pith. sign in

arxiv: 2605.18165 · v1 · pith:QGMVEJZKnew · submitted 2026-05-18 · 💻 cs.LG

Elastic-dLLM: Position Preserving Context Compression and Augmentation of Diffusion LLMs

Junyi Wu , Tianchen Zhao , Shaoqiu Zhang , Linfeng Zhang , Guohao Dai , Yu Wang This is my paper

Pith reviewed 2026-05-20 13:09 UTC · model grok-4.3

classification 💻 cs.LG
keywords diffusion LLMsMASK token compressionparallel decodingcontext augmentationlong-context scalingdLLMredundancy reduction
0
0 comments X

The pith

Compressing redundant MASK tokens in diffusion LLMs speeds up decoding while preserving position and structural information.

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

The paper establishes that diffusion large language models incur high costs from jointly denoising large chunks of MASK tokens, much of which repeats similar feature representations for preceding context and many MASK positions. Systematic checks confirm both the redundancy and the tokens' role in supplying needed structural cues. The proposed position-preserving compression cuts repeated work and adds terminal-aware augmentation, yielding faster generation and a direct path to longer effective contexts without expanding the model's fixed input window.

Core claim

dLLMs spend substantial compute repeatedly processing the same preceding context and many MASK tokens that carry nearly identical features; position-preserving compression of these redundant MASK computations accelerates decoding, while terminal-aware augmentation of a protected terminal MASK token improves quality for block-wise models and supports context-folding-style long-context scaling for full-sequence models under fixed input-length limits.

What carries the argument

Position-preserving MASK token compression together with terminal-aware augmentation, which removes duplicate feature computations on MASK positions while retaining their placement and critical structural signals.

Load-bearing premise

That many MASK tokens share essentially the same feature representations so that dropping some of them loses little essential information for the denoising process.

What would settle it

Measure whether generation quality on standard benchmarks drops when the number of retained MASK tokens is halved versus the uncompressed baseline at identical denoising steps.

Figures

Figures reproduced from arXiv: 2605.18165 by Guohao Dai, Junyi Wu, Linfeng Zhang, Shaoqiu Zhang, Tianchen Zhao, Yu Wang.

Figure 1
Figure 1. Figure 1: Compared with autoregressive decoding, dLLMs repeatedly compute bidirectional attention [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: [MASK] Attention Redundancy. At￾tention maps during denoising show sparse and regular patterns: [MASK] tokens mainly aggre￾gate locally among themselves, while their inter￾actions with decoded tokens are relatively weak. 30 20 10 0 10 20 t-SNE Dimension 1 30 20 10 0 10 20 t-SNE Dimension 2 Feature Space Layer 31, Step 16 Prompt (25) Decoded (16) [MASK] (496) 40 30 20 10 0 10 20 30 t-SNE Dimension 1 30 20 1… view at source ↗
Figure 5
Figure 5. Figure 5: Terminal Position Signal. EOS pre￾diction during denoising. The red curve marks the [MASK] position with the highest EOS prob￾ability, and the green dotted line marks the final decoded EOS position. The early peak near the end indicates that the final RoPE position pro￾vides a terminal signal. 3.2 [MASK] tokens contain structural information [MASK] tokens carry RoPE positional information. Despite similar … view at source ↗
Figure 6
Figure 6. Figure 6: Overview of Elastic-dLLM. The observations show that [PITH_FULL_IMAGE:figures/full_fig_p005_6.png] view at source ↗
read the original abstract

Unlike autoregressive models, which generate one token at a time, dLLMs denoise a chunk of [MASK] tokens jointly and sample one or more tokens per step; despite enabling parallel decoding, this process incurs substantial computational cost due to the large chunk size of masked tokens. We observe that much of this cost is spent on repeatedly processing the preceding context and many [MASK] tokens with the same feature representations, indicating considerable computational redundancy. In this work, we revisit dLLM's redundancy from the perspective of [MASK] tokens. Through systematic analysis, we verify the redundancy of [MASK] tokens while revealing their critical role in providing structural information. Guided by these findings, we propose position-preserving [MASK] token compression and terminal-aware augmentation. By compressing redundant [MASK] computation, this approach accelerates decoding and further provides a natural extension toward context-folding-like long-context scaling under limited input-length constraints for full-sequence dLLMs such as LLaDA-8B-Instruct and LLaDA-1.5. Moreover, for block dLLMs such as LLaDA2.0-mini, it augments the context with a protected terminal [MASK] token to enhance generation quality with negligible overhead.

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

2 major / 2 minor

Summary. The manuscript proposes Elastic-dLLM for diffusion LLMs, observing that repeated processing of preceding context and many [MASK] tokens with identical feature representations creates computational redundancy. It introduces position-preserving [MASK] token compression to accelerate decoding while retaining structural information, enabling context-folding-like long-context scaling under input-length limits for full-sequence models (e.g., LLaDA-8B-Instruct), and terminal-aware [MASK] augmentation for block dLLMs (e.g., LLaDA2.0-mini) to improve quality with low overhead.

Significance. If the compression preserves output distributions and quality, the work offers a practical efficiency gain for non-autoregressive dLLM inference and long-context handling, addressing a key deployment bottleneck. The systematic analysis of [MASK] redundancy and their structural role is a clear strength, as is the distinction between full-sequence and block dLLM variants.

major comments (2)
  1. [§4] §4 (Method, position-preserving compression): The central claim that compression removes only redundant computation while retaining structural information for correct joint denoising rests on observed feature similarity. However, since attention and cross-token dependencies evolve across the noise schedule in iterative denoising, feature matches in early layers do not automatically imply invariance of the sampled conditional distribution over the chunk; direct verification (e.g., output distribution comparison or KL divergence on held-out generations) is required to support the claim.
  2. [Experimental results] Experimental section (results on LLaDA models): The abstract and description reference systematic analysis and acceleration, but load-bearing quantitative evidence—such as tables reporting exact speedup factors, perplexity deltas, or generation quality metrics (e.g., MAUVE or human eval) with vs. without compression—is needed to confirm that structural information is preserved and that the approach scales as claimed under limited input lengths.
minor comments (2)
  1. [Abstract] Abstract: The description of terminal-aware augmentation for block dLLMs could explicitly note the negligible overhead in terms of added tokens or FLOPs to clarify the efficiency claim.
  2. [Introduction] Notation: The distinction between full-sequence dLLMs and block dLLMs is introduced late; a brief upfront definition or table contrasting their [MASK] handling would improve readability.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for their constructive and detailed comments, which have helped us strengthen the manuscript. We address each major comment below and have revised the paper to incorporate additional validation and quantitative evidence as suggested.

read point-by-point responses

  1. Referee: [§4] §4 (Method, position-preserving compression): The central claim that compression removes only redundant computation while retaining structural information for correct joint denoising rests on observed feature similarity. However, since attention and cross-token dependencies evolve across the noise schedule in iterative denoising, feature matches in early layers do not automatically imply invariance of the sampled conditional distribution over the chunk; direct verification (e.g., output distribution comparison or KL divergence on held-out generations) is required to support the claim.

    Authors: We appreciate the referee's point that feature similarity alone does not guarantee invariance of the conditional distribution given the evolving nature of attention across denoising steps. While our systematic analysis identified redundancy through feature representations, we acknowledge the need for direct distributional verification. In the revised manuscript, we have added experiments that compute KL divergence between output distributions (with vs. without compression) on held-out generations across multiple noise levels. These results confirm that the sampled distributions remain closely aligned, supporting the claim that structural information is preserved for joint denoising. The new analysis is included in Section 4. revision: yes

  2. Referee: [Experimental results] Experimental section (results on LLaDA models): The abstract and description reference systematic analysis and acceleration, but load-bearing quantitative evidence—such as tables reporting exact speedup factors, perplexity deltas, or generation quality metrics (e.g., MAUVE or human eval) with vs. without compression—is needed to confirm that structural information is preserved and that the approach scales as claimed under limited input lengths.

    Authors: We agree that explicit quantitative tables are necessary to make the claims of acceleration and quality preservation fully load-bearing. The original manuscript emphasized systematic analysis of [MASK] redundancy and its structural role, but we have now expanded the experimental section with detailed tables. These report exact speedup factors, perplexity deltas, MAUVE scores, and other quality metrics comparing compressed and baseline variants on LLaDA-8B-Instruct and LLaDA2.0-mini. Additional results demonstrate the context-folding-like scaling under input-length constraints. The tables and accompanying discussion have been added to the revised experimental section. revision: yes

Circularity Check

0 steps flagged

No significant circularity; proposal grounded in empirical observations of redundancy

full rationale

The paper's central proposal for position-preserving [MASK] token compression and terminal-aware augmentation is derived directly from systematic analysis and stated observations that many [MASK] tokens exhibit identical feature representations while still providing structural information. No load-bearing step reduces by construction to a fitted parameter renamed as prediction, a self-citation chain, or a self-definitional equivalence. The method is presented as an engineering response to verified computational redundancy in dLLM decoding, with the derivation chain remaining self-contained against external benchmarks such as the observed feature similarity and the explicit goal of accelerating joint denoising without altering the underlying diffusion process.

Axiom & Free-Parameter Ledger

0 free parameters · 1 axioms · 0 invented entities

The approach rests on empirical observations of redundancy and the structural role of MASK tokens; no free parameters or invented entities are mentioned.

axioms (1)
  • domain assumption MASK tokens exhibit substantial computational redundancy while retaining critical structural and positional information.
    Derived from systematic analysis of dLLM decoding process.

pith-pipeline@v0.9.0 · 5764 in / 1102 out tokens · 35797 ms · 2026-05-20T13:09:36.704867+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

38 extracted references · 38 canonical work pages · 10 internal anchors

  1. [1]

    Chiu, Zhihan Yang, Zhixuan Qi, Jiaqi Han, Subham Sekhar Sahoo, and V olodymyr Kuleshov

    Marianne Arriola, Aaron Gokaslan, Justin T. Chiu, Zhihan Yang, Zhixuan Qi, Jiaqi Han, Subham Sekhar Sahoo, and V olodymyr Kuleshov. Block Diffusion: Interpolating Between Autoregressive and Diffusion Language Models. InICLR, 2025

    work page 2025

  2. [2]

    arXiv preprint arXiv:2408.07055 , year=

    Yushi Bai, Jiajie Zhang, Xin Lv, Linzhi Zheng, Siqi Zhu, Lei Hou, Yuxiao Dong, Jie Tang, and Juanzi Li. Longwriter: Unleashing 10,000+ word generation from long context llms.arXiv preprint arXiv:2408.07055, 2024

    work page arXiv 2024

  3. [3]

    Tiwei Bie, Maosong Cao, Kun Chen, Lun Du, Mingliang Gong, Zhuochen Gong, Yanmei Gu, Jiaqi Hu, Zenan Huang, Zhenzhong Lan, et al. Llada2. 0: Scaling up diffusion language models to 100b.arXiv preprint arXiv:2512.15745, 2025

    work page internal anchor Pith review Pith/arXiv arXiv 2025

  4. [4]

    PyramidKV: Dynamic KV Cache Compression based on Pyramidal Information Funneling

    Zefan Cai, Yichi Zhang, Bofei Gao, Yuliang Liu, Tianyu Liu, Keming Lu, Wayne Xiong, Yue Dong, Baobao Chang, Junjie Hu, and Wen Xiao. PyramidKV: Dynamic KV cache compression based on pyramidal information funneling.arXiv preprint arXiv:2406.02069, 2024

    work page internal anchor Pith review Pith/arXiv arXiv 2024

  5. [5]

    Mark Chen, Jerry Tworek, Heewoo Jun, Qiming Yuan, Henrique Ponde de Oliveira Pinto, Jared Kaplan, Harri Edwards, Yuri Burda, Nicholas Joseph, Greg Brockman, Alex Ray, Raul Puri, Gretchen Krueger, Michael Petrov, Heidy Khlaaf, Girish Sastry, Pamela Mishkin, Brooke Chan, Scott Gray, Nick Ryder, Mikhail Pavlov, Alethea Power, Lukasz Kaiser, Mohammad Bavarian...

    work page 2021

  6. [6]

    Dpad: Efficient diffusion language models with suffix dropout.arXiv preprint arXiv:2508.14148,

    Xinhua Chen, Sitao Huang, Cong Guo, Chiyue Wei, Yintao He, Jianyi Zhang, Hai "Hellen" Li, and Yiran Chen. Dpad: Efficient diffusion language models with suffix dropout, 2025. URL https://arxiv.org/abs/2508.14148

    work page arXiv 2025

  7. [7]

    Flashblock: Attention caching for efficient long-context block diffusion.arXiv preprint arXiv:2602.05305, 2026

    Zhuokun Chen, Jianfei Cai, and Bohan Zhuang. Flashblock: Attention caching for efficient long-context block diffusion.arXiv preprint arXiv:2602.05305, 2026

    work page arXiv 2026

  8. [8]

    Training Verifiers to Solve Math Word Problems

    Karl Cobbe, Vineet Kosaraju, Mohammad Bavarian, Mark Chen, Heewoo Jun, Lukasz Kaiser, Matthias Plappert, Jerry Tworek, Jacob Hilton, Reiichiro Nakano, et al. Training verifiers to solve math word problems.arXiv preprint arXiv:2110.14168, 2021

    work page internal anchor Pith review Pith/arXiv arXiv 2021

  9. [9]

    Flashdlm: Accelerating diffusion language model inference via efficient kv caching and guided diffusion.arXiv preprint arXiv:2505.21467, 2025

    Zhanqiu Hu, Jian Meng, Yash Akhauri, Mohamed S Abdelfattah, Jae-sun Seo, Zhiru Zhang, and Udit Gupta. Flashdlm: Accelerating diffusion language model inference via efficient kv caching and guided diffusion.arXiv preprint arXiv:2505.21467, 2025

    work page arXiv 2025

  10. [10]

    D2 cache: Accelerating diffusion-based llms via dual adaptive caching.arXiv preprint arXiv:2509.23094, 2025

    Yuchu Jiang, Yue Cai, Xiangzhong Luo, Jiale Fu, Jiarui Wang, Chonghan Liu, and Xu Yang. D2 cache: Accelerating diffusion-based llms via dual adaptive caching.arXiv preprint arXiv:2509.23094, 2025

    work page arXiv 2025

  11. [11]

    Inan, Lukas Wutschitz, Yanzhi Chen, Robert Sim, and Saravan Rajmohan

    Minki Kang, Wei-Ning Chen, Dongge Han, Huseyin A. Inan, Lukas Wutschitz, Yanzhi Chen, Robert Sim, and Saravan Rajmohan. Acon: Optimizing context compression for long-horizon llm agents.arXiv preprint arXiv:2510.00615, 2025. doi: 10.48550/arXiv.2510.00615

    work page doi:10.48550/arxiv.2510.00615 2025

  12. [12]

    SnapKV: LLM knows what you are looking for before generation

    Yuhong Li, Yingbing Huang, Bowen Yang, Bharat Venkitesh, Acyr Locatelli, Hanchen Ye, Tianle Cai, Patrick Lewis, and Deming Chen. SnapKV: LLM knows what you are looking for before generation. InNeurIPS, 2024

    work page 2024

  13. [13]

    Let’s verify step by step

    Hunter Lightman, Vineet Kosaraju, Yuri Burda, Harrison Edwards, Bowen Baker, Teddy Lee, Jan Leike, John Schulman, Ilya Sutskever, and Karl Cobbe. Let’s verify step by step. InThe twelfth international conference on learning representations, 2023

    work page 2023

  14. [14]

    dllm-cache: Accelerating diffusion large language models with adaptive caching.arXiv preprint arXiv:2506.06295, 2025

    Zhiyuan Liu, Yicun Yang, Yaojie Zhang, Junjie Chen, Chang Zou, Qingyuan Wei, Shaobo Wang, and Linfeng Zhang. dllm-cache: Accelerating diffusion large language models with adaptive caching.arXiv preprint arXiv:2506.06295, 2025. 10

    work page arXiv 2025

  15. [15]

    Zichang Liu, Aditya Desai, Fangshuo Liao, Weitao Wang, Victor Xie, Zhaozhuo Xu, Anastasios Kyrillidis, and Anshumali Shrivastava. Scissorhands: Exploiting the persistence of impor- tance hypothesis for llm kv cache compression at test time.Advances in Neural Information Processing Systems, 36:52342–52364, 2023

    work page 2023

  16. [16]

    KIVI: A Tuning-Free Asymmetric 2bit Quantization for KV Cache

    Zirui Liu, Jiayi Yuan, Hongye Jin, Shaochen Zhong, Zhaozhuo Xu, Vladimir Braverman, Beidi Chen, and Xia Hu. Kivi: A tuning-free asymmetric 2bit quantization for kv cache.arXiv preprint arXiv:2402.02750, 2024

    work page internal anchor Pith review Pith/arXiv arXiv 2024

  17. [17]

    Discrete Diffusion Modeling by Estimating the Ratios of the Data Distribution

    Aaron Lou, Chenlin Meng, and Stefano Ermon. Discrete Diffusion Modeling by Estimating the Ratios of the Data Distribution. InICML, 2024

    work page 2024

  18. [18]

    dkv-cache: The cache for diffusion language models.arXiv preprint arXiv:2505.15781,

    Xinyin Ma, Runpeng Yu, Gongfan Fang, and Xinchao Wang. dKV-Cache: The Cache for Diffusion Language Models.arXiv preprint arXiv:2505.15781, 2025

    work page arXiv 2025

  19. [19]

    A diverse corpus for evaluating and developing english math word problem solvers

    Shen-yun Miao, Chao-Chun Liang, and Keh-Yih Su. A diverse corpus for evaluating and developing english math word problem solvers. InProceedings of the 58th Annual Meeting of the Association for Computational Linguistics, pages 975–984, 2020

    work page 2020

  20. [20]

    Attention is all you need for kv cache in diffusion llms.arXiv preprint arXiv:2510.14973, 2025

    Quan Nguyen-Tri, Mukul Ranjan, and Zhiqiang Shen. Attention is all you need for kv cache in diffusion llms.arXiv preprint arXiv:2510.14973, 2025

    work page arXiv 2025

  21. [21]

    Large Language Diffusion Models

    Shen Nie, Fengqi Zhu, Zebin You, Xiaolu Zhang, Jingyang Ou, Jun Hu, Jun Zhou, Yankai Lin, Ji-Rong Wen, and Chongxuan Li. Large Language Diffusion Models.arXiv preprint arXiv:2502.09992, 2025

    work page internal anchor Pith review Pith/arXiv arXiv 2025

  22. [22]

    Tinyzero

    Jiayi Pan, Junjie Zhang, Xingyao Wang, Lifan Yuan, Hao Peng, and Alane Suhr. Tinyzero. https://github.com/Jiayi-Pan/TinyZero, 2025. Accessed: 2025-01-24

    work page 2025

  23. [23]

    Chiu, Alexander Rush, and V olodymyr Kuleshov

    Subham Sekhar Sahoo, Marianne Arriola, Yair Schiff, Aaron Gokaslan, Edgar Marroquin, Justin T. Chiu, Alexander Rush, and V olodymyr Kuleshov. Simple and Effective Masked Diffusion Language Models. InNeurIPS, 2024

    work page 2024

  24. [24]

    Scaling long-horizon llm agent via context-folding.arXiv preprint arXiv:2510.11967, 2025

    Weiwei Sun, Miao Lu, Zhan Ling, Kang Liu, Xuesong Yao, Yiming Yang, and Jiecao Chen. Scaling long-horizon llm agent via context-folding.arXiv preprint arXiv:2510.11967, 2025. doi: 10.48550/arXiv.2510.11967

    work page doi:10.48550/arxiv.2510.11967 2025

  25. [25]

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

    Jiaming Tang, Yilong Zhao, Kan Zhu, Guangxuan Xiao, Baris Kasikci, and Song Han. Quest: Query-aware sparsity for efficient long-context llm inference.arXiv preprint arXiv:2406.10774, 2024

    work page internal anchor Pith review Pith/arXiv arXiv 2024

  26. [26]

    SparseD: Sparse Attention for Diffusion Language Models

    Zeqing Wang, Gongfan Fang, Xinyin Ma, Xingyi Yang, and Xinchao Wang. SparseD: Sparse Attention for Diffusion Language Models. InICLR, 2026

    work page 2026

  27. [27]

    Fast-dLLM: Training-free Acceleration of Diffusion LLM by Enabling KV Cache and Parallel Decoding

    Chengyue Wu, Hao Zhang, Shuchen Xue, Zhijian Liu, Shizhe Diao, Ligeng Zhu, Ping Luo, Song Han, and Enze Xie. Fast-dLLM: Training-free Acceleration of Diffusion LLM by Enabling KV Cache and Parallel Decoding.arXiv preprint arXiv:2505.22618, 2025

    work page internal anchor Pith review Pith/arXiv arXiv 2025

  28. [28]

    Quantcache: Adaptive importance-guided quantization with hierarchical latent and layer caching for video generation

    Junyi Wu, Zhiteng Li, Zheng Hui, Yulun Zhang, Linghe Kong, and Xiaokang Yang. Quantcache: Adaptive importance-guided quantization with hierarchical latent and layer caching for video generation. InProceedings of the IEEE/CVF International Conference on Computer Vision, pages 15035–15044, 2025

    work page 2025

  29. [29]

    Resum: Unlocking long-horizon search intelligence via context summarization.arXiv preprint arXiv:2509.13313, 2025

    Xixi Wu, Kuan Li, Yida Zhao, Liwen Zhang, Litu Ou, Huifeng Yin, Zhongwang Zhang, Xinmiao Yu, Dingchu Zhang, Yong Jiang, Pengjun Xie, Fei Huang, Minhao Cheng, Shuai Wang, Hong Cheng, and Jingren Zhou. Resum: Unlocking long-horizon search intelligence via context summarization.arXiv preprint arXiv:2509.13313, 2025. doi: 10.48550/arXiv.2509.13313

    work page doi:10.48550/arxiv.2509.13313 2025

  30. [30]

    Infllm: Unveiling the intrinsic capacity of llms for under- standing extremely long sequences with training-free memory.arXiv, 2024

    Chaojun Xiao, Pengle Zhang, Xu Han, Guangxuan Xiao, Yankai Lin, Zhengyan Zhang, Zhiyuan Liu, Song Han, and Maosong Sun. Infllm: Unveiling the intrinsic capacity of llms for under- standing extremely long sequences with training-free memory.arXiv, 2024

    work page 2024

  31. [31]

    Efficient streaming language models with attention sinks

    Guangxuan Xiao, Yuandong Tian, Beidi Chen, Song Han, and Mike Lewis. Efficient streaming language models with attention sinks. InICLR, 2024. 11

    work page 2024

  32. [32]

    Dream 7B: Diffusion Large Language Models

    Jiacheng Ye, Zhihui Xie, Lin Zheng, Jiahui Gao, Zirui Wu, Xin Jiang, Zhenguo Li, and Lingpeng Kong. Dream 7B: Diffusion Large Language Models.arXiv preprint arXiv:2508.15487, 2025

    work page internal anchor Pith review Pith/arXiv arXiv 2025

  33. [33]

    Quant- dllm: Post-training extreme low-bit quantization for diffusion large language models.arXiv preprint arXiv:2510.03274, 2025

    Tianao Zhang, Zhiteng Li, Xianglong Yan, Haotong Qin, Yong Guo, and Yulun Zhang. Quant- dllm: Post-training extreme low-bit quantization for diffusion large language models.arXiv preprint arXiv:2510.03274, 2025

    work page arXiv 2025

  34. [34]

    H2O: Heavy-hitter oracle for efficient generative inference of large language models

    Zhenyu Zhang, Ying Sheng, Tianyi Zhou, Tianlong Chen, Lianmin Zheng, Ruisi Cai, Zhao Song, Yuandong Tian, Christopher Re, Clark Barrett, Zhangyang Wang, and Beidi Chen. H2O: Heavy-hitter oracle for efficient generative inference of large language models. InNeurIPS, 2023

    work page 2023

  35. [35]

    d1: Scaling reasoning in diffusion large language models via reinforcement learning.arXiv preprint arXiv:2504.12216, 2025

    Siyan Zhao, Devaansh Gupta, Qinqing Zheng, and Aditya Grover. d1: Scaling reasoning in diffusion large language models via reinforcement learning.arXiv preprint arXiv:2504.12216, 2025

    work page arXiv 2025

  36. [36]

    Instruction-Following Evaluation for Large Language Models

    Jeffrey Zhou, Tianjian Lu, Swaroop Mishra, Siddhartha Brahma, Sujoy Basu, Yi Luan, Denny Zhou, and Le Hou. Instruction-following evaluation for large language models.arXiv preprint arXiv:2311.07911, 2023

    work page internal anchor Pith review Pith/arXiv arXiv 2023

  37. [37]

    LLaDA 1.5: Variance-Reduced Preference Optimization for Large Language Diffusion Models

    Fengqi Zhu, Rongzhen Wang, Shen Nie, Xiaolu Zhang, Chunwei Wu, Jun Hu, Jun Zhou, Jianfei Chen, Yankai Lin, Ji-Rong Wen, et al. Llada 1.5: Variance-reduced preference optimization for large language diffusion models.arXiv preprint arXiv:2505.19223, 2025

    work page internal anchor Pith review Pith/arXiv arXiv 2025

  38. [38]

    Es-dllm: Efficient inference for diffusion large language models by early-skipping.arXiv preprint arXiv:2603.10088, 2026

    Zijian Zhu, Fei Ren, Zhanhong Tan, and Kaisheng Ma. Es-dllm: Efficient inference for diffusion large language models by early-skipping.arXiv preprint arXiv:2603.10088, 2026. 12

    work page arXiv 2026