Pith. sign in

REVIEW 4 major objections 6 minor 29 references

A continuous-space diffusion language model in which each token follows its own noise-to-data time schedule can generate competitive text in very few steps while avoiding the errors of sampling tokens independently.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

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

T0 review · deepseek-v4-flash

2026-08-02 14:41 UTC pith:EKLZYBRK

load-bearing objection Per-token times via Beta quantiles are a genuine, well-proved mechanism with a strong Sudoku ablation, but the language-model claims overreach because the entropy-ranking rule is never ablated at scale on OpenWebText. the 4 major comments →

arxiv 2607.14106 v1 pith:EKLZYBRK submitted 2026-05-07 cs.CL cs.AI

Token Time Continuous Diffusion for Language Modeling

Parikshit Bansal , Sujay Sanghavi This is my paper
classification cs.CL cs.AI
keywords diffusion language modelscontinuous-space diffusionper-token timesflow matchingfew-step generationentropy-based rank assignmentshortcut distillationprefix-conditioned generation
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

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

This paper introduces token time continuous diffusion (TTCD), a diffusion language model that starts from Gaussian noise and deterministically evolves token embeddings to a final text canvas. The central idea is per-token times: each position has its own local time, so tokens the model is more sure about move from noise to data earlier and can guide the rest. This is designed to fix two known problems: discrete diffusion models that denoise several tokens at once sample from product marginals instead of the true joint distribution, and global-time continuous models treat every token as equally noisy. The paper reports that TTCD matches or beats comparable discrete and continuous diffusion language models at high speedups, and that it is especially strong on prefix-conditioned generation because prompt tokens can be pinned to clean time 1 while canvas tokens progress. The same mechanism also gives large gains on Sudoku solving at very low step counts.

Core claim

The central claim is that replacing a single global noise time with per-token times—drawn once per sequence and ordered by per-token predictive entropy—lets a continuous-space flow model generate language in one to eight steps with quality comparable to or better than discrete diffusion baselines, and with substantially better conditional generation. Each token's local time is computed from a Beta quantile function F(t_g, u), chosen so that the average local time equals the global time and higher-rank tokens are denoised earlier. At inference, one forward pass over the fully noised canvas ranks tokens by the entropy of their output distributions, and those ranks are fixed for the whole denoi

What carries the argument

The key machinery is the per-token time function F(t_g, u), the quantile function of a Beta(1/(1 - t_g), 1/t_g) distribution, together with an entropy-based rank assignment that orders tokens by how sure the model is at full noise. F is designed so every local time starts at 0 and ends at 1, increases with global time and with rank, and has mean equal to the global time. Each token then takes its own step size Delta_l = F(t_g + Delta, u_l) - F(t_g, u_l), and its update is the difference between the model's expected token embedding and the current embedding, scaled by 1/(1 - t_l). The architectural change is per-token adaptive layer norm: each token's local time modulates the transformer bloc

Load-bearing premise

The load-bearing premise is that the entropy of the model's output distribution over a fully noised canvas, measured in one forward pass, reliably tells which tokens will be easiest to denoise across the whole trajectory, so the ranks can be fixed once and never updated.

What would settle it

Train the same TTCD model on language data with random rank assignment instead of entropy-based ranks and compare generative perplexity at 2-4 steps; if random ranks match entropy ranks, the claimed benefit of entropy-based per-token scheduling is not real. Alternatively, re-rank tokens at every denoising step; if this improves quality, the fixed-rank assumption is wrong.

Watch this falsifier. Get emailed when new claim-graph text bears on it.

If this is right

  • Few-step diffusion language models should assign per-token times rather than a single global schedule, because continuous per-token evolution avoids the factorization error of sampling multiple tokens independently.
  • Prefix-conditioned generation becomes natural: prompt tokens stay at clean time 1 while only the output canvas is denoised, which the paper shows gives large quality gains at 2-4 steps.
  • The model can be distilled into one- or two-step generators with a shortcut self-consistency loss, making fast deterministic generation practical.
  • Entropy-based rank assignment from a single forward pass is a workable way to choose which tokens to denoise first, with no re-ranking during generation.
  • The mechanism generalizes beyond text: on Sudoku solving it lifts two-step accuracy from near zero or single digits to about 31%, versus about 12% for the best discrete baseline at the same step count.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • The paper itself notes that all language experiments are at a relatively small model scale and have not been validated at billions of parameters; if the comparative advantage shrinks at larger scales, the practical case for per-token times weakens.
  • A random-rank token-time baseline on language would isolate how much of the gain comes from per-token times rather than from the entropy-ranking rule; the paper only runs that ablation on Sudoku.
  • Because the schedule is a Beta quantile function with monotone likelihood ratio, the same construction could be applied to other flow-based discrete generators, including uniform-noise discrete diffusion models, which the paper names as future work.

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

4 major / 6 minor

Summary. TTCD proposes token-time continuous diffusion, a flow-matching language model in which each token has its own local time t_l = F(t_g, u_l), driven by a rank u_l. The paper chooses F as the quantile function of Beta(1/(1-t_g), 1/t_g) and states Lemma 1, which is proved in Appendix A.1. Training (Algo. 1) interpolates each token embedding between Gaussian noise and its clean embedding using the local time, and optimizes a token-level cross-entropy loss conditioned on the full vector of token times. Inference (Algo. 2) is deterministic: ranks are assigned by the entropy of the model's output distribution on the fully noised canvas, are fixed thereafter, and the canvas is evolved with the Euler-style update of Eq. (5); prompt tokens are pinned at local time 1. A shortcut/consistency distillation is described in Algo. 3, and an adaLN modification conditions each token on its own local time without adding parameters. Experiments include 9x9 Sudoku, a 160M-parameter OpenWebText model, and QM9.

Significance. If the claims hold, per-token continuous schedules are a simple and useful design for few-step diffusion language models, particularly for conditional generation, and they offer a concrete way around the 'factorization problem' of discrete diffusion at high speedups. The paper has genuine strengths: Lemma 1 is a real proof, not a restatement; the Sudoku ablation includes a random-rank arm, which directly tests the entropy-ranking rule; the architectural change is parameter-free; and the pseudocode is complete enough to reimplement. The principal weaknesses are empirical control and reporting: on the main language experiments, the per-token-time mechanism is not separated from entropy-based rank ordering or from prompt-pinning, and no error bars are given for the headline comparisons. The model-scale inconsistency (160M vs 100M) also needs correction.

major comments (4)
  1. [§3.3 / Algo. 2 vs §3.2 / Algo. 1; Tables 2–5, Fig. 3] All OpenWebText TTCD generations use entropy-ordered ranks, while training (Algo. 1) draws ranks iid uniform. No random-rank TTCD arm is reported on language, so the language results cannot separate the per-token-time mechanism from the entropy-ranking rule. The Sudoku ablation (Table 1) is a clean controlled test and strongly supports the entropy rule there (31.51% vs 1.01% at 2 steps; 61.33 vs 22.27 at 4 steps), but it is a 6M-parameter toy task. Please add a random-rank/uniform-rank TTCD ablation on OpenWebText at least at 2 and 4 steps, and report how entropy ranks correlate with final token confidence.
  2. [§4.2, Fig. 4, Tables 4–5; Algo. 2 line 5] The prefix-conditioned comparison is confounded by prompt pinning: TTCD pins all prompt tokens at t=1, whereas FLM and Duo use a single global time for prompt and canvas. The claim that per-token times 'naturally allow prefix-conditioned generation' is therefore not directly supported; the observed conditional gains could come from pinning alone. Add an ablation with TTCD where the prompt is denoised with the canvas, or a global-time continuous baseline with the prompt pinned clean, to separate these factors.
  3. [Checklist item 7; Tables 2–5] The headline 'comparable/outperform' claims rest on point estimates without error bars or per-seed values; the checklist states that averages over three seeds are used but no variance is reported. At several operating points the differences that matter for the comparison are not large relative to typical seed noise (e.g., Table 3, 4 steps: TTCD PPL 249.20 at temperature 1.0 vs Duo-DCD 305.75; the ordering reverses at temperature 0.8). Please report standard deviations or confidence intervals for the central OWT and Sudoku results, or state which differences are robust across seeds.
  4. [Abstract, §1, §4.2, §5] The model scale is stated inconsistently: the abstract and §4.2 say a 160M parameter model, while §1 and §5 say 100M. Since the comparisons claim 'similar size' to baselines, this inconsistency affects the interpretation of the empirical claims. Please state the exact architecture and parameter count, and update the abstract/limitations accordingly.
minor comments (6)
  1. [Related Work] The sentence 'Prior work like CCDD CADD, CANDI' appears garbled; the model names should be enumerated and cited cleanly.
  2. [Table 3] The label 'TTCD w/ Shortucut' is a typo for 'Shortcut'.
  3. [§4.1] The text says TTCD is 'just behind' discrete entropy-based unmasking at higher steps, but Table 1 shows 65.85 vs 92.90 at 8 steps and 68.46 vs 97.30 at 16 steps. Please rephrase to avoid overstating closeness.
  4. [Algo. 2 / §3.3] The initialization 't_g, Δ ← 0, 1/N' is ambiguous about whether Δ is the step size or an increment variable; please clarify the global schedule notation.
  5. [Checklist item 5 / Appendix B] No code is released and the Sudoku/QM9 training details are compressed to 'follow prior work.' Since the paper introduces new inference logic, releasing code or providing exact hyperparameters for the 6M Sudoku model would improve reproducibility.
  6. [§3.4 / Algo. 3] The shortcut objective is described as minimizing KL between one-step and two-step unrolled probabilities, but the paper does not state whether the teacher network is frozen or whether stop-gradient is used for both target branches; Algo. 3 says stopgrad on target logits, but the text should be equally explicit.

Circularity Check

0 steps flagged

No significant circularity: TTCD's theoretical lemma and empirical comparisons are self-contained.

full rationale

Walking the derivation chain, I find no step that reduces a stated prediction to its own inputs. Lemma 1 is a genuine proof: F is defined as the Beta quantile and properties A-D are verified mathematically; requirement D holds because the Beta mean equals tg under the chosen parameterization, but that is a design check, not a restatement of the target result. The entropy-based rank assignment in Algo 2 (lines 6-7) is self-referential in flavor: 'surety' is defined as low output entropy and ranks are sorted accordingly, so the statement that sure tokens are denoised earlier is true by construction. However, the paper's actual performance claims are not derived from that definition. The usefulness of the ranking rule is empirically anchored by the Sudoku ablation (Table 1: token-time with entropy 31.51% vs. token-time random 1.01% at 2 steps), which is a controlled comparison against random ranks. On OpenWebText the absence of a random-rank token-time arm means entropy-ranking and prompt-pinning effects are not fully isolated from per-token times, but this is an experimental confound/ablation gap, not circularity: the OWT results compare TTCD against independently trained MDLM, Duo, and FLM baselines. Shortcut distillation is standard self-distillation with a ground-truth infinitesimal-step cross-entropy term (Algo 3 lines 13-14); teacher and student share parameters, but the objective is not a fitted constant and the baselines are external. There are no author self-citations used as load-bearing evidence and no imported uniqueness theorem. Therefore the derivation is self-contained.

Axiom & Free-Parameter Ledger

2 free parameters · 4 axioms · 1 invented entities

The method's genuine contribution is the per-token time schedule plus entropy ranking; the ledger shows the price: two hand-chosen design elements (Beta family, entropy rule) that are not fitted but are also not ablated at scale, plus standard continuous-LM flow assumptions inherited from prior work.

free parameters (2)
  • Per-token time function F (Beta quantile family)
    F(t_g,u) is chosen by hand as the quantile of Beta(1/(1-t_g),1/t_g) (§3.1, Lemma 1). Only properties A-D are required, and many F would satisfy them; the specific Beta form is never ablated against alternatives, so the method's success is pinned to this un-fitted design choice.
  • Entropy-based rank ordering rule
    Algo 2 lines 6-7 sort ranks by model entropy at t_g=0. This rule is chosen by hand; no alternatives (confidence margin, learned scheduler, iterative re-ranking) are tested on OpenWebText, and the random-rank baseline exists only in the Sudoku experiments.
axioms (4)
  • domain assumption Lower entropy at t_g=0 identifies tokens that are genuinely easier, and denoising them earlier improves sample quality.
    Load-bearing for Algo 2 and all main LM results; supported only by the 6M-parameter Sudoku ablation (Table 1), not by any OpenWebText ablation of rank assignment.
  • domain assumption The cross-entropy flow of Eqns 3/5 gives a faithful conditional-mean estimate of the token embeddings, so the per-token step-size update is an accurate transport step.
    Inherited from prior continuous LMs (CDCD, FLM, cited as [5,10]); TTCD's per-token step sizes Delta_l amplify any bias in this estimate, making the assumption more consequential than in global-time models.
  • ad hoc to paper The specific Beta family (not just any F satisfying A-D) is the right per-token time design.
    The motivation in §1/§3.1 (sudden transition zone, sure tokens) is qualitative; there is no derivation selecting the Beta quantile, and no ablation over F.
  • domain assumption Deterministic ODE coupling through the model encodes joint token dependencies, so final argmax decoding avoids the product-marginal error of discrete parallel sampling.
    Central motivation of §1; stated as a hypothesis ('We call this the factorization problem') with no formal analysis linking ODE determinism to joint-sample quality.
invented entities (1)
  • Per-token rank u_l (token time) independent evidence
    purpose: Assigns each canvas position its own noise-to-data schedule; ranks are sorted by model entropy so sure tokens denoise first, and prompt tokens are fixed at rank time 1.
    The construct is operationally fully specified (Eqn 1-5, Algo 2) and has falsifiable handles: within the paper, the Sudoku ablation shows random-rank token time (1.01%) vs entropy-rank (31.51%) at 2 steps; outside, it predicts that entropy-ranked per-token schedules should transfer to other diffusion LMs and datasets.

pith-pipeline@v1.3.0-alltime-deepseek · 4580 in / 4761 out tokens · 252469 ms · 2026-08-02T14:41:53.842588+00:00 · methodology

0 comments
read the original abstract

In this paper we introduce token time continuous diffusion (TTCD), a new diffusion language model which (a) operates in continuous space, deterministically mapping Gaussian noise to a final token canvas with no further sampling, and crucially (b) incorporates a new notion of per-token times, with some tokens proceeding from noise to token at a faster rate than others. Continuous space modeling helps TTCD avoid the parallel sampling of multiple tokens, which is a key source of inaccuracy at high speedups for models that iterate purely in discrete space. The notion of per-token times helps TTCD to better model conditional generation, allows for more sure tokens to proceed at a faster rate, and allows for differentiated inter-token influences during refinement. TTCD outperforms discrete models at high speedups. We train a 160M parameter TTCD model on OpenWebText, and then self-distill it; we find that at high speedups we are comparable in unconditional generation quality, and outperform in conditional generation, several existing models of similar size trained, on the same data, and self-distilled. We achieve similar gains in Sudoku solving as well.

Figures

Figures reproduced from arXiv: 2607.14106 by Parikshit Bansal, Sujay Sanghavi.

Figure 1
Figure 1. Figure 1: In this figure we evaluate unconditional generation performance of distilled diffusion LMs trained on OpenWebText (see Sec 4.2). We generate samples from the different methods considered, and plot their generative perplexity (y-axis, lower is better) and text entropy (x-axis, higher is better), for one to eight step generation. Duo w/ DCD is a distilled uniform-noise (discrete) diffusion model and Flow map… view at source ↗
Figure 2
Figure 2. Figure 2: Per-token times:(a) Left: This shows how local times t l change as global time tg varies from 0 to 1, for different ranks u l (given by colors).(b) Right: For a final generated sequence “token times", this represents the distribution of intermediate vectors z as a function of global time tg for given token ranks u l . At tg = 0 they are all the same Gaussian, at tg = 1 they are on the respective data token… view at source ↗
Figure 3
Figure 3. Figure 3: Unconditional generation: In this figure we study the unconditional generation of the different diffusion-based language models trained on the OpenWebText (OWT) dataset (see Sec 4.2). We compare the perplexity of the generated samples with varying step (forward passes through the diffusion model) budget. The plot shows that the baseline continuous diffusion model’s (FLM) entropy on varying it’s inference p… view at source ↗
Figure 4
Figure 4. Figure 4: Prefix-conditioned generation on canvas of 32 length. (a) Left: We evaluate different models on conditional generation of a canvas of 32 tokens (see Sec 4.2). The plots demonstrate the efficacy of TTCD (w/ shortuct) over the baselines diffusion models for two (16x speedup) and four (8x speedup) generation steps. (b) Right: The figure shows a qualitative example of the prefix-conditioned generation for Duo … view at source ↗
Figure 5
Figure 5. Figure 5: Classifier-free guidance on QM9 We evaluate classifier-free guidance on the molecule data QM9 [20]. The methods are trained on the QM9 dataset following prior work set of training hyper parameters [20]. The base￾line methods are (a) UDLM is a discrete uniform noising model and (b) MDLM is masked discrete diffusion model. We consider two continuous diffusion baselines (c) Continuous w/ sequence time and (d)… view at source ↗
Figure 6
Figure 6. Figure 6: Warping function used for global-only time continuous sudoku model’s training and [PITH_FULL_IMAGE:figures/full_fig_p014_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: In this figure we evaluate the continuous method on varying the two sampling parameters [PITH_FULL_IMAGE:figures/full_fig_p015_7.png] view at source ↗
Figure 8
Figure 8. Figure 8: In this figure we present numbers for different canvas lengths, concretely, 32 tokens (top [PITH_FULL_IMAGE:figures/full_fig_p017_8.png] view at source ↗
Figure 9
Figure 9. Figure 9: QED Mean figure 20 [PITH_FULL_IMAGE:figures/full_fig_p020_9.png] view at source ↗
Figure 10
Figure 10. Figure 10: Generated sample from TTCD for 8 step generation [PITH_FULL_IMAGE:figures/full_fig_p021_10.png] view at source ↗
Figure 11
Figure 11. Figure 11: Generated sample from TTCD for 32 step generation [PITH_FULL_IMAGE:figures/full_fig_p022_11.png] view at source ↗
Figure 12
Figure 12. Figure 12: Generated sample from TTCD (w/ Shortcut) for 2 step generation [PITH_FULL_IMAGE:figures/full_fig_p023_12.png] view at source ↗
Figure 13
Figure 13. Figure 13: Generated sample from TTCD (w/ Shortcut) for 8 step generation [PITH_FULL_IMAGE:figures/full_fig_p024_13.png] view at source ↗
Figure 14
Figure 14. Figure 14: Prefix-conditioned generated sample from TTCD (w/ Shortcut) and Duo w/ DCD for 4 [PITH_FULL_IMAGE:figures/full_fig_p025_14.png] view at source ↗
Figure 15
Figure 15. Figure 15: Prefix-conditioned generated sample from TTCD (w/ Shortcut) and Duo w/ DCD for 4 [PITH_FULL_IMAGE:figures/full_fig_p026_15.png] view at source ↗

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

29 extracted references · 12 linked inside Pith

  1. [1]

    Structured denoising diffusion models in discrete state-spaces.Advances in neural information processing systems, 34:17981–17993, 2021

    Jacob Austin, Daniel D Johnson, Jonathan Ho, Daniel Tarlow, and Rianne Van Den Berg. Structured denoising diffusion models in discrete state-spaces.Advances in neural information processing systems, 34:17981–17993, 2021

    2021

  2. [2]

    Dirichlet diffusion score model for biological sequence generation

    Pavel Avdeyev, Chenlai Shi, Yuhao Tan, Kseniia Dudnyk, and Jian Zhou. Dirichlet diffusion score model for biological sequence generation. InInternational Conference on Machine Learning, pages 1276–1301. PMLR, 2023

    2023

  3. [3]

    How to build a consistency model: Learning flow maps via self-distillation.arXiv preprint arXiv:2505.18825, 2025

    Nicholas M Boffi, Michael S Albergo, and Eric Vanden-Eijnden. How to build a consistency model: Learning flow maps via self-distillation.arXiv preprint arXiv:2505.18825, 2025

    arXiv 2025

  4. [4]

    Fisher flow matching for generative modeling over discrete data.Advances in Neural Information Processing Systems, 37:139054–139084, 2024

    Oscar Davis, Samuel Kessler, Mircea Petrache, ˙Ismail ˙I Ceylan, Michael Bronstein, and Avishek J Bose. Fisher flow matching for generative modeling over discrete data.Advances in Neural Information Processing Systems, 37:139054–139084, 2024

    2024

  5. [5]

    Continu- ous diffusion for categorical data.arXiv preprint arXiv:2211.15089, 2022

    Sander Dieleman, Laurent Sartran, Arman Roshannai, Nikolay Savinov, Yaroslav Ganin, Pierre H Richemond, Arnaud Doucet, Robin Strudel, Chris Dyer, Conor Durkan, et al. Continu- ous diffusion for categorical data.arXiv preprint arXiv:2211.15089, 2022

    Pith/arXiv arXiv 2022

  6. [6]

    One step diffusion via shortcut models.arXiv preprint arXiv:2410.12557, 2024

    Kevin Frans, Danijar Hafner, Sergey Levine, and Pieter Abbeel. One step diffusion via shortcut models.arXiv preprint arXiv:2410.12557, 2024

    Pith/arXiv arXiv 2024

  7. [7]

    Likelihood-based diffusion language models

    Ishaan Gulrajani and Tatsunori B Hashimoto. Likelihood-based diffusion language models. Advances in Neural Information Processing Systems, 36:16693–16715, 2023

    2023

  8. [8]

    Continuous diffusion model for language modeling

    Jaehyeong Jo and Sung Ju Hwang. Continuous diffusion model for language modeling. InThe Thirty-ninth Annual Conference on Neural Information Processing Systems

  9. [9]

    Train for the worst, plan for the best: Understanding token ordering in masked diffusions.arXiv preprint arXiv:2502.06768, 2025

    Jaeyeon Kim, Kulin Shah, Vasilis Kontonis, Sham Kakade, and Sitan Chen. Train for the worst, plan for the best: Understanding token ordering in masked diffusions.arXiv preprint arXiv:2502.06768, 2025

    Pith/arXiv arXiv 2025

  10. [10]

    One-step language modeling via continuous denoising.arXiv preprint arXiv:2602.16813, 2026

    Chanhyuk Lee, Jaehoon Yoo, Manan Agarwal, Sheel Shah, Jerry Huang, Aditi Raghunathan, Seunghoon Hong, Nicholas M Boffi, and Jinwoo Kim. One-step language modeling via continuous denoising.arXiv preprint arXiv:2602.16813, 2026

    Pith/arXiv arXiv 2026

  11. [11]

    Diffusion-lm improves controllable text generation.Advances in neural information processing systems, 35:4328–4343, 2022

    Xiang Li, John Thickstun, Ishaan Gulrajani, Percy S Liang, and Tatsunori B Hashimoto. Diffusion-lm improves controllable text generation.Advances in neural information processing systems, 35:4328–4343, 2022

    2022

  12. [12]

    Stop-think-autoregress: Language modeling with latent diffusion planning.arXiv preprint arXiv:2602.20528, 2026

    Justin Lovelace, Christian Belardi, Sofian Zalouk, Adhitya Polavaram, Srivatsa Kundurthy, and Kilian Q Weinberger. Stop-think-autoregress: Language modeling with latent diffusion planning.arXiv preprint arXiv:2602.20528, 2026

    arXiv 2026

  13. [13]

    Latent diffusion for language generation.Conference on Neural Information Processing Systems (NeurIPS), 2023

    Justin Lovelace, Varsha Kishore, Chao Wan, Eliot Shekhtman, and Kilian Q Weinberger. Latent diffusion for language generation.Conference on Neural Information Processing Systems (NeurIPS), 2023

    2023

  14. [14]

    Large language diffusion models.arXiv preprint arXiv:2502.09992, 2025

    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

    Pith/arXiv arXiv 2025

  15. [15]

    Scalable diffusion models with transformers

    William Peebles and Saining Xie. Scalable diffusion models with transformers. InProceedings of the IEEE/CVF international conference on computer vision, pages 4195–4205, 2023

    2023

  16. [16]

    Candi: Hybrid discrete-continuous diffusion models.arXiv preprint arXiv:2510.22510, 2025

    Patrick Pynadath, Jiaxin Shi, and Ruqi Zhang. Candi: Hybrid discrete-continuous diffusion models.arXiv preprint arXiv:2510.22510, 2025

    Pith/arXiv arXiv 2025

  17. [17]

    Simple and effective masked diffusion language models.Advances in Neural Information Processing Systems, 37:130136–130184, 2024

    Subham Sahoo, Marianne Arriola, Yair Schiff, Aaron Gokaslan, Edgar Marroquin, Justin Chiu, Alexander Rush, and V olodymyr Kuleshov. Simple and effective masked diffusion language models.Advances in Neural Information Processing Systems, 37:130136–130184, 2024. 10

    2024

  18. [18]

    The diffusion duality.arXiv preprint arXiv:2506.10892, 2025

    Subham Sekhar Sahoo, Justin Deschenaux, Aaron Gokaslan, Guanghan Wang, Justin Chiu, and V olodymyr Kuleshov. The diffusion duality.arXiv preprint arXiv:2506.10892, 2025

    arXiv 2025

  19. [19]

    Progressive distillation for fast sampling of diffusion models

    Tim Salimans and Jonathan Ho. Progressive distillation for fast sampling of diffusion models. arXiv preprint arXiv:2202.00512, 2022

    Pith/arXiv arXiv 2022

  20. [20]

    Simple guidance mechanisms for discrete diffusion models

    Yair Schiff, Subham Sekhar Sahoo, Hao Phung, Guanghan Wang, Alexander Rush, V olodymyr Kuleshov, Hugo Dalla-Torre, Sam Boshar, Bernardo P de Almeida, and Thomas Pierrot. Simple guidance mechanisms for discrete diffusion models. In... International Conference on Learning Representations, volume 2025, page 44153, 2025

    2025

  21. [21]

    Causal language modeling can elicit search and reasoning capabilities on logic puzzles.Advances in Neural Information Processing Systems, 37:56674–56702, 2024

    Kulin Shah, Nishanth Dikkala, Xin Wang, and Rina Panigrahy. Causal language modeling can elicit search and reasoning capabilities on logic puzzles.Advances in Neural Information Processing Systems, 37:56674–56702, 2024

    2024

  22. [22]

    Dirichlet flow matching with applications to dna sequence design

    Hannes Stark, Bowen Jing, Chenyu Wang, Gabriele Corso, Bonnie Berger, Regina Barzilay, and Tommi Jaakkola. Dirichlet flow matching with applications to dna sequence design. In International Conference on Machine Learning, pages 46495–46513. PMLR, 2024

    2024

  23. [23]

    Fast-dllm: Training-free acceleration of diffusion llm by enabling kv cache and parallel decoding.arXiv preprint arXiv:2505.22618, 2025

    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

    Pith/arXiv arXiv 2025

  24. [24]

    Dream 7b: Diffusion large language models.arXiv preprint arXiv:2508.15487, 2025

    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

    Pith/arXiv arXiv 2025

  25. [25]

    Dinoiser: Diffused conditional sequence learning by manipulating noises.arXiv preprint arXiv:2302.10025, 2023

    Jiasheng Ye, Zaixiang Zheng, Yu Bao, Lihua Qian, and Mingxuan Wang. Dinoiser: Diffused conditional sequence learning by manipulating noises.arXiv preprint arXiv:2302.10025, 2023

    Pith/arXiv arXiv 2023

  26. [26]

    Continuously augmented discrete diffusion model for categorical generative modeling.arXiv preprint arXiv:2510.01329, 2025

    Huangjie Zheng, Shansan Gong, Ruixiang Zhang, Tianrong Chen, Jiatao Gu, Mingyuan Zhou, Navdeep Jaitly, and Yizhe Zhang. Continuously augmented discrete diffusion model for categorical generative modeling.arXiv preprint arXiv:2510.01329, 2025

    arXiv 2025

  27. [27]

    Masked diffusion models are secretly time-agnostic masked models and exploit inaccurate categorical sampling.arXiv preprint arXiv:2409.02908, 2024

    Kaiwen Zheng, Yongxin Chen, Hanzi Mao, Ming-Yu Liu, Jun Zhu, and Qinsheng Zhang. Masked diffusion models are secretly time-agnostic masked models and exploit inaccurate categorical sampling.arXiv preprint arXiv:2409.02908, 2024

    Pith/arXiv arXiv 2024

  28. [28]

    It’s cool, get 20 degrees

    Cai Zhou, Chenxiao Yang, Yi Hu, Chenyu Wang, Chubin Zhang, Muhan Zhang, Lester Mackey, Tommi Jaakkola, Stephen Bates, and Dinghuai Zhang. Coevolutionary continuous discrete diffu- sion: Make your diffusion language model a latent reasoner.arXiv preprint arXiv:2510.03206, 2025. A Theoretical Results A.1 Proof of Lemma 1 Proof.Fort g ∈(0,1), write α(tg) := ...

    Pith/arXiv arXiv 2025

  29. [29]

    • Depending on the country in which research is conducted, IRB approval (or equivalent) may be required for any human subjects research

    Institutional review board (IRB) approvals or equivalent for research with human subjects Question: Does the paper describe potential risks incurred by study participants, whether such risks were disclosed to the subjects, and whether Institutional Review Board (IRB) approvals (or an equivalent approval/review based on the requirements of your country or ...