KLASS: KL-Guided Fast Inference in Masked Diffusion Models

We sincerely thank you for your comprehensive and positive comments on our manuscript. We are especially grateful for your recognition of our key contributions:

• The conceptual elegance and practical power of our simple, training-free method.
• The strength and breadth of our empirical study, demonstrating consistent quality improvements and significant speed-ups across diverse domains.
• Its reproducibility, ensured by our provision of code and detailed experimental settings.

We have carefully considered your suggestions for improvement and provide our detailed responses below.

[W4-1] Theoretical Guarantee

We thank the reviewer for their helpful comments on Theorem 4.3. We agree that its primary role was to build intuition. Motivated by your feedback, we have developed a new, rigorous information-theoretic framework that formally proves the effectiveness of KLASS. This now serves as the primary theoretical foundation in our revised manuscript.

The primary goal of the original Theorem 4.3 was to provide a high-level intuition for why KLASS is effective. Our intended definitions were as follows:

• "Mild context change": We intended this to mean a change between two contexts, $x_{-i}$ and $x'_{-i}$, where the Hamming distance is bounded by a small integer $k$. This is empirically grounded, as standard samplers like Top-1 or Top-2 correspond to k=1 or k=2. Our own adaptive method, KLASS, also results in a small average k (typically 2-4 tokens per step), confirming the changes are indeed mild.

• "δ-approximation norm": This refers to the distance between our model's distribution, $p_\theta$, and an ideal target distribution, $\pi$, as measured by KL Divergence. Formally, the assumption is that the model's approximation error is bounded such that $D_{KL}(\pi \| p_\theta) \le \delta$.

Formalizing these concepts as you suggested indeed makes the original theorem's structure more rigorous. Taking your valuable feedback a step further, we were motivated to develop an entirely new proof that provides a more direct and quantitatively verifiable guarantee, which we believe better solidifies the theoretical foundations of our work.

Provable Framework for Token Unmasking Safety

To complement this high-level intuition with a rigorous proof, we have developed a new, self-contained theoretical framework built on standard principles of information theory. This new information-theoretic approach directly analyzes the risk of unmasking a stable token. It provides a provable, quantitative upper bound on this risk, thereby offering a formal and rigorous complement to the high-level intuition provided by our original theorem.

To make our new framework clear, we first define the key terms used in our theorem:

• Risk Measure (CMI), $I_t(i)$: We quantify the risk of unmasking token $i$ at step $t$ using the Conditional Mutual Information, $I_t(i) = I(X_i; Z \mid V_t)$, which measures its dependency on all other masked tokens $Z$. A low CMI means the token is "safe" to unmask.
• Cone Entry Step, $t_\star(i)$: This is the first step where token $i$ becomes 'stable' by satisfying both KL stability ($K_{t,n}(i) \leq \varepsilon$) and confidence ($\max_v p_t(X_i=v) \geq \tau$) thresholds.
• Post-Stability KL Tail, $\delta_t(i)$: This term aggregates the small, confidence-weighted KL divergences ($\Delta$) of a token after it has entered the stability cone: $$\delta_t(i) = \sum_{k=t_\star(i)}^{t-1} [1-c_{k+1}(i)]\Delta_{k+1}(i)$$
• Initial Entropy, $H_{t_\star(i)}(X_i)$: This is the uncertainty (entropy) of token $i$ at the precise moment it becomes stable. Due to the high confidence requirement for entering the cone, this value is provably small.

Theorem (Provable Bound for Token Safety) Let $i$ be a token that is inside the stability cone at step $t$, and let $t_\star(i)$ be the step at which it first entered the cone. The mutual information $I_t(i)$ is bounded by:

We are grateful for your detailed and constructive feedback. We were pleased to see you recognize that our work:

• Rests on a rational principle supported by theoretical analysis.

• Validates its effectiveness with comprehensive experimental results.

We have carefully considered your comments and provide our detailed responses below.

[W3-1] Relationship between KL metric and confidence

We agree their relationship is critical. Below, we clarify its relationship.

1. Correlation between confidence and KL metric

To better understand the relationship between confidence and KL divergence, we measured the Pearson correlation coefficient between the two metric. We conducted this analysis by recording the confidence and KL divergence of the decoded tokens. As shown in the table below, the correlation between these two metrics is relatively low across all datasets. This indicates that the confidence and KL metrics capture different aspects of model behavior, thus making them complementary rather than overlapping.

2. Impact of using only one metric

As shown in Table 1 of our manuscript, when we rely on either confidence or KL divergence alone, the model's performance drops compared to when both metrics are utilized together. This supports the argument that while each metric has its own utility, their combined use yields a more robust decision-making process.

3. Semantic Distinction

Our empirical and theoretical analyses confirm that confidence and KL divergence are complementary, not redundant. Their Pearson correlation is consistently low across all datasets, and as we show in response [W2-1], correctly generated tokens exhibit both higher confidence and lower KL divergence. This motivates their combined use, as they capture fundamentally different aspects of the model's prediction process.

Confidence provides a static snapshot of certainty at a single timestep ($p_\theta(x_{t, \ell} | z_t)$). It measures the peak of a probability distribution, but it contains no information about the iterative process that led to that state. Relying on confidence alone can lead to errors from premature commitments to transient spikes of certainty.

Divergence, by contrast, measures the dynamics of the model's predictions over time by comparing consecutive distributions ($p_\theta(\cdot | z_t)$ and $p_\theta(\cdot | z_{t-1})$). A low KL score is a direct signal that the model's iterative refinement has concluded for that token—its thought process has converged and stabilized. It is a measure of temporal consistency.

KLASS requires both a confident state and a stable flow. This two-factor check allows us to fully leverage the diffusion model's iterative refinement capability, ensuring we unmask tokens only when the model's prediction is both certain and has verifiably ceased to fluctuate.

[W3-2] Peformance gap with MaskGIT setting

Thank you for the constructive feedback. While FID score improvement seems marginal compared to the baseline method, significance of KLASS is particularly evident in the Inception Score (IS) [1]. Specifically, KLASS achieved a 9.2-point increase in IS over the strong baseline (43% relative gain) which indicates substantial improvement in sample quality and diversity [2]. By waiting for a prediction to become stable (low KL-divergence) rather than exploiting only confidence value, KLASS enables generating complex classes where confidence-based samplers might often fail, which is indicated by the high IS score. Furthermore, we would like to point out that the reported value has further room for improvement as we conducted only naive hyper-parameter search for the experiment. We will include this discussion in our final manuscript.

[W3-3] Hyperparameter selection of threshold

We fully agree that the practicality of our method hinges on a systematic hyperparameter selection process. To this end, all our experiments were guided by a lightweight, three-step guideline we developed for this purpose. Below, we outline this practical guideline—which requires only a small validation set—and demonstrate its low sensitivity using Dream on MBPP as a concrete example. We kindly ask you to refer [W1-1] for additional results on hyperparameter sensitivity.

Step 1. KL Threshold Estimation: Run baseline decoding on the validation set and collect KL values. Set the initial KL threshold to the median of this distribution (e.g. $\approx$ 0.01 for reasoning tasks).

Step 2. Confidence Threshold Search: Fix the KL threshold from Step 1. Sweep confidence from 0.9 down to 0.6, selecting the one that maximizes accuracy while maintaining efficiency. For Dream on MBPP, fixing KL = 0.01 yields:

Both 0.9 and 0.95 tie in accuracy, but confidence = 0.9 uses fewer evaluations. We therefore select conf = 0.9.

Step 3. KL Threshold Refinement: Fix confidence at the value from Step 2 and perform a finer KL sweep around the initial estimate. Results with fixing conf = 0.9 are:

Peak accuracy at KL = 0.001 confirms both the initial estimate and its optimal refinement, giving the final configuration (conf=0.9, kl=0.001).

By optimizing confidence and KL in this order, we converge on strong configurations with far fewer evaluations than a full grid search. Moreover, since thresholded decoding reduces per-sample latency, the search process is fast and practical.

Using the selected configuration, we achieve 64.59% accuracy with 111.24 steps—outperforming the top-confidence baseline (63.81% accuracy with 256 steps). Notably, performance remains stable across nearby settings, demonstrating robustness and low sensitivity to hyperparameter choices. The table below shows results on the full dataset:

Dream on MBPP

Each cell: Accuracy (Steps)

Accuracy stays within 1% of the selected configuration across variations, with potential for even further tuning. This stability, combined with minimal search cost, highlights the practicality and generalizability of our guideline across tasks and models.

[Q3-1] Final total sampling steps

Thank you for raising insightful question. KLASS operates within a predefined maximum number of sampling steps but shortens the actual steps used by unmasking a variable number of tokens per step based on the KLASS criterion. If no tokens meet the criterion, the top-k most confident tokens are selected to satisfy a minimum number of unmaskings per step—determined to ensure full sequence completion within the step budget. This strategy allows KLASS to often finish earlier while never exceeding the predefined maximum.

Additionally, KLASS can be configured to match a fixed number of steps exactly by selecting the required number of tokens from those satisfying the KLASS criterion. To enforce a predictable lower bound on the total sampling steps, we may guide the process by the analytically computable expectation of token transitions from the predefined reverse process. At each step, we decode all tokens that meet the KLASS criterion and then fill any deficit with top-k decoding to meet the target number for that step.

We have added further results demonstrating that our method also performs well under a shorter predefined budget of 128 steps. While respecting this maximum, KLASS often completes generation in fewer steps and achieves higher accuracy than the top-k confidence baseline.

Each cell: Accuracy (Steps)

We will include above discussion and results in our final manuscript to strengthen the paper.

References

[1] Brock et al. "Large Scale GAN Training for High Fidelity Natural Image Synthesis." ICLR2019.

[2] Zhou et al. "Activation Maximization Generative Adversarial Nets." ICLR2018.

1. Effectiveness in MaskGIT

We sincerely thank the reviewer for the constructive and insightful feedback. We acknowledge the reviewer's valid point regarding the importance of the FID metric and that the improvement presented initially might appear marginal. To better address this concern, we would like to offer a deeper analysis that we believe clarifies the core contribution and effectiveness of our proposed KLASS sampler. This analysis is directly motivated by a key finding within the original MaskGIT paper itself.

In MaskGIT [1], the authors observed that for their confidence-based scheduling function, "more iterations are not necessarily better." They found that performance tends to peak at a sweet spot before worsening again. Crucially, they hypothesized that this is because "too many iterations may discourage the model from keeping less confident predictions, which worsens the token diversity." This points to a fundamental limitation of relying solely on confidence: it can lead to a loss of sample diversity (i.e., mode collapse) as the model prematurely converges on a narrow set of high-probability tokens, especially when more refinement steps are applied. By evaluating a prediction's stability rather than just its peak confidence, KLASS avoids the greedy, premature commitment that causes the baseline to fail. This preserves token diversity and enables stable, robust image generation even as the number of refinement steps increases.

Our new results below demonstrate this point clearly. As the number of steps increases from 64 to 128, the baseline confidence scheduler's performance collapses, with its FID score worsening from 46.84 to 53.23. In stark contrast, KLASS not only delivers a significantly better FID at both settings, but its performance remains remarkably stable and robust.

Therefore, the primary contribution of KLASS lies in its ability to solve this fundamental stability-diversity issue, preventing the loss of sample diversity and performance collapse when using longer, more refined sampling schedules. This specific capability is what explains the superior and robust performance of KLASS on MaskGIT, as demonstrated in our experiments.

To further substantiate that the effectiveness of KLASS is not confined to a narrow sweet spot where the original MaskGIT performs optimally, we will include expanded experimental results in the final manuscript, testing KLASS on additional generative models (e.g. MMaDA[2]).

2. Practicality of KLASS

Thank you for raising this important concern regarding the practicality of our method. To address this concern, we present results using a single, non-optimized configuration of KLASS (confidence = 0.9, KL = 0.01) across all models and datasets. As shown below, this default setting consistently outperforms the top-confidence baseline in both accuracy and efficiency. This demonstrates that users can adopt KLASS with a robust default setting for immediate performance gains without any task-specific tuning, making it just as general and easy-to-use as traditional confidence-guided strategy, yet with significantly better performance.

For users seeking to maximize performance, KLASS accommodates an optional, lightweight tuning process. As detailed in our response [W1-1], this procedure requires only a small validation set (~100 examples) and is computationally inexpensive. For LLaDA on MATH500, for example, the tuning process takes around 4.2 hours, which is even less than the time required for a single inference run using a top-1 confidence baseline (5.3 hours). Therefore, KLASS offers a practical and generalizable sampling strategy that achieves strong performance without task-specific tuning, while also supporting lightweight tuning for additional gains.

References

[1] Chang et al. "MaskGIT: Masked Generative Image Transformer." arXiv:2202.04200 (2022).

[2] Yang et al. "MMaDA: Multimodal Large Diffusion Language Models." arXiv:2505.15809 (2025).

Dear Reviewer yKia,

Thank you for the continued discussion, which has pushed us to clarify the effectiveness of our method. To directly address your valid concerns about the performance gains on MaskGIT [1], we have conducted a new experiment on MMaDA [2], a larger-scale (8B) and more recent state-of-the-art masked diffusion model that surpasses prior work like SDXL [3].

For this experiment, we evaluated image generation performance by generating 10,000 images at 512x512 resolution from 1,000 ImageNet class prompts. For KLASS, we used a KL threshold of 0.1 and a confidence threshold of 0.3. The results, using MMaDA's default 15-step schedule for a fair comparison, are summarized below:

As the results clearly demonstrate, KLASS provides a substantial improvement over the standard confidence-based sampler. Specifically, KLASS achieves a 7.75-point reduction in FID and a 10.02-point increase in IS. Achieving such a significant performance gain on a challenging, large-scale model like MMaDA is particularly meaningful, as it underscores the practical value and effectiveness of KLASS on current state-of-the-art systems. This result highlights the fundamental value of our stability-based criterion. We will conduct additional experiments using larger image sets and a wider range of step settings, and include the results in the manuscript.

We believe this new evidence on a distinct and highly relevant model definitively demonstrates that the effectiveness of KLASS is not marginal, but substantial and generalizable. We hope this addresses your concerns and would be grateful to hear any further thoughts you might have.

Thank you once again for your valuable feedback.

References

[1] Chang et al. "MaskGIT: Masked Generative Image Transformer." arXiv:2202.04200 (2022).

[2] Yang et al. "MMaDA: Multimodal Large Diffusion Language Models." arXiv:2505.15809 (2025).

[3] Podell et al. "SDXL: Improving Latent Diffusion Models for High-Resolution Image Synthesis." arXiv:2307.01952 (2023).

We are deeply grateful for your thoughtful and constructive review, which has been invaluable in improving our work. We were particularly encouraged by your acknowledgment of:

• The theoretical soundness, empirical motivation, and intuitive nature of our approach.

• Its consistent performane improvements across all testeed modalites and metrics, demonstrating practical effectiveness.

We have carefully considered the concerns you raised and provide our detailed responses below.

[W2-1] Empirical Justification of the Stability Criterion

We appreciate the reviewer’s insightful feedback regarding the interpretation of the plots and the relationship between internal model confidence, KL divergence, and correctness.

To address this, we conducted a newly measured statistical analysis across four reasoning benchmarks, comparing average confidence and KL divergence between correct and incorrect samples for both LLaDA and Dream. As detailed in the tables below, correct samples consistently exhibit higher confidence and lower KL divergence compared to incorrect samples across all tasks and models. This consistent pattern across multiple benchmarks and both models provides empirical evidence that internal confidence and KL divergence serve as reliable indicators of correctness throughout the decoding process.

LLaDA: Confidence

LLaDA: KL Divergence

Dream: Confidence

Dream: KL Divergence

We will include comprehensive results across all datasets in the final manuscript to further substantiate these findings.

Furthermore, to provide a theoretical explanation for why these empirical patterns are reliable indicators of correctness, we have developed a new information-theoretic framework.

We formalize the "unmasking risk" of a token $i$ using Conditional Mutual Information (CMI), $I_t(i) = I(X_i; Z \mid V_t)$, which measures the token's dependency on other unknown parts of the sequence. Our core theoretical result proves that the criteria used by KLASS (low cumulative KL divergence and high confidence) are sufficient to guarantee that this CMI is bounded by a provably value, $\Gamma_t(i) \le 0.6 \, \text{bits}$.

In essence, this means a token's distributional stability—which we show empirically in the tables above—is a strong indicator of its informational independence. A token that is informationally independent is "safe" to unmask and highly likely to be correct, as its value does not depend on other unknowns. This provides the formal link between our empirical observations and the success of our method.

For the full derivation and detailed discussion of this new framework, we kindly refer the reviewer to our response to Reviewer 5Uk3 [w4-1].

[W2-2] Comparison with Distillation-Based Methods

We agree that distillation can achieve greater speedups. However, we argue that our training-free method, KLASS, is a complementary and orthogonal approach, not a competing one.

A distilled model still relies on an iterative sampling process, and KLASS, as a universal sampler, is designed to optimize this remaining step. This strategy of stacking optimizations has been shown to be effective; for example, Di4C [1] further accelerated the already-distilled SDTT [2] model.

The two methods address different aspects of optimization. Distillation aims to create a new, faster model artifact through a costly, model-specific training procedure. In contrast, KLASS is a training-free sampler that optimizes the inference process for any given model. This makes KLASS a low-cost, universal, and plug-and-play tool that can be applied immediately for efficiency gains, complementing the more intensive approach of distillation.

[Q2-1]What is "Halton-based" (line 247)?

This refers to the experimental setup from the paper "Halton Scheduler for Masked Generative Image Transformer" [3] For our image generation task, we adapted the publicly released code for their confidence-based MaskGIT implementation to serve as our baseline.

[Q2-2]Is there a reason the appendix seems to just contain a completely different (assumed earlier) version of the paper?

Thank you for pointing it out. The slight differences you may have noted are intended to improve overall readability for the reviewers.

[Q2-3]Lines 146-147: This should say "KL score for corrects tends to be significantly lower", right?

Thank you for the careful reading and the helpful suggestion. You are correct. Our intended meaning is that the KL score for correct tokens is significantly lower, as shown in Figure 1b. We will revise the text to make this clearer in the final version.

References

[1] Hayakawa et al. "Distillation of Discrete Diffusion through Dimensional Correlations." ICML2025.

[2] Deschenaux et al. "Beyond Autoregression: Fast LLMs via Self-Distillation Through Time." ICLR2025.

[3] Besnier et al. "Halton Scheduler for Masked Generative Image Transformer" ICLR2025.

We sincerely appreciate your insightful review and are grateful for your acknowledgement that our work:

• Proposes a novel and intuitve method
• Provides strong and broad empricial validation across diverse domains, demonstrating both significant speedups and peformance improvements.
• Offeres a practical, training-free and plug-and-play sampler. We have carefully considered your concerns and provide our detailed responses below.

[W1-1] Hyperparameter Sensitivity

We clarify below how we select hyperparameters in practice and provide additional evidence that KLASS is robust to variations around the chosen values.

1. Hyperparameter Selection Guideline

For hyperparameter selection process, we follow a lightweight three-step search procedure using a small validation set, around 100 examples.

• Step 1. Initial KL Threshold Estimation: We estimate a reasonable KL threshold by inspecting the distribution of KL values during decoding.
• Step 2. Confidence Threshold Search: With a fixed KL threshold, we sweep confidence values (e.g., 0.9 down to 0.6) to find the best balance of accuracy and decoding speed.
• Step 3. KL Threshold Refinement: With a fixed confidence threshold, we perform a finer search for the optimal KL value around the initial estimate.

This process is fast, requiring only a small number of samples and negligible computation compared to training.

2. Hyperparameter Sensitivity Grid Search

To address sensitivity, we performed a grid search across diverse models and tasks (LLaDA/Dream for reasoning, MDLM for molecular), demonstrating KLASS's robustness.

Our findings show that configurations near the selected optimum consistently yield high accuracy with significantly reduced sampling steps. For instance, on HumanEval with LLaDA, settings near (confidence=0.9, KL=0.01) maintain or improve upon the baseline accuracy of 39.63% while using far fewer than its 256 steps. Moreover, these searches often reveal that slightly different settings can outperform our main reported configuration—such as (confidence=0.95, KL=0.005) on MBPP with Dream—indicating both robustness and potential for further tuning.

HumanEval (LLaDA) : Conf = 0.9, KL = 0.01 \ Baseline Performance: 39.63 (NFE=256)

MBPP (LLaDA): Conf = 0.7, KL = 0.01 \ Baseline Performance: 48.64 (NFE=256)

HumanEval (Dream): Conf = 0.8, KL = 0.001 \ Baseline Performance: 58.53 (NFE=256)

MBPP (Dream): Conf = 0.9, KL = 0.001 \ Baseline Performance: 63.81 (NFE=256)

Each cell: Accuracy (Steps)

Molecule (QED): Conf = 0.98, KL = 0.001 \ Baseline Performance: 0.526 (NFE=32)

Each cell: QED (Steps)

We will include the results for all datasets in our final version of the manuscript.

[W1-2] Computational Overhead of KL Divergence

The overhead is minimal because it is a lightweight post-processing step on existing logits, requiring no additional forward pass. For the set of masked tokens $I_m = \lbrace i \mid z^i_t = m \rbrace$, the process computes the KL score $d^i_t = D_{KL}(p^i_t || p^i_{t-1})$ and caches the prior distribution. This yields a combined computational $\Delta C$ and memory overhead of $O(|I_m| \cdot |V|)$. This linear cost is negligible, as the simple vector operations are insignificant compared to the expensive matrix multiplications and multi-gigabyte footprint of the main diffusion step.

Empirically, our wall-clock time measurements validate this conclusion, showing memory overheads under 1.57% and latency overheads under 0.21% per decoding step. Additional results for sequence lengths up to 1,024 are provided in [W4-4].

We will include this analysis in our revised manuscript.

[W1-3] Comparison to Other Adaptive Sampling Methods

We thank the reviewer for the valuable suggestion. We agree that a broader discussion of adaptive sampling methods, including those for continuous diffusion, will better contextualize our work. We will incorporate a full comparative analysis, summarized below, into our final manuscript.

KLASS operates dynamically, creating a sample-specific schedule on-the-fly guided by an internal measure of temporal stability (low KL divergence). This self-planning approach contrasts with methods that use fixed, pre-computed schedules from theoretical error bounds (e.g., JYS [1], DUS [2]) or rely on heavy external models for path guidance and error-correction remasking (P2 [3], DDPD [4]). Furthermore, its stability criterion advances beyond simpler heuristics like low confidence (ReMDM [5]) or high entropy (Diffusion-EAGS [6]) used for single-token correction, as it allows KLASS to finalize multiple converged tokens at once to efficiently shorten the generation path.

Furthermore, this step-reduction strategy is orthogonal and complementary to techniques that reduce per-step computational cost. Methods like caching (FreeCache [7], dKV-Cache [8]), which make each denoising step cheaper, can be combined with KLASS to compound acceleration.

While KLASS shares an adaptive spirit with solvers for continuous diffusion, their mechanisms are fundamentally different. Continuous solvers like DDIM [9] and DPM-Solver [10] reframe the reverse process as an ODE, using numerical methods to modulate a global sampling trajectory's step size based on local error, enabling 10-20 step generation. In contrast, KLASS operates in the discrete domain by adapting the inference lifespan of individual tokens—a mechanism tailored for token-level convergence.

In summary, KLASS is a lightweight, training-free sampler that uses a unique temporal stability criterion to dynamically shorten inference, setting it apart from methods that rely on fixed schedules, external planners, simpler heuristics, or orthogonal computational optimizations.

[W1-4] Wall Clock Time

Yes, we have reported wall-clock latency for MATH in the appendix (Table 8). We have now newly measured and included results for all datasets:

These results validate that KLASS achieves substantial speedups with comparable or improved accuracy across diverse reasoning tasks. Notably, we observe over 2× speedup on GSM8K for LLaDA, and over 2× speedup on HumanEval for both models, confirming that our method provides meaningful efficiency gains in practical settings.

[Typo] Thank you for pointing it out. We will thoroughly proofread and refine the final version.

References

[1] Park et al. "Jump Your Steps: Optimizing Sampling Schedule of Discrete Diffusion Models." ICLR2025.

[2] Luxembourg et al. "Plan for Speed: Dilated Scheduling for Masked Diffusion Language Models." arXiv:2506.19037 (2025).

[3] Peng et al. "Path Planning for Masked Diffusion Model Sampling." arXiv:2502.03540 (2025).

[4] Liu et al. "THINK WHILE YOU GENERATE: DISCRETE DIFFUSION WITH PLANNED DENOISING." ICLR2025.

[5] Wang et al. "Remasking Discrete Diffusion Models with Inference-Time Scaling." arXiv:2503.00307 (2025).

[6] Koh et al. "PLM-Based Discrete Diffusion Language Models with Entropy-Adaptive Gibbs Sampling." arXiv:2411.06438 (2024).

[7] Hu et al. "Accelerating Diffusion Language Model Inference via Efficient KV Caching and Guided Diffusion." arXiv:2505.21467 (2025).

[8] Ma et al. "dKV-Cache: The Cache for Diffusion Language Models." arXiv:2505.15781 (2025).

[9]Song et al. "Denoising Diffusion Implicit Models." ICLR2021.

[10]Lu et al. "DPM-Solver: A Fast ODE Solver for Diffusion Probabilistic Model Sampling in Around 10 Steps." NeurIPS2022.