- 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
-
The paper is very clearly written, and the motivation is reasonable.
-
Experiments on LLaDA and Dream demonstrate the effectiveness of the proposed method.
-
The paper provides an ablation study.
Critical weaknesses
-
Although the actual implementation differs from previous work, the core idea and motivation of this paper share many similarities with previous work.
-
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.
-
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.
-
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
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 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.
| Model | Method | GSM8K | HumanEval | MBPP | MATH500 | IFEval | Avg |
|---|---|---|---|---|---|---|---|
| LLaDA | Standard | 77.18 / 5.28 | 45.12 / 7.43 | 39.20 / 2.07 | 38.40 / 7.27 | 56.75 / 14.70 | 51.33 / 7.35 |
| Confidence | 76.35 / 17.02 | 45.12 / 16.31 | 39.20 / 13.12 | 38.80 / 18.06 | 56.75 / 26.97 | 51.24 / 18.30 | |
| KLASS | 76.95 / 12.14 | 44.51 / 10.96 | 38.60 / 8.78 | 36.80 / 13.34 | 56.93 / 22.39 | 50.76 / 13.52 | |
| WINO | 74.53 / 15.78 | 43.90 / 15.96 | 39.60 / 7.22 | 37.80 / 15.30 | 56.19 / 27.62 | 50.40 / 16.38 | |
| DCD | 77.33 / 17.69 | 45.12 / 15.98 | 40.00 / 16.19 | 37.80 / 18.67 | 56.75 / 28.42 | 51.40 / 19.39 | |
| D2D (Fixed) | 78.09 / 24.42 | 45.12 / 22.36 | 37.80 / 17.04 | 38.80 / 24.53 | 56.75 / 32.07 | 51.31 / 24.08 | |
| D2D (Tuned) | 78.24 / 25.15 | 45.12 / 22.36 | 39.20 / 18.92 | 38.80 / 24.53 | 56.93 / 32.86 | 51.66 / 24.76 | |
| Dream | Standard | 80.52 / 3.60 | 56.71 / 7.38 | 55.20 / 1.71 | 45.20 / 6.13 | 55.27 / 14.03 | 58.58 / 6.57 |
| Confidence | 80.59 / 13.67 | 59.15 / 22.65 | 55.60 / 10.32 | 46.20 / 16.47 | 56.19 / 25.50 | 59.55 / 17.72 | |
| KLASS | 81.05 / 7.06 | 60.98 / 13.90 | 56.00 / 5.98 | 46.80 / 10.26 | 53.79 / 21.22 | 59.72 / 11.68 | |
| WINO | 77.18 / 15.40 | 57.93 / 23.21 | 55.20 / 9.25 | 45.40 / 18.31 | 51.20 / 26.94 | 57.38 / 18.62 | |
| DCD | 80.67 / 16.23 | 59.15 / 25.15 | 56.80 / 15.17 | 45.80 / 17.88 | 55.27 / 27.91 | 59.54 / 20.47 | |
| D2D (Fixed) | 79.23 / 19.36 | 57.93 / 30.78 | 54.20 / 16.87 | 45.80 / 22.89 | 54.34 / 29.74 | 58.30 / 23.93 | |
| D2D (Tuned) | 80.67 / 18.34 | 58.54 / 31.65 | 56.20 / 16.97 | 46.60 / 21.38 | 55.64 / 30.75 | 59.53 / 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.
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 across five benchmarks.
| Benchmark | Confidence Acc/Pass@1 | TPS | D2D Acc/Pass@1 | TPS |
|---|---|---|---|---|
| GSM8K | 80.36 | 108.28 | 80.59 | 130.81 |
| HumanEval | 62.80 | 84.39 | 61.59 | 95.15 |
| MBPP | 49.60 | 11.86 | 49.20 | 15.44 |
| MATH500 | 45.20 | 108.69 | 45.60 | 127.38 |
| IFEval | 65.43 | 68.59 | 66.17 | 78.66 |
| Avg. | 60.68 | 76.36 | 60.63 | 89.49 |
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 on the dParallel-distilled LLaDA, under the same evaluation settings.
| Benchmark | LLaDA + D2D Acc | TPS | dParallel-LLaDA + Confidence Acc | TPS |
|---|---|---|---|---|
| GSM8K | 78.24 | 25.15 | 76.35 | 27.53 |
| HumanEval | 45.12 | 22.36 | 39.63 | 24.47 |
| MBPP | 39.20 | 18.92 | 38.40 | 14.47 |
| MATH500 | 38.80 | 24.53 | 33.80 | 25.53 |
| IFEval | 56.93 | 32.86 | 53.23 | 27.77 |
| Avg. | 51.66 | 24.76 | 48.28 | 23.95 |
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 and D2D on exactly the same dParallel-distilled LLaDA checkpoint, isolating the effect of the decoding strategy.
| Benchmark | Confidence Acc | TPS | D2D Acc | TPS | Speedup |
|---|---|---|---|---|---|
| GSM8K | 76.35 | 27.53 | 75.97 | 36.34 | 1.32× |
| HumanEval | 39.63 | 24.47 | 39.63 | 37.87 | 1.55× |
| MBPP | 38.40 | 14.47 | 38.20 | 19.82 | 1.37× |
| MATH500 | 33.80 | 25.53 | 33.40 | 35.95 | 1.41× |
| IFEval | 53.23 | 27.77 | 55.64 | 36.60 | 1.32× |
| Avg. | 48.28 | 23.95 | 48.57 | 33.32 | 1.39× |
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.
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.