedings of the 61st annual meeting of the association for computational linguistics (volume 1: Long papers) , pages=

  54. [54]

    DiffuSeq: Sequence to Sequence Text Generation with Diffusion Models

    Diffuseq: Sequence to sequence text generation with diffusion models , author=. arXiv preprint arXiv:2210.08933 , year=

    work page internal anchor Pith review Pith/arXiv arXiv

  55. [55]

    Advances in neural information processing systems , volume=

    Diffusion-lm improves controllable text generation , author=. Advances in neural information processing systems , volume=

  56. [56]

    Advances in Neural Information Processing Systems , volume=

    Simple and effective masked diffusion language models , author=. Advances in Neural Information Processing Systems , volume=

  57. [57]

    Discrete Diffusion Modeling by Estimating the Ratios of the Data Distribution

    Discrete diffusion modeling by estimating the ratios of the data distribution , author=. arXiv preprint arXiv:2310.16834 , year=

    work page internal anchor Pith review Pith/arXiv arXiv

  58. [58]

    Proceedings of the 62nd annual meeting of the association for computational linguistics (volume 1: Long papers) , pages=

    Longbench: A bilingual, multitask benchmark for long context understanding , author=. Proceedings of the 62nd annual meeting of the association for computational linguistics (volume 1: Long papers) , pages=

  59. [59]

    OpenCompass: A Universal Evaluation Platform for Foundation Models , author=

  60. [60]

    , title =

    Kamradt, G. , title =. GitHub repository , howpublished =. 2023 , publisher =

    2023

  61. [61]

    Proceedings of the 3rd ACM SIGPLAN International Workshop on Machine Learning and Programming Languages , pages=

    Triton: an intermediate language and compiler for tiled neural network computations , author=. Proceedings of the 3rd ACM SIGPLAN International Workshop on Machine Learning and Programming Languages , pages=

  62. [62]

    Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers) , pages=

    Bench: Extending long context evaluation beyond 100K tokens , author=. Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers) , pages=

  63. [63]

    and Ermon, Stefano and Rudra, Atri and R

    Dao, Tri and Fu, Daniel Y. and Ermon, Stefano and Rudra, Atri and R. Flash. Advances in Neural Information Processing Systems (NeurIPS) , year=

  64. [64]

    Advances in Neural Information Processing Systems , volume=

    Flashattention-3: Fast and accurate attention with asynchrony and low-precision , author=. Advances in Neural Information Processing Systems , volume=

  65. [65]

    Transactions of the Association for Computational Linguistics , volume=

    Lost in the middle: How language models use long contexts , author=. Transactions of the Association for Computational Linguistics , volume=

  66. [66]

    2024 , url=

    Dong, Juechu and Feng, Boyuan and Guessous, Driss and Liang, Yanbo and He, Horace , journal=. 2024 , url=

    2024

  67. [67]

    xFormers: A modular and hackable Transformer modelling library , author=

  68. [68]

    Proceedings of Machine Learning and Systems , volume=

    Flashinfer: Efficient and customizable attention engine for llm inference serving , author=. Proceedings of Machine Learning and Systems , volume=

  69. [69]

    YaRN: Efficient Context Window Extension of Large Language Models

    YaRN: Efficient Context Window Extension of Large Language Models , author=. arXiv preprint arXiv:2309.00071 , year=

    work page internal anchor Pith review Pith/arXiv arXiv

  70. [70]

    Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing: System Demonstrations , pages=

    Transformers: State-of-the-Art Natural Language Processing , author=. Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing: System Demonstrations , pages=

    2020

  71. [71]

    LoSA: Locality Aware Sparse Attention for Block-Wise Diffusion Language Models

    LoSA: Locality Aware Sparse Attention for Block-Wise Diffusion Language Models , author=. arXiv preprint arXiv:2604.12056 , year=

    work page internal anchor Pith review Pith/arXiv arXiv

  72. [72]

    arXiv preprint arXiv:2602.02159 , year=

    Focus-dLLM: Accelerating Long-Context Diffusion LLM Inference via Confidence-Guided Context Focusing , author=. arXiv preprint arXiv:2602.02159 , year=

    work page arXiv

  73. [73]

    Proceedings of the 41st International Conference on Machine Learning , articleno =

    An, Chenxin and Huang, Fei and Zhang, Jun and Gong, Shansan and Qiu, Xipeng and Zhou, Chang and Kong, Lingpeng , title =. Proceedings of the 41st International Conference on Machine Learning , articleno =. 2024 , publisher =

    2024