Pith. sign in

REVIEW 4 major objections 5 minor 25 references

Ripple-Pivot Search: Active Parallel Decoding for Diffusion Large Language Models

T0 review · 4 major / 5 minor · reviewed 2026-08-16 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗ Review JSON

Pith's one-line read This paper claims that committing a mid-entropy pivot position first—rather than the most confident one—unlocks 4–10x faster diffusion LLM decoding while largely preserving accuracy.

desk verdict A real extension—searching what to commit, not just where—with an honest but GSM8K-tied empirical case; deserves review, but ask for code and a held-out hyperparameter check. read the letter →

arxiv 2608.11742 v1 pith:GHB6JKG3 submitted 2026-08-12 cs.CL

Yushi Ye , Xu Chen , Haoyun Jiang , Jinsong Lan , Haihong Tang , Bo Han , Ivor Tsang , Yanfeng Wang
show 2 more authors
Bo Zheng Jiangchao Yao
This is my paper · ORCID
classification cs.CL
keywords diffusionlanguagemodelsparalleldecodingrippleeffectmid-entropypivotlookaheadevaluationtraining-freeaccelerationKVcachingreasoningandcodegeneration
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

This paper is trying to establish that decoding order in diffusion large language models can be actively scheduled for speed: committing a single 'pivot' token whose predictive distribution sits in the mid-entropy regime—neither already certain nor too diffuse—reduces uncertainty at the other masked positions, so later steps can unmask more tokens in parallel. If true, a training-free scheduler can cut wall-clock inference time by 4–10x on reasoning and code-generation tasks while preserving or even improving accuracy, and by up to 18x when combined with KV caching. The proposed Ripple-Pivot Search (RPS) makes this concrete by searching for the best pivot position and, unlike earlier lookahead methods, also searching over which token to assign to that pivot instead of always taking the model's top-1 prediction. The paper reports this speedup consistently across three models and four benchmarks, with the largest accuracy gain (up to 5.49%) coming on code generation relative to the previous lookahead baseline.

What carries the argument

The central mechanism is the ripple effect, the observed drop in entropy at remaining masked positions after a mid-entropy pivot is committed. The algorithm that carries the argument is RPS: pivot selection maximizes truncated entropy $-\sum_{v\in T_i} p_i(v)\log p_i(v)$ over positions whose retained top-$k_{\max}$ probability mass $\mu_i$ is at least $\tau_{\mathrm{pivot}}$, isolating the mid-entropy regime; lookahead scoring then builds one branch per plausible token plus a [MASK] anchor in a single packed forward pass, scoring branch $c$ by $-\frac{1}{|\mathcal{M}|-1}\sum_{i\in\mathcal{M}\setminus\{i^\star\}} H(p_i^c) + \lambda \log p_{\mathrm{anchor}}(c)$. Proposition 1 (entropy-certified parallelism) gives the entropy term a direct speed meaning: the mean downstream entropy lower-bounds the number of next-step commits eligible under a confidence threshold $\tau$, so minimizing it certifies more parallelism; Proposition 2 (plausibility-adjusted selection margin) shows the plausibility term is a Lagrangian safeguard requiring less plausible tokens to compensate with a proportionally larger entropy reduction.

What would settle it

Recompute the oracle analysis behind Figure 1 on a held-out diffusion LM and task: if the largest mean entropy reductions at other masked positions do not occur when mid-entropy pivots are committed, or if RPS's speedup collapses when its hyperparameters are transferred without retuning, the central claim fails.

Watch

Extended reading notes

Core claim

The paper's central claim is the ripple effect: in masked diffusion LLM decoding, committing a deliberately chosen mid-entropy position reduces predictive entropy at the other still-masked positions more than committing a confident or a diffuse position does. RPS operationalizes this with a two-stage search: pivot selection picks the masked position that maximizes truncated entropy subject to a retained-probability-mass threshold, and lookahead scoring evaluates candidate token assignments jointly in a single forward pass with an isolated attention mask, scoring each candidate by the mean downstream entropy it induces plus a plausibility term that guards against degenerate tokens; the pivot is committed only when the best candidate beats leaving it masked. On LLaDA-8B-Instruct, Dream-v0-Instruct-7B, and LLaDA-1.5 across GSM8K, MATH500, HumanEval, and MBPP, the paper reports 4.24–9.80x tokens-per-second speedup over the one-token-per-step decoder with largely preserved accuracy, up to 5.49% accuracy improvement over the LoPA lookahead baseline at comparable throughput, and up to 17.82x TPS speedup when combined with KV caching.

Load-bearing premise

The method rests on the empirical ripple effect—that committing a mid-entropy position first reliably reduces uncertainty at the other masked positions—a pattern observed on a subset of GSM8K rather than proved, so if that pattern or the fixed hyperparameters do not transfer to a new model or task, the speedups will not either.

Editorial extensions

If this is right

  • RPS achieves 4.24–9.80x tokens-per-second speedup over the one-token-per-step Default decoder across three diffusion LMs and four reasoning and code benchmarks while largely preserving accuracy.
  • RPS improves accuracy over the previous lookahead baseline LoPA by up to 5.49%, with the largest gains on HumanEval (4.27% on LLaDA and 5.49% on Dream) at comparable or higher throughput.
  • The plausibility safeguard is load-bearing: removing it ($\lambda = 0$) costs 2.1% accuracy on GSM8K, while $\lambda \in [0.1, 0.5]$ forms a stable plateau with little speed change.
  • Combined with KV prefix caching, RPS reaches up to 17.82x TPS speedup with less than 0.5% accuracy change, showing iteration-level and per-forward optimizations are complementary.
  • RPS remains the strongest quality–efficiency trade-off when generation length varies from 128 to 512 tokens, including on a short-length HumanEval setting where aggressive baselines degrade.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • If the ripple effect is a general property of masked diffusion models, the same mid-entropy pivot selection could accelerate token-parallel generative models beyond language, such as masked image or audio diffusion transformers.
  • The decision to search over token assignments (not just positions) suggests a transferable principle: any lookahead scheduler that fixes top-1 assignments could gain from evaluating a small candidate set with a plausibility regularizer, which could be tested by inserting this step into WINO or ETE.
  • A learned predictor of downstream entropy reduction could replace the fixed $\lambda$ and $\tau_{\mathrm{pivot}}$, making the scheduler fully adaptive; the paper's own limitation section notes $\lambda$ still needs per-task selection within a plateau.
  • The reported 85% non-top-1 oracle rate at mid-entropy pivots, if it generalizes, implies that confidence-maximizing schedulers may systematically defer the most informative tokens; one could test this by comparing RPS's commit order against confidence-based orders on longer reasoning chains.
Share X Bluesky LinkedIn Reddit HN

Signed reviews

No signed human review yet.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

4 major / 5 minor

Summary. The manuscript introduces Ripple-Pivot Search (RPS), a training-free decoding scheduler for diffusion large language models. RPS first selects a 'mid-entropy' pivot position using truncated entropy under a probability-mass threshold, then evaluates a small set of candidate token assignments in a single packed lookahead forward pass, scoring each branch by downstream mean entropy plus a plausibility regularizer. The authors report 4–10x wall-clock speedups over a standard decoder across three dLLMs and four reasoning/code benchmarks, accuracy gains of up to 5.49% over the LoPA lookahead baseline, and further gains when combined with KV caching. The paper also proves two propositions about the lookahead objective: one lower-bounding next-step eligible commit counts by mean downstream entropy, and one characterizing the plausibility-adjusted selection margin.

Significance. If the empirical claims hold, RPS is a practically useful, training-free accelerator for diffusion language model inference, and the idea of jointly optimizing 'where' and 'what' to commit is a reasonable contribution over confidence- or entropy-based schedulers and the LoPA/ETE lookahead family. The paper is also commendably explicit about the conservativeness of its theory and about the empirical, non-theoretical status of the ripple-effect motivation. However, the central validation is weakened by configuration selection on the same benchmark that is then reported as a headline result, by the lack of any variance or significance information, and by the fact that the motivating phenomenon is only demonstrated on a subset of GSM8K. These issues are addressable with additional experiments rather than being fundamental flaws in the method's design.

major comments (4)
  1. [§5.2, Table 3 and Table 1] The hyperparameters k_max=10, r=0.1, τ_pivot=0.9/0.95, and λ∈[0.1,0.5] are selected via ablations on LLaDA GSM8K, and the same configuration is then evaluated on GSM8K in Table 1. The GSM8K rows are therefore not an independent test of the configuration, and the reported GSM8K speedups partly re-describe the configuration search. Please report results with hyperparameters chosen on a held-out validation split or a different benchmark, and provide error bars over multiple seeds.
  2. [§1 and Appendix E] The ripple-effect motivation rests on Fig. 1, an oracle analysis conducted on a subset of GSM8K, and Appendix E explicitly states that the characterization is 'empirical and qualitative rather than theoretically derived' and is not shown to quantitatively generalize. This is load-bearing because the mid-entropy pivot-selection rule is motivated entirely by this pattern. Please provide quantitative evidence that the mid-entropy advantage holds across tasks and models, for example by reporting entropy-reduction curves or pivot-strategy ablations on held-out benchmarks.
  3. [Tables 1 and 2, §5.1] No variance or statistical significance information is reported. Accuracy differences such as the 1.22% HumanEval gain over Default and the 4.27%/5.49% gains over LoPA on HumanEval could be within sampling noise, especially on HumanEval's 164 problems. Please report multiple seeds or bootstrap confidence intervals, and state the number of evaluation samples per benchmark.
  4. [§4.2, Proposition 1, Eq. (5)] Proposition 1 lower-bounds the number of positions eligible for commitment in the next step, but the abstract and §5.1 headline end-to-end NFE and TPS speedups; the manuscript does not establish a formal link between the certified bound and end-to-end wall-clock speedup. Since the authors explicitly acknowledge that the bound 'does not directly bound end-to-end NFE,' either connect N_τ(c) to a per-step cost model or present Proposition 1 explicitly as heuristic motivation rather than as a speedup guarantee.
minor comments (5)
  1. [§1, contributions bullet] There is a typo in the second contribution bullet: 'Curicially' should be 'Critically'.
  2. [Figure 3] The three panels of Figure 3 appear to be identical in the rendered manuscript; please ensure the left, middle, and right panels show the intended distinct plots.
  3. [Table 2] The column headers 'Acc', 'NFE Sp.', and 'TPS Sp.' run together in the rendered table (e.g., '30.491.00×1.00×'); add clear separators between the accuracy and speedup values.
  4. [§5.3, Table 4] The forward-pass breakdown in Table 4 is reported for LLaDA on GSM8K only; please state this restriction in the caption or text.
  5. [General] No code or reproduction details are provided. Releasing the implementation and evaluation scripts, including seeds, would substantially improve reproducibility of the reported speedups and accuracy numbers.

Circularity Check

0 steps flagged · score 2.0 of 10

No construct-level circularity; measured speedups on external benchmarks carry the central claim, with only minor self-citations and a mild GSM8K tuning caveat.

full rationale

The derivation chain is not circular at the level of construction. Propositions 1 and 2 are proven from the definitions of predictive entropy, confidence thresholds, and the RPS scoring rule; they do not assume the speedups they are used to interpret. The 4-10x TPS/NFE speedups and accuracy comparisons are wall-clock and accuracy measurements against Default, Confidence, KLASS, EB-Sampler, WINO, and LoPA on GSM8K, MATH500, HumanEval, and MBPP, so the headline results are not derived from the scoring objective by an equality. The only mildly concerning spots are the selection of k_max=10, r=0.1, tau_pivot in {0.9,0.95}, and lambda in [0.1,0.5] by ablations on LLaDA GSM8K (Section 5.2, Table 3), after which the same configuration is reported on GSM8K in Table 1; and the motivating ripple-effect analysis in Section 1 is conducted on a subset of GSM8K. Appendix E explicitly concedes both points, calling the ripple-effect characterization 'empirical and qualitative rather than theoretically derived' and noting that lambda 'still requires per-task selection within that range'. These are generalization and selection-bias risks, not reductions by construction: the HumanEval, MATH500, and MBPP rows use untuned parameters and remain external checks. The self-citations (WINO in Hong et al., 2025; rejection mixing in Ye et al., 2026) appear only as baselines or related work and do not carry the argument. No self-citation chain forces the reported gains.

Assumptions & free parameters 5 free parameters · 4 assumptions · 0 invented entities

The central claim rests on an empirically observed ripple effect (not derived), hand-set hyperparameters tuned on GSM8K, and the modeling assumption that packed lookahead branches faithfully simulate individual assignments. No new physical or architectural entities are introduced; the named ripple effect is a label for an empirical uncertainty-reduction pattern. The appendix propositions are proven from standard entropy bounds.

free parameters (5)
  • plausibility weight lambda = 0.3 (selected from [0.1,0.5] via GSM8K ablation)
    Eq. (4) balances downstream entropy reduction against anchor-branch plausibility; the interval is chosen in §5.2 on LLaDA GSM8K and then applied broadly, so the GSM8K main result partially reflects tuning on the same benchmark.
  • probability-mass threshold tau_pivot = 0.9 (LLaDA), 0.95 (Dream)
    Eq. (3) excludes high-entropy positions; set per model family based on ablations in Table 3.
  • candidate budget k_max = 10
    Truncated support size in Eq. (3); Table 3 shows strong sensitivity, with k_max=20 dropping accuracy to 74.91.
  • reachability ratio r = 0.1
    Candidate set threshold C={v: p_i*(v) >= r * P_max_i*}; robust across 0.05-0.3 in Table 3.
  • standard commit threshold tau = 0.9 (assumed, not explicitly reported for RPS)
    Algorithm 1 step 6 uses a confidence-aware commit threshold tau; the paper follows Fast-dLLM's Confidence setting, but the RPS-specific tau is not called out in the main text.
assumptions (4)
  • domain assumption The masked diffusion LLM's per-position predictive distributions are reliable and reusable across steps, including logits carried over from lookahead branches.
    Algorithm 1 reuses the winner branch's logits for the next step, and the packed attention-mask lookahead assumes each branch faithfully simulates model predictions under a single candidate assignment.
  • ad hoc to paper The ripple effect: committing a mid-entropy pivot materially reduces downstream masked-position entropy.
    Stated as an empirical finding in §1 and Fig. 1 (right), but Appendix E concedes it is qualitative and derived from a GSM8K subset, not theoretically established.
  • ad hoc to paper Top-k truncation with k_max and reachability ratio r retains the potentially correct tokens at the pivot.
    Eq. (3) and the candidate set definition assume the correct token lies within the truncated support; Table 3 shows k_max=20 hurts accuracy, indicating weak candidates pollute the set.
  • standard math Standard properties of Shannon entropy used in the two proposition proofs.
    Appendix B relies on entropy decomposition and monotonicity of the binary entropy function on [1/2,1]; these are unproblematic.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Ripple-Pivot Search: Active Parallel Decoding for Diffusion Large Language Models." pith.science (2026). https://pith.science/paper/GHB6JKG3

@misc{pith2026260811742,
  author       = {Pith},
  title        = {Pith review of: Ripple-Pivot Search: Active Parallel Decoding for Diffusion Large Language Models},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/GHB6JKG3}},
  note         = {Machine review of arXiv:2608.11742}
}
abstract

Diffusion Large Language Models (dLLMs) have emerged as a competitive alternative to autoregressive language models, offering the potential for substantially faster inference through parallel decoding. Existing parallel decoding schedulers typically commit positions only after they meet a per-position criterion, overlooking how early commitments may benefit subsequent decoding. We identify a ripple effect in dLLM decoding: proactively committing a mid-entropy pivot position can induce a pronounced reduction in uncertainty across the remaining masked positions. This uncertainty reduction allows subsequent steps to unmask more tokens in parallel, thereby accelerating the overall decoding process. To exploit the ripple effect, we propose Ripple-Pivot Search (RPS), a novel training-free decoding method that seeks mid-entropy positions as promising candidate pivots (where to decode), and determines their token assignment that yields the greatest downstream benefit via lookahead evaluation (what to decode). Across 3 dLLMs and 4 reasoning and code-generation benchmarks, RPS achieves 4-10$\times$ wall-clock speedup over the standard decoder while preserving generation quality, and improves accuracy over the previous lookahead baseline by up to 5.49% while delivering higher throughput in most settings. When integrated with KV caching, RPS further achieves up to 18$\times$ wall-clock speedup over the standard decoder.

Figures

Figures reproduced from arXiv: 2608.11742 by the authors.

Figure 1
Figure 1. Left: Comparison of parallel decoding schedulers. Right: Starting from a partially decoded block, we commit each remaining masked position to its oracle token and measure the resulting entropy reduction ∆Hi at other masked positions (y-axis) against the pre-commit entropy Hbefore of the resolved position (x-axis). Colors indicate the oracle token’s rank under the model distribution, and the shaded band marks the mid… view at source ↗
Figure 2
Figure 2. Overview of RPS. Left: One-step example of RPS decoding. The upper region illustrates pivot selection, while the lower region shows lookahead scoring. Right: Customized attention mask for lookahead forward pass. The packed sequence contains the shared context with candidate branches. Light-blue cells indicate allowed attention and blank cells indicate blocked attention. 4 METHOD 4.1 RIPPLE-PIVOT SEARCH Our RPS intro… view at source ↗
Figure 3
Figure 3. Left: Accuracy and TPS speedup of RPS on GSM8K with LLaDA under different plau￾sibility weights λ. Middle: Accuracy and TPS trade-off of different pivot-selection and scoring strategies on GSM8K with LLaDA Right: Failure-length comparison on HumanEval for Dream and LLaDA. For cases where exactly one of LoPA and RPS succeeds, we report the fraction for which the failing completion contains fewer non-empty source line… view at source ↗
Figures from the paper (1 more)
Figure 4
Figure 4. Figure 4: Left: Decoding snapshots of LoPA and RPS on HumanEval/57 (monotonic). Coloured tokens denote newly committed tokens, underlined tokens denote lookahead commits, and “■” denotes still-masked positions. Right: TPS speedup relative to Default and Confidence with and witho…

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

25 extracted references · 3 canonical work pages

  1. [1]

    Nye, Maarten Bosma, Henryk Michalewski, David Do- han, Ellen Jiang, Carrie J

    Jacob Austin, Augustus Odena, Maxwell I. Nye, Maarten Bosma, Henryk Michalewski, David Do- han, Ellen Jiang, Carrie J. Cai, Michael Terry, Quoc V . Le, and Charles Sutton. Program synthesis with large language models.CoRR, abs/2108.07732,

    arXiv

  2. [4]

    Llada2.0: Scaling up diffusion language models to 100b

    Tiwei Bie, Maosong Cao, Kun Chen, Lun Du, Mingliang Gong, Zhuochen Gong, Yanmei Gu, Jiaqi Hu, Zenan Huang, Zhenzhong Lan, Chengxi Li, Chongxuan Li, Jianguo Li, Zehuan Li, Huabin Liu, Lin Liu, Guoshan Lu, Xiaocheng Lu, Yuxin Ma, Jianfeng Tan, Lanning Wei, Ji-Rong Wen, Yipeng Xing, Xiaolu Zhang, Junbo Zhao, Da Zheng, Jun Zhou, Junlin Zhou, Zhanchao Zhou, Li...

    arXiv

  3. [5]

    Tom B. Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared Kaplan, Prafulla Dhari- wal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, Sandhini Agarwal, Ariel Herbert-V oss, Gretchen Krueger, Tom Henighan, Rewon Child, Aditya Ramesh, Daniel M. Ziegler, Jeffrey Wu, Clemens Winter, Christopher Hesse, Mark Chen, Eric Sigler, Mateusz Lit...

    work page 2020

  4. [7]

    SDAR: A synergistic diffusion- autoregression paradigm for scalable sequence generation.CoRR, abs/2510.06303,

    Shuang Cheng, Yihan Bian, Dawei Liu, Linfeng Zhang, Qian Yao, Zhongbo Tian, Wenhai Wang, Qipeng Guo, Kai Chen, Biqing Qi, and Bowen Zhou. SDAR: A synergistic diffusion- autoregression paradigm for scalable sequence generation.CoRR, abs/2510.06303,

    arXiv

  5. [8]

    Training verifiers to solve math word problems.CoRR, abs/2110.14168,

    Karl Cobbe, Vineet Kosaraju, Mohammad Bavarian, Mark Chen, Heewoo Jun, Lukasz Kaiser, Matthias Plappert, Jerry Tworek, Jacob Hilton, Reiichiro Nakano, Christopher Hesse, and John Schulman. Training verifiers to solve math word problems.CoRR, abs/2110.14168,

    arXiv

  6. [9]

    From bits to rounds: Parallel decoding with exploration for diffusion language models.CoRR, abs/2511.21103,

    Hengyu Fu, Baihe Huang, Virginia Adams, Charles Wang, Venkat Srinivasan, and Jiantao Jiao. From bits to rounds: Parallel decoding with exploration for diffusion language models.CoRR, abs/2511.21103,

    arXiv

  7. [10]

    Wide-in, narrow-out: Revokable decoding for efficient and effective dllms.CoRR, abs/2507.18578,

    10 Preprint Feng Hong, Geng Yu, Yushi Ye, Haicheng Huang, Huangjie Zheng, Ya Zhang, Yanfeng Wang, and Jiangchao Yao. Wide-in, narrow-out: Revokable decoding for efficient and effective dllms.CoRR, abs/2507.18578,

    arXiv

  8. [11]

    KLASS: kl-guided fast inference in masked diffusion models.CoRR, abs/2511.05664,

    Seo Hyun Kim, Sunwoo Hong, Hojung Jung, Youngrok Park, and Se-Young Yun. KLASS: kl-guided fast inference in masked diffusion models.CoRR, abs/2511.05664,

    arXiv

Show all 25 references
  1. [12]

    Why diffusion language models struggle with truly parallel (non-autoregressive) decoding?CoRR, abs/2602.23225,

    Pengxiang Li, Dilxat Muhtar, Tianlong Chen, Lu Yin, and Shiwei Liu. Why diffusion language models struggle with truly parallel (non-autoregressive) decoding?CoRR, abs/2602.23225,

  2. [15]

    Large language diffusion models.CoRR, abs/2502.09992,

    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.CoRR, abs/2502.09992,

    arXiv

  3. [16]

    Accelerating dif- fusion large language models with slowfast sampling: The three golden principles.CoRR, abs/2506.10848,

    Qingyan Wei, Yaojie Zhang, Zhiyuan Liu, Dongrui Liu, and Linfeng Zhang. Accelerating dif- fusion large language models with slowfast sampling: The three golden principles.CoRR, abs/2506.10848,

  4. [17]

    Fast-dllm: Training-free acceleration of diffusion LLM by enabling KV cache and parallel decoding.CoRR, abs/2505.22618,

    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.CoRR, abs/2505.22618,

    arXiv

  5. [18]

    Free draft-and-verification: Toward lossless parallel decoding for diffusion large language models.CoRR, abs/2510.00294,

    Shutong Wu and Jiawei Zhang. Free draft-and-verification: Toward lossless parallel decoding for diffusion large language models.CoRR, abs/2510.00294,

  6. [19]

    Lopa: Scaling dllm inference via lookahead parallel decoding.CoRR, abs/2512.16229,

    Chenkai Xu, Yijie Jin, Jiajun Li, Yi Tu, Guoping Long, Dandan Tu, Mingcong Song, Hongjie Si, Tianqi Hou, Junchi Yan, and Zhijie Deng. Lopa: Scaling dllm inference via lookahead parallel decoding.CoRR, abs/2512.16229,

  7. [20]

    Qwen3 technical report.CoRR, abs/2505.09388,

    An Yang, Anfeng Li, Baosong Yang, Beichen Zhang, Binyuan Hui, Bo Zheng, Bowen Yu, Chang Gao, Chengen Huang, Chenxu Lv, Chujie Zheng, Dayiheng Liu, Fan Zhou, Fei Huang, Feng Hu, Hao Ge, Haoran Wei, Huan Lin, Jialong Tang, Jian Yang, Jianhong Tu, Jianwei Zhang, Jianxin Yang, Jia...

    arXiv

  8. [21]

    Dream 7b: Diffusion large language models.CoRR, abs/2508.15487,

    Jiacheng Ye, Zhihui Xie, Lin Zheng, Jiahui Gao, Zirui Wu, Xin Jiang, Zhenguo Li, and Lingpeng Kong. Dream 7b: Diffusion large language models.CoRR, abs/2508.15487,

    arXiv

  9. [22]

    Rejection mixing: Fast semantic propagation of mask tokens for efficient DLLM inference

    Yushi Ye, Feng Hong, Huangjie Zheng, Xu Chen, Zhiyong Chen, Yanfeng Wang, and Jiangchao Yao. Rejection mixing: Fast semantic propagation of mask tokens for efficient DLLM inference. CoRR, abs/2602.22868,

  10. [23]

    Llada 1.5: Variance-reduced preference optimization for large language diffusion models.CoRR, abs/2505.19223,

    Fengqi Zhu, Rongzhen Wang, Shen Nie, Xiaolu Zhang, Chunwei Wu, Jun Hu, Jun Zhou, Jianfei Chen, Yankai Lin, Ji-Rong Wen, and Chongxuan Li. Llada 1.5: Variance-reduced preference optimization for large language diffusion models.CoRR, abs/2505.19223,

    arXiv

  11. [24]

    Accelerating masked image generation by learning latent controlled dynamics.CoRR, abs/2602.23996,

    11 Preprint Kaiwen Zhu, Quansheng Zeng, Yuandong Pu, Shuo Cao, Xiaohui Li, Yi Xin, Qi Qin, Jiayang Li, Yu Qiao, Jinjin Gu, and Yihao Liu. Accelerating masked image generation by learning latent controlled dynamics.CoRR, abs/2602.23996,

  12. [25]

    KLASS (Kim et al., 2025), EB- Sampler (Ben-Hamu et al., 2025), and WINO (Hong et al.,

    with thresholdτ= 0.9. KLASS (Kim et al., 2025), EB- Sampler (Ben-Hamu et al., 2025), and WINO (Hong et al.,

    2025

  13. [2020]

    Mark Chen, Jerry Tworek, Heewoo Jun, Qiming Yuan, Henrique Pond ´e 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...

    arXiv

  14. [2021]

    Learning to parallel: Accelerating diffu- sion large language models via adaptive parallel decoding.CoRR, abs/2509.25188,

    Wenrui Bao, Zhiben Chen, Dan Xu, and Yuzhang Shang. Learning to parallel: Accelerating diffu- sion large language models via adaptive parallel decoding.CoRR, abs/2509.25188,

  15. [2024]

    DAWN: dependency-aware fast inference for diffusion llms.CoRR, abs/2602.06953,

    Lizhuo Luo, Zhuoran Shi, Jiajun Luo, Zhi Wang, Shen Ren, Wenya Wang, and Tianwei Zhang. DAWN: dependency-aware fast inference for diffusion llms.CoRR, abs/2602.06953,

  16. [2025]

    Accelerated sampling from masked diffusion models via entropy bounded unmasking.CoRR, abs/2505.24857,

    Heli Ben-Hamu, Itai Gat, Daniel Severo, Niklas Nolte, and Brian Karrer. Accelerated sampling from masked diffusion models via entropy bounded unmasking.CoRR, abs/2505.24857,

    arXiv

  17. [2026]

    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, ICLR 2024, Vienna, Austria, May 7-11,

    2024

Pith tools

Reviewed August 16, 2026 · model on record in the stance chip above.