acceptodds
Under review as a conference paper at ICLR 2027

Decouple to Decode: Parallel Decoding with Deferred Proposal Verification for Diffusion Language Models

Abstract

Discrete diffusion language models (dLLMs) alleviate the sequential bottleneck of autoregressive generation by predicting multiple masked tokens in parallel at each denoising step. However, existing parallel decoding methods typically make single-step commitment decisions, coupling whether a prediction is reliable enough to commit with whether it is worth retaining for further evidence, thereby tying broader proposal coverage to looser commitment criteria. Indeed, we observe two complementary regimes: one where predictions highly agree with the final tokens and are suitable for immediate commitment, and another where candidates of varying reliability remain mixed but become more separable under cross-step re-evaluation. Motivated by this observation, we propose Decouple to Decode (D2D), a training-free framework that decouples immediate commitment from deferred verification. D2D uses conditional risk budgets to commit a Core and retain a masked Shell for one-step verification; in the next ordinary forward pass, Shell proposals are accepted if their token identities remain unchanged and confidence does not decrease, requiring no extra verifier or model evaluation. Experiments on LLaDA and Dream across five benchmarks spanning mathematical reasoning, code generation, and instruction following demonstrate up to wall-clock speedup over standard decoding and % over confidence decoding, while maintaining comparable task performance. Combined with KV caching, D2D achieves up to speedup over standard decoding.

Primary area:
foundation or frontier models, including LLMs
Keywords:
Diffusion Language Models, Efficient Inference

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_vxv7OpenReview ↗
Rating
1
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes Decouple to Decode (D2D), a training-free framework that decouples immediate commitment from deferred verification. D2D uses conditional risk budgets to commit a Core and retain a masked Shell for one-step verification; in the next ordinary forward pass, Shell proposals are accepted if their token identities remain unchanged and confidence does not decrease, requiring no extra verifier or model evaluation. Experiments on LLaDA and Dream across multiple benchmarks demonstrate that the proposed method achieves significantly high speedup while maintaining comparable task performance.

Critical strengths
  1. The paper is very clearly written, and the motivation is reasonable.

  2. Experiments on LLaDA and Dream demonstrate the effectiveness of the proposed method.

  3. The paper provides an ablation study.

Critical weaknesses
  1. Although the actual implementation differs from previous work, the core idea and motivation of this paper share many similarities with previous work.

  2. D2D requires manual setting of decoding hyperparameters, and different models and benchmarks use different hyperparameter configurations. This indicates that the proposed method is sensitive to decoding hyperparameters, reducing its generalization ability and practical value.

  3. The experimental results are not presented on block diffusion-based dLLMs. Since the most advanced dLLMs are currently based on block diffusion, demonstrating the effectiveness of the proposed method on block dLLMs (such as LLaDA 2.0, Fast-DLLM-v2) is crucial.

  4. Does the proposed method outperform step-distilled methods, such as dparallel or d3llm? If the proposed method is a decoding algorithm orthogonal to step-distilling, can D2D also be applied to a step-distilled dLLM?

Questions for authors

Please see the weaknesses.

6 comments
AuthorsOpenReview ↗

Reply to weakness1

We thank the reviewer for this insightful question.

Response to W1: Distinction from Prior Work

We would like to clarify that D2D differs from prior methods in both its motivation and decoding mechanism.

First, regarding motivation, existing methods primarily focus on identifying reliable tokens for immediate commitment, using either current-step confidence or historical stability signals. However, these methods largely treat decoding as a commit-or-repredict decision, where predictions not committed at the current step are not explicitly retained as proposals. As illustrated in Figure 1, many such predictions already agree with the final output, while cross-step re-evaluation provides additional evidence for distinguishing reliable candidates. This motivates us to introduce an intermediate state between immediate commitment and ordinary re-prediction, where specific token proposals are explicitly retained for subsequent verification. Unlike directly incorporating historical stability into the current commitment criterion, this intermediate state preserves both the proposed token identity and its confidence as explicit evidence for future reassessment, enabling potentially useful predictions to be exploited without premature commitment.

Second, regarding the decoding mechanism, D2D realizes this intermediate state through a unified risk-constrained Core–Shell framework, rather than simply combining confidence- and stability-based criteria. The Core determines immediate commitment by controlling the aggregate risk of jointly committed tokens, while the Shell extends proposal coverage through a controlled relaxation of the same risk constraint. Importantly, Shell proposals remain masked and are verified against their stored identities and confidence scores in the next ordinary forward pass, followed by budget-constrained acceptance. This design decouples proposal admission from immediate commitment, allowing broader proposal coverage without directly relaxing the commitment risk budget or introducing additional model evaluations.

This distinction also separates D2D from existing approaches. KLASS, STDec, and LESS incorporate temporal evidence into immediate commitment decisions, whereas D2D explicitly preserves proposals in a separate intermediate state. DCD defers uncertain positions for re-prediction without verifying the same stored assignments, while WINO exposes tentative predictions before potential re-masking. In contrast, D2D retains specific proposals without exposing them to the decoding context until verification.

Thus, D2D differs from prior work in how uncertain predictions are handled: rather than immediately committing, discarding, or exposing tentative predictions, D2D retains them as explicit masked proposals for subsequent verification. By organizing this process within a unified risk-constrained framework, D2D decouples proposal admission from immediate commitment, rather than simply combining existing reliability criteria.

AuthorsOpenReview ↗

Reply to weakness2

Response to W2: Hyperparameter Sensitivity and Generalization

We would like to clarify that D2D supports a task-agnostic default configuration, while task-specific tuning provides additional flexibility to optimize the quality–throughput trade-off.

Specifically, D2D's two hyperparameters have distinct, task-independent roles: $\delta$ controls the aggregate risk of immediate commitment, while $\epsilon_A$ regulates the additional risk allowance for Shell admission. Since admitted Shell proposals must still pass deferred verification and budget-constrained acceptance, a relatively conservative configuration can be reused across different downstream tasks without requiring benchmark-specific parameter tuning. Meanwhile, different tasks exhibit varying sensitivities to commitment errors, leaving room for further task-specific optimization.

In Table 1, we use task-specific configurations to better explore the achievable quality–throughput trade-off across different benchmarks, rather than because such tuning is necessary for D2D to be effective. We also consider hyperparameter choices for competing methods, including KLASS and WINO, while retaining the standard Confidence threshold $\tau=0.9$ as a default reference. We further evaluate Confidence under different thresholds in our additional quality–throughput analysis (Figure 5), confirming that D2D's advantages persist across different operating points.

To validate the effectiveness of a task-agnostic configuration, we additionally evaluate D2D using a single fixed setting (δ=0.4,ϵA=0.7)(\delta=0.4,\epsilon_A=0.7) across all five benchmarks on both LLaDA and Dream below, without task-specific tuning, and compare it against all existing baselines.Each entry reports Accuracy/Pass@1 (%) / TPS.

ModelMethodGSM8KHumanEvalMBPPMATH500IFEvalAvg
LLaDAStandard77.18 / 5.2845.12 / 7.4339.20 / 2.0738.40 / 7.2756.75 / 14.7051.33 / 7.35
Confidence76.35 / 17.0245.12 / 16.3139.20 / 13.1238.80 / 18.0656.75 / 26.9751.24 / 18.30
KLASS76.95 / 12.1444.51 / 10.9638.60 / 8.7836.80 / 13.3456.93 / 22.3950.76 / 13.52
WINO74.53 / 15.7843.90 / 15.9639.60 / 7.2237.80 / 15.3056.19 / 27.6250.40 / 16.38
DCD77.33 / 17.6945.12 / 15.9840.00 / 16.1937.80 / 18.6756.75 / 28.4251.40 / 19.39
D2D (Fixed)78.09 / 24.4245.12 / 22.3637.80 / 17.0438.80 / 24.5356.75 / 32.0751.31 / 24.08
D2D (Tuned)78.24 / 25.1545.12 / 22.3639.20 / 18.9238.80 / 24.5356.93 / 32.8651.66 / 24.76
DreamStandard80.52 / 3.6056.71 / 7.3855.20 / 1.7145.20 / 6.1355.27 / 14.0358.58 / 6.57
Confidence80.59 / 13.6759.15 / 22.6555.60 / 10.3246.20 / 16.4756.19 / 25.5059.55 / 17.72
KLASS81.05 / 7.0660.98 / 13.9056.00 / 5.9846.80 / 10.2653.79 / 21.2259.72 / 11.68
WINO77.18 / 15.4057.93 / 23.2155.20 / 9.2545.40 / 18.3151.20 / 26.9457.38 / 18.62
DCD80.67 / 16.2359.15 / 25.1556.80 / 15.1745.80 / 17.8855.27 / 27.9159.54 / 20.47
D2D (Fixed)79.23 / 19.3657.93 / 30.7854.20 / 16.8745.80 / 22.8954.34 / 29.7458.30 / 23.93
D2D (Tuned)80.67 / 18.3458.54 / 31.6556.20 / 16.9746.60 / 21.3855.64 / 30.7559.53 / 23.82

The results show that D2D-Fixed achieves the highest throughput across all ten model–benchmark settings, outperforming every competing baseline without task-specific tuning. Compared with Confidence decoding, its average TPS improves by 31.6% on LLaDA and 35.0% on Dream. Even compared with DCD, the strongest competing baseline in average throughput, D2D-Fixed remains 24.2% and 16.9% faster, respectively. Meanwhile, average task performance remains close to Standard decoding, with only 0.02 and 0.28 percentage-point decreases.

Task-specific tuning further recovers generation quality while preserving these acceleration advantages. Compared with D2D-Fixed, D2D-Tuned improves average task performance by 0.35 and 1.23 percentage points on LLaDA and Dream, respectively, exceeding Standard decoding on both models. Importantly, its average TPS remains 35.3% and 34.4% higher than Confidence decoding.

Overall, D2D does not require benchmark-specific hyperparameter tuning to achieve consistent acceleration, as a single configuration outperforms all competing baselines in throughput across both models and all five tasks. Task-specific tuning further recovers generation quality while preserving these gains. These results support the practical applicability of D2D with a reusable default configuration, rather than relying on extensive per-task tuning.

AuthorsOpenReview ↗

Reply to weakness3&4

Response to W3: Generalization to Block-Diffusion dLLMs

D2D extends naturally to modern block-diffusion architectures. Its Core–Shell mechanism acts on active masked positions, enabling deferred verification without retraining or extra model evaluations.

We validate this on Fast-dLLM v2 using its official default settings and KV cache, comparing D2D against Confidence decoding (τ=0.9)(\tau=0.9) across five benchmarks.

BenchmarkConfidence Acc/Pass@1TPSD2D Acc/Pass@1TPS
GSM8K80.36108.2880.59130.81
HumanEval62.8084.3961.5995.15
MBPP49.6011.8649.2015.44
MATH50045.20108.6945.60127.38
IFEval65.4368.5966.1778.66
Avg.60.6876.3660.6389.49

The results demonstrate consistent acceleration across all five benchmarks while largely preserving generation quality. D2D achieves 1.13–1.30× speedups, improving average throughput from 76.36 to 89.49 TPS (+17.2%), with nearly unchanged average task performance (60.68% vs. 60.63%). Notably, D2D improves throughput by 30.2% on MBPP and 20.8% on GSM8K, while also improving accuracy on GSM8K, MATH500, and IFEval. These results show that deferred verification continues to provide additional acceleration opportunities within block-diffusion architectures.

Overall, these results establish D2D's effectiveness on modern block-diffusion dLLMs, extending its demonstrated applicability beyond LLaDA and Dream.

Response to W4: Comparison and Compatibility with Step-Distilled dLLMs

D2D provides competitive training-free acceleration and is complementary to step distillation, enabling further efficiency improvements on distilled models. While step-distillation methods improve parallel prediction capabilities through training, D2D exploits additional decoding opportunities through decoupled commitment and deferred verification, without modifying model parameters.

We test both claims using dParallel's official step-distilled LLaDA checkpoint.

1. Comparison with Step-Distilled Decoding.

We first compare D2D on the original LLaDA with Confidence decoding (τ=0.9)(\tau=0.9) on the dParallel-distilled LLaDA, under the same evaluation settings.

BenchmarkLLaDA + D2D AccTPSdParallel-LLaDA + Confidence AccTPS
GSM8K78.2425.1576.3527.53
HumanEval45.1222.3639.6324.47
MBPP39.2018.9238.4014.47
MATH50038.8024.5333.8025.53
IFEval56.9332.8653.2327.77
Avg.51.6624.7648.2823.95

As shown in the table above, D2D achieves higher task performance across all five benchmarks, improving average quality from 48.28% to 51.66%. Despite requiring no additional training, D2D also achieves slightly higher average throughput (24.76 vs. 23.95 TPS). These results demonstrate that D2D provides a competitive training-free alternative to step-distilled decoding.

2. Compatibility with Step Distillation.

More importantly, D2D can be directly applied to step-distilled models to achieve additional acceleration. To verify this, we evaluate Confidence decoding (τ=0.9)(\tau=0.9) and D2D on exactly the same dParallel-distilled LLaDA checkpoint, isolating the effect of the decoding strategy.

BenchmarkConfidence AccTPSD2D AccTPSSpeedup
GSM8K76.3527.5375.9736.341.32×
HumanEval39.6324.4739.6337.871.55×
MBPP38.4014.4738.2019.821.37×
MATH50033.8025.5333.4035.951.41×
IFEval53.2327.7755.6436.601.32×
Avg.48.2823.9548.5733.321.39×

As shown in the table above, D2D consistently improves throughput across all five benchmarks, achieving 1.32–1.55× speedups. Average TPS increases from 23.95 to 33.32 (+39.1%), while average task performance remains comparable (48.28% vs. 48.57%). Notably, these improvements require no further training or modifications to the distilled model.

Overall, D2D is competitive with the evaluated step-distilled baseline and can be applied orthogonally to step distillation. It surpasses dParallel-LLaDA + Confidence in average quality and slightly in throughput without extra training. On the same distilled checkpoint, D2D provides a further 1.39× speedup at comparable quality, confirming that the two approaches can be combined effectively.

Reviewer_oL2VOpenReview ↗
Rating
3
Confidence
3
Informativeness
1
Presentation
2
Soundness
3
Read review
Paper summary

This paper studies the problem of early commitment in diffusion language model. It identified the issue of existing confidence-based method, i.e., among the medium-conf tokens, there exists the final correct options. Therefore, it proposes a fine-grained heuristic to further filters these tokens by comparing the consistency of token between two decoding steps. It conducted experiments on 5 benchmarks and demonstrate good performance and speedups.

Critical strengths
  1. the experimental side is solid, for the main results, I saw 4-5 baselines, 5 benchmarks are compared including the most relevant one KLASS. It is obvious the method strength lies on the speedup improvement.
  2. The ablation study is thorough to covers hyperparameter, design choices which justifies the claims of the methods.
Critical weaknesses
  1. The method is to combines the confidence-based method and stabilty-based method in a reasonable way. That is, confidence-based commits high-conf tokens, stability-based method for med-conf tokens, and for the stability-based criterion, they use the one that argmax is unchanged and argmax token confidence is not decreasing. it is admittedly conceptually incremental even with good improvement and speedup.
Questions for authors

I see the algorithm itself brings difficulty to tune the confidence hyper for shell set selection. There exists a hyperparameter study in Table3. But from Table 3, it looks like the range of medium "shell" set can be pretty large? Only the full remaining set of shell set hurts performance, and 0.9 can even improve, why does that happen? Is that means only applying stabilty-method on the extremely low conf token hurt?

4 comments

We thank the reviewer for this insightful question.

Response to Weakness: On the Novelty of D2D

D2D is not simply a combination of confidence-based and stability-based decoding, but a structural redesign motivated by the limitations of existing decoding strategies.

First, regarding motivation, existing methods primarily focus on identifying reliable tokens for immediate commitment using current-step confidence or historical stability signals. However, these methods largely couple proposal selection with token commitment, leaving potentially reliable predictions just below the commitment boundary underutilized. As shown in Figure 1, many such predictions already agree with the final output, while cross-step re-evaluation provides additional evidence for distinguishing reliable candidates. This motivates D2D to introduce an intermediate state between immediate commitment and ordinary re-prediction, allowing potentially useful proposals to be retained without premature commitment.

Second, regarding the decoding mechanism, D2D realizes this intermediate state through a unified risk-constrained Core–Shell framework. The Core controls the aggregate risk of immediate commitment, while the Shell extends proposal coverage under a relaxed risk budget. Unlike KLASS, STDec, and LESS, which incorporate temporal evidence into commitment decisions, D2D explicitly retains historical proposals for subsequent verification. It also differs from DCD, which defers uncertain positions for re-prediction, and WINO, which exposes tentative predictions before potential re-masking. Instead, D2D keeps Shell proposals masked and verifies their stored identities and confidence scores in the next ordinary forward pass, without additional model evaluations.

Thus, the novelty of D2D lies in decoupling proposal admission from immediate commitment through an explicit intermediate proposal state, rather than simply combining existing reliability criteria.

Response to Question: On Shell Admission and Hyperparameter Tuning

A larger Shell expands the pool of proposals awaiting verification rather than directly increasing token commitment, and D2D does not require task-specific tuning to achieve substantial acceleration.

Regarding the relatively broad Shell range, as specified in Eqs. (4)–(5), $\epsilon_A$ controls the conditional risk allowance for Shell admission, while admitted proposals must subsequently pass identity consistency, non-decreasing confidence, and budget-constrained acceptance. Therefore, increasing $\epsilon_A$ creates more opportunities to exploit reliable proposals, but does not necessarily increase the number of accepted tokens.

The degradation at ϵA =1.0\epsilon_A\ = 1.0 can be further explained by the risk constraint. When ϵA <1\epsilon_A\ < 1, Shell admission maintains a positive lower bound on the product of candidate confidence scores. At ϵA =1.0\epsilon_A\ = 1.0, this uniform protection disappears, allowing substantially lower-confidence proposals to enter when the Core and verified-set risks are small. Some unreliable proposals may still pass verification, increasing the risk of erroneous commitment. Thus, the degradation does not imply that only extremely low-confidence tokens are harmful; rather, it reflects the potential risk of overly permissive Shell admission. Different tasks may also exhibit different sensitivities to such errors, motivating the flexibility provided by $\epsilon_A$.

Regarding hyperparameter tuning, we additionally evaluate D2D using a single fixed configuration (δ=0.4\delta=0.4,$\epsilon_A=0.7$) across all five benchmarks on both LLaDA and Dream, without task-specific tuning.

ModelMethodAvg. Acc/Pass@1Avg. TPS
LLaDAStandard51.337.35
Confidence51.2418.30
D2D (Fixed)51.3124.08
D2D (Tuned)51.6624.76
DreamStandard58.586.57
Confidence59.5517.72
D2D (Fixed)58.3023.93
D2D (Tuned)59.5323.82

The results show that D2D retains substantial acceleration advantages under fixed hyperparameters, achieving 31.6% and 35.0% higher average TPS than Confidence decoding on LLaDA and Dream, respectively, with only 0.02 and 0.28 percentage-point quality decreases relative to Standard decoding. Task-specific tuning further improves average quality by 0.35 and 1.23 percentage points, while preserving these throughput advantages.

Overall, a broad Shell can provide additional verification opportunities without directly increasing commitment, while its admission risk is controlled by $\epsilon_A$. More importantly, the fixed-parameter results demonstrate that D2D's acceleration benefits do not depend on task-specific tuning; such tuning primarily provides additional flexibility to recover generation quality.

Thank you again for your detailed review and valuable suggestions. As the discussion period draws to a close, we would be grateful if you could let us know whether our responses have addressed your concerns. If anything remains unclear, we would be happy to discuss it further.

I thank the authors' rebuttal, which convinces me the propose method is more of a solid algorithmic trick instead of a combination. I also thank the additional results. I will maintain my positive rating (3), but it doesn't meet the expectation of 4 compared to other papers in my batch. I encourage other reviewers check the rebuttal.

We sincerely thank the reviewer for your constructive feedback and positive reassessment. We greatly appreciate your time and thoughtful consideration.

Reviewer_bhVvOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes D2D, a training-free decoding method for masked diffusion language models. It commits a confidence-budgeted Core immediately while retaining masked Shell proposals for verification in the next regular forward pass. Experiments on two models and five benchmarks show improved throughput with generally comparable task performance.

Critical strengths
  1. The method is simple, training-free and requires no additional model forward pass for verification.
  2. The Experiments show clearly strong performance on a variety of tasks.
Critical weaknesses
  1. The individual contributions of risk-budgeted selection and deferred verification remain unclear, as the ablations do not fully isolate their effects.

  2. The reported results demonstrate throughput gains at selected operating points, but do not clearly establish an improved quality–throughput frontier over existing methods.

Questions for authors

Questions:

  1. Theorem 3.1 establishes one-step selection optimality, while Proposition 3.2 guarantees non-worsening confidence scores for retained proposals. Can the authors clarify what these results establish about deferred verification beyond budgeted Core-only decoding, and provide empirical evidence of its additional benefit?
  2. Table 1 uses a fixed Confidence threshold of 0.9, while D2D uses task-specific hyperparameters (Appendix C). Could the authors also sweep the Confidence threshold and compare quality–throughput curves to assess whether D2D’s gains persist at matched quality?
1 comment

We thank the reviewer for this insightful question.

Response to W1/Q1: Individual Contributions of Budgeted Core and Deferred Verification

Deferred verification provides additional acceleration beyond budgeted Core-only decoding, while both components are jointly designed within a unified risk-constrained framework. Theorem 3.1 characterizes Core selection under the original risk budget, while Proposition 3.2 establishes non-worsening confidence-based reward–risk terms for verified proposals. Importantly, the aggregate-risk formulation of Core also provides the basis for Shell admission: the Shell extends candidate coverage through a controlled relaxation of the same risk constraint, while deferred verification restores the original budget for subsequent commitment. Thus, the two components form a coherent mechanism for expanding proposal coverage without directly relaxing immediate commitment.

To validate these contributions, we introduce a Budgeted Core-only baseline that retains the same risk-budgeted Core selection but disables Shell admission and deferred verification. With Table 3, we compare four configurations:

MethodGSM8K AccTPSHumanEval Pass@1TPS
(1) Confidence76.3517.0245.1216.31
(2) Confidence + Joint Verification77.4120.5342.6821.11
(3) Budgeted Core-only77.5621.6644.5119.91
(4) Full D2D78.2425.1545.1222.36

The additional benefit of deferred verification is directly demonstrated by comparing (3) and (4). With the same Budgeted Core, introducing deferred verification improves throughput by 16.1% on GSM8K and 12.3% on HumanEval, while also improving task performance by 0.68 and 0.61 percentage points, respectively. This confirms that retaining and verifying proposals enables acceleration beyond Core-only decoding.

Meanwhile, comparisons (1) vs. (3) and (2) vs. (4) demonstrate the benefit of Budgeted Core selection. It improves throughput even without deferred verification and improves both quality and throughput when combined with the same Shell verification mechanism.

Overall, these results demonstrate that deferred verification provides additional benefits beyond Core-only decoding, while the shared risk-constrained design enables Core selection and Shell verification to work together effectively. Rather than simply combining two independent techniques, D2D uses budgeted commitment to establish a reliable boundary and deferred verification to safely exploit additional proposals beyond it.

Response to W2/Q2: Quality–Throughput Trade-off

D2D achieves a more favorable quality–throughput trade-off than Confidence decoding, and its acceleration advantage persists even at matched or higher generation quality. This advantage is not limited to the fixed Confidence threshold used in Table 1.

We conduct additional hyperparameter sweeps on LLaDA GSM8K and HumanEval. We vary Confidence threshold $\tau$ from 0.5 to 0.9 and explore different combinations of D2D's commitment budget $\delta$ and Shell admission parameter $\epsilon_A$. Figure 5 presents the resulting quality–throughput curves, while the table provides representative quality-matched comparisons. For each Confidence operating point shown in the table, we select the evaluated D2D configuration with highest TPS while maintaining equal or higher task performance.

BenchmarkConfidence QualityTPSD2D QualityTPSΔ QualityTPS Gain
GSM8K76.3517.0276.5731.05+0.22+82.4%
77.1021.9677.6328.82+0.53+31.2%
75.1327.3375.5135.54+0.38+30.0%
HumanEval45.1216.3145.1222.61±0.00+38.6%
41.4620.5742.0726.52+0.61+28.9%
35.9825.5437.2037.02+1.22+44.9%

The results show that D2D consistently achieves higher throughput at matched or improved quality across these operating points. On GSM8K, D2D achieves 30.0%–82.4% higher throughput while maintaining higher accuracy. On HumanEval, throughput improves by 28.9%–44.9%, including a 38.6% improvement at identical Pass@1 (45.12%). The trade-off curves further show that these gains extend beyond individual parameter configurations, with D2D maintaining a more favorable quality–throughput relationship across a broad range of operating points.

These improvements are consistent with D2D's decoupled decoding design: unlike directly relaxing the Confidence threshold to increase parallel commitment, D2D retains additional proposals for deferred verification, enabling greater decoding progress without immediately committing uncertain predictions.

Overall, these results directly address the quality–throughput concern: D2D's acceleration advantage persists after tuning the Confidence baseline and remains substantial at matched or higher generation quality.

Comments on the paper

2 comments

We sincerely thank the reviewer for your constructive feedback and positive reassessment. We greatly appreciate your time and thoughtful consideration.

AuthorsOpenReview ↗

Rebuttal Summary by Authors

We sincerely thank all reviewers for their careful evaluation and constructive feedback. We have addressed their concerns through additional experiments, controlled comparisons, theoretical clarifications, and manuscript revisions. Below, we summarize the recognized strengths of D2D and the key improvements made during the rebuttal.

1. Common Ground & Highlights

  • Promising Decoding Efficiency (bhVv, oL2V): Reviewers acknowledged the substantial acceleration achieved by D2D. Reviewer bhVv further confirmed that our additional Core-only ablation and threshold sweeps addressed their main concerns and strengthened the paper.

  • Meaningful Algorithmic Contribution (oL2V): Following our clarifications, reviewer oL2V recognized D2D as a solid algorithmic technique rather than a simple combination of existing decoding strategies, supporting the distinction of our decoupled decoding design.

  • Positive Reassessment (bhVv, oL2V): Reviewer bhVv raised their recommendation from Weak Reject to Weak Accept, while reviewer oL2V reaffirmed their positive Weak Accept rating after considering our rebuttal.

2. Summary of Revisions

In response to the reviewers' feedback, we conducted additional experiments and clarified the manuscript (revisions highlighted in blue):

  • Individual Contributions and Theoretical Clarification (bhVv): We added a Budgeted Core-only ablation (Table 3), isolating the benefits of Core selection and deferred verification. The results confirm that deferred verification provides additional acceleration beyond Core-only decoding. We also clarified how both components form a unified risk-constrained framework rather than a simple combination of independent techniques.

  • Quality–Throughput Trade-offs (bhVv): We conducted additional hyperparameter sweeps and quality-matched comparisons on GSM8K and HumanEval (Figure 5), demonstrating that D2D retains substantial throughput advantages at comparable or higher generation quality, even after tuning the Confidence baseline.

  • Novelty and Distinction from Prior Work (oL2V, vxv7): We clarified D2D's motivation and distinctions from history-aware, deferred-commitment, and revocable decoding methods. We emphasized the explicit masked proposal state and decoupled commitment–verification mechanism, establishing the methodological difference from existing reliability-based decoding strategies.

  • Hyperparameter Robustness and Shell Admission (oL2V, vxv7): We added unified-hyperparameter evaluations across two models and five benchmarks (Appendix), showing that D2D maintains its throughput advantages without task-specific tuning, while tuning further recovers generation quality. We also clarified how the Shell admission budget balances proposal coverage and verification risk.

  • Generalization to Modern Block-Diffusion Models (vxv7): We extended our evaluation to Fast-dLLM v2 with native KV caching (Appendix). The results demonstrate consistent acceleration with comparable generation quality, supporting D2D's applicability to modern block-diffusion architectures.

  • Comparison and Compatibility with Step Distillation (vxv7): We added comparisons using the official dParallel-distilled LLaDA checkpoint (Appendix), demonstrating competitive training-free performance and confirming that D2D can be combined orthogonally with step distillation to achieve additional acceleration.

We sincerely appreciate the positive feedback and encouraging reassessments from bhVv and oL2V, particularly their recognition of our additional experiments and clarifications. We also welcome further discussion with the remaining reviewer and would be happy to address any additional questions or concerns.

We believe these revisions have strengthened the empirical evidence, methodological clarity, and practical applicability of D2D. Detailed point-by-point responses have been provided under each review, and we remain open to further feedback.

Best regards,
Authors of D2D

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.