acceptodds
Under review as a conference paper at ICLR 2027

RAPD: Repulsion-Attraction Energy for Parallel Decoding of Diffusion Large Language Models

Abstract

Parallel decoding accelerates masked diffusion large language models by committing multiple positions per forward pass, but selecting a set of mutually compatible positions remains challenging. Existing methods exploit inter-position cues such as locality and dependency, but most of them model these interactions in only one direction, either promoting or suppressing joint commitment. We propose repulsion-attraction parallel decoding (RAPD), a training-free strategy that combines individual commitment evidence and signed pairwise interactions in a single energy function for commit-set selection. Nearby positions attract each other and promote joint commitment, whereas positions that compete for the same candidate tokens repel each other and suppress conflicting commitment. RAPD iteratively refines commitment probabilities using variational updates, without additional model forward passes. On LLaDA-8B-Instruct, RAPD uses a single configuration across GSM8K, MATH, HumanEval, and MBPP, achieving the fewest forward passes among the compared methods while maintaining competitive accuracy. On HumanEval, RAPD matches the pass@1 of Fast-dLLM with 27.8% fewer forward passes and achieves 4.73x the throughput of sequential decoding. On the same benchmark, threshold sweeps further show that RAPD matches or exceeds the accuracy of Fast-dLLM and LocalLeap with fewer forward passes across all tested baseline thresholds. Our analysis shows that attraction reduces decoding steps, while repulsion mitigates the accuracy loss from attraction-only decoding.

Primary area:
foundation or frontier models, including LLMs
Keywords:
diffusion language models, parallel decoding, energy-based modeling, variational inference

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_WLcsOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper found that there are conflicts among the individually confident predictions during parallel decoding in dLLMs. To handle this problem, this paper proposed RAPD, which is a training-free method to deal with selections based on energy. The energy combines each position's unary commitment evidence with signed pairwise interactions. The results on LLaDA-8B-Instruct and Dream-7B-Instruct across 4 benchmarks show that RAPD achieves the fewest forward passes and the highest throughput on most of the experiments.

Critical strengths
  1. Well-motivated. The collision example shows why independent per-position thresholding fails. Hence, it is natural to frame commit-set selection as an energy with attractive and repulsive interactions.
  2. Light and practical. RAPD is a training-free method and needs no extra forward passes.
  3. Strong results in efficiency. On LLaDA-8B, RAPD has the lowest NFE and highest TPS on all four benchmarks with a single configuration.
Critical weaknesses
  1. Limited generalization. The paper noted that their method's weights were selected on HumanEval, where it also achieved the best performance. However, on other benchmarks, the accuracy is at or near the bottom among the compared methods, which is not stable.
  2. Mismatch between the motivation and the method's effect. The paper motivates RAPD to reduce prediction collisions. Repulsion is the only mechanism in RAPD that could prevent a conflicting commitment, but repulsion alone does not improve on no coupling.
  3. Limited robustness. All main results use a single generation shape (length 256, block size 32). However, when the generation length is increased to 512 on HumanEval, every baseline improves by 4.9-6.1 points, whereas RAPD shows no improvement.
Questions for authors
  1. Why were the weights of RAPD tuned on HumanEval? If the weights were selected on another dataset, would the resulting configuration and the reported gains remain the same?
  2. Why does RAPD have no benefit from a longer generation length on HumanEval? Does re-tuning the threshold or weights for longer lengths recover this gap?
  3. Is RAPD compatible with KV-Cache (e.g., Fast-dLLM’s prefix or dual cache)? How does performance compare in that setting and at larger batch sizes?
3 comments

We thank the reviewer for the thoughtful review and helpful comments.


[Q1/W1] Configuration selection on a different dataset.

Thank you for the question. HumanEval served as our initial development benchmark for selecting the configurations, which we then fixed across all four tasks. To directly test whether the gains depend on this choice, we repeated configuration selection using development data, without using test results for selection.

Using only the MBPP training and validation splits (464 problems), we reselected the coupling parameters, unary criterion, readout layer, kernel width, and per-block step cap. We selected the setting with the highest development accuracy, breaking ties by fewer forward passes. Most settings remained unchanged: κ\kappa increased from 2 to 4, with the repulsion weight adjusted to preserve the coupling budget, and the per-block step cap increased from 16 to 32.

We evaluated this single configuration across all four test sets; each entry below reports (problems solved / NFE).

BenchmarkRAPD (paper)RAPD (dev-selected)Fast-dLLM
HumanEval (164)66 / 56.363 / 65.866 / 78.0
MBPP (500)205 / 49.4210 / 51.0211 / 69.1
GSM8K (1,319)1,026 / 55.91,041 / 55.21,034 / 78.5
MATH (5,000)1,578 / 69.51,622 / 71.21,649 / 97.1

The development-selected configuration uses 16–30% fewer forward passes than Fast-dLLM, with no significant accuracy difference on any benchmark. These results show that RAPD’s efficiency gains persist when configuration selection uses separate development data.


[W2] Effect of Repulsion.

Thank you for this observation. We apologize for any confusion. The role of repulsion is to preserve accuracy as attraction expands the commit set. This is the effect stated in the abstract: repulsion "mitigates the accuracy loss from attraction-only decoding."

The limited standalone benefit is consistent with the stability of high-confidence predictions under threshold decoding. We first examined co-committed pairs on HumanEval and 200 problems each from MBPP and GSM8K. Committing one position and rerunning the model changed the other position’s most likely token in only 370 of 472,453 pairs (0.08%).

The attraction mechanism expands the commit set to include positions below the confidence threshold. Repulsion becomes effective in this more aggressive regime, as shown by the controlled ablation below. Each entry reports (problems solved / NFE).

ConfigurationHumanEval (164)MBPP (500)
No coupling59 / 64.4201 / 56.5
Repulsion only60 / 68.4200 / 59.2
Attraction only55 / 51.9191 / 46.3
RAPD (both)66 / 56.3205 / 49.4

Adding repulsion to attraction improves accuracy on both code benchmarks, while retaining approximately 13% fewer forward passes than no coupling. On HumanEval, it also reduces adjacent same-token co-commitments from 822 to 536, with the same-token share among all co-committed pairs decreasing from 5.35% to 4.25%.

These results support the intended interaction: attraction increases parallel commitment, while repulsion recovers the accuracy lost under attraction-only decoding and preserves the efficiency gain. We will clarify these complementary roles in the revision.

[W3/Q2] Performance at longer generation lengths.

We sincerely thank the reviewer for highlighting this issue. The reviewer's feedback helped us close the accuracy gap at length 512 by revisiting the decoding settings.

We'd like to note that increasing the generation length also substantially increases the decoding cost. For the three baselines, extending the length from 256 to 512 improves HumanEval accuracy by 4.9–6.1 percentage points but requires 2.03–2.11× as many forward passes. The longer setting therefore represents a substantially larger computational budget, and its accuracy gains should be assessed alongside this cost. Given our focus on efficient decoding, we investigated whether RAPD can recover accuracy at this longer length while retaining its forward-pass savings.

The accuracy gap at length 512 is closed by reselecting the commitment threshold and per-block step cap on development data. Here, we take the MBPP development set at length 512 and the same selection rule as in Q1. The resulting HumanEval score increases significantly from 66 to 77 of 164 problems at generation length 512.

MethodHumanEval (164)GSM8K (1,319)
RAPD (paper settings)66 / 117.21,027 / 70.0
RAPD (dev-selected threshold and cap)77 / 1391,025 / 81
Fast-dLLM74 / 164.91,031 / 95.1
LocalLeap75 / 124.31,026 / 72.8
DAWN72 / 130.01,009 / 70.4

The new RAPD setting shows no significant accuracy difference from any baseline on either benchmark, while requiring approximately 15–16% fewer forward passes than Fast-dLLM.

Our analysis indicates that the original decoding settings commit too aggressively at length 512. The commitment threshold and per-block step cap should be selected considering the generation length and block size; the threshold controls how aggressively positions are committed, while the step cap sets the decoding budget for each block. We will clarify this distinction in the revised manuscript.


[Q3] Compatibility with KV caching and larger batch sizes.

KV caching. Yes, RAPD is compatible with Fast-dLLM’s PrefixCache and DualCache. We evaluated both decoding methods using the same cache implementation on LLaDA-8B-Instruct with a single RTX 4090.

Each entry reports RAPD / Fast-dLLM. Accuracy is the number of problems solved, and TPS is measured end to end at batch size 1.

BenchmarkCacheProblems solvedNFETPS
HumanEval (164)Prefix70 / 6970.5 / 84.588.9 / 76.0
HumanEval (164)Dual58 / 5877.9 / 90.761.5 / 55.1
MBPP (500)Prefix203 / 21159.9 / 69.656.9 / 50.4
MBPP (500)Dual189 / 19359.3 / 69.345.0 / 40.3
GSM8K (1,319)Prefix1,034 / 1,02568.2 / 81.263.9 / 54.7
GSM8K (1,319)Dual1,035 / 1,04272.2 / 85.352.5 / 46.1
MATH (5,000)Prefix1,615 / 1,64986.6 / 100.865.4 / 57.4
MATH (5,000)Dual1,615 / 1,63592.8 / 107.150.3 / 45.0

Across all eight combinations, RAPD requires 13–17% fewer forward passes and achieves 12–17% higher TPS, with no significant accuracy difference from Fast-dLLM. These results demonstrate the effectiveness of RAPD on top of KV-cache acceleration.

Batch size. We conducted a trace-based analysis of block-synchronous batching using per-problem decoding traces collected under PrefixCache. We formed random batches and counted the maximum number of steps required by their constituent sequences within each block.

At the paper’s settings, the estimated NFE reduction over Fast-dLLM increases from approximately 30% at batch size 1 to 35% at batch size 16, using all 164 HumanEval problems and the GSM8K problems. RAPD’s per-block step cap limits the waiting time caused by slower sequences.

We sincerely thank the reviewer again for suggesting these valuable experiments, which helped us strengthen the evaluation of RAPD under practical inference settings.

Dear Reviewer WLcs,

Thank you again for your time and constructive feedback.

Your suggestions have substantially improved our paper. In response, we made several additions, most notably repeating the configuration selection on held-out development data, clarifying the role of repulsion through controlled ablations, re-selecting the decoding settings for longer generations, and evaluating RAPD with KV caching. The revised manuscript has been uploaded, with all changes highlighted in red.

As the discussion period ends soon, we would sincerely appreciate your feedback on whether our responses have addressed your concerns. We would be glad to answer any remaining questions.

Best regards,

The Authors

Reviewer_V5RNOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

RAPD selects parallel commit sets with a signed pairwise energy. Locality supplies attraction. Embedding decision directions supply repulsion. Mean-field updates refine commit probabilities without extra backbone evaluations. A contraction bound controls the coupling strength. Tests cover LLaDA, Dream, and four reasoning or coding benchmarks.

Critical strengths
  1. The signed energy unifies promotion and suppression of simultaneous commitments. The contraction analysis makes refinement stability checkable. This is a useful design principle.

  2. The paper reports throughput as well as forward counts. Appendix G isolates attraction and repulsion at matched settings. Two backbones provide some evidence of transfer.

Critical weaknesses
  1. The main HumanEval configuration was selected on the reported 164-problem test set (Appendix C). About ten coupling arms were tested. Other reported tasks also informed design choices. This weakens the strongest frontier claim. A separate development set is needed.

  2. The novelty and empirical scope remain limited. Mean-Field Parallel Decoding [1] already uses binary energies and variational fixed-point commitment updates. The signed interactions are a plausible extension. Their incremental value needs a direct comparison. Evaluation covers only LLaDA 1.0 and Dream. Transfer to newer backbones remains untested. Appendix H also reports weaker HumanEval accuracy at larger blocks and length 512. Together, these gaps limit the general acceleration claim.

[1] Mean-Field Parallel Decoding for Discrete Diffusion Language Models

Questions for authors
  1. What gains remain after selecting all parameters on separate development data? Please report the untouched HumanEval result.

  2. How does RAPD compare with Mean-Field Parallel Decoding under matched latency and tuning budgets? Does decision-direction similarity predict harmful co-commitments beyond confidence and distance?

  3. Do quality and wall-clock gains persist on newer backbones, such as LLaDA 2.1 or SDAR? Please compare their native decoders under matched settings. If incompatible, which assumptions restrict applicability?

  4. The comparison and discussion of revision-based decoding need more depth. WINO [2] verifies and revokes drafts. ReMix [3] refines continuous token states. How does RAPD’s prevention of conflicting commitments differ from these approaches? The distinction from dependency-aware selectors, including DAWN [4] and LocalLeap [5], also needs clarification. Mean-Field Parallel Decoding [1] is the closest empirical comparator. Revision methods provide complementary context.

[2] Wide-In, Narrow-Out: Revokable Decoding for Efficient and Effective DLLMs

[3] Rejection Mixing: Fast Semantic Propagation of Mask Tokens for Efficient DLLM Inference

[4] DAWN: Dependency-Aware Fast Inference for Diffusion LLMs

[5] Accelerating Diffusion LLM Inference via Local Determinism Propagation

5 comments

We thank the reviewer for the thoughtful review and for recognizing the signed-energy formulation and stability analysis. The reviewer’s questions helped us strengthen both the empirical evaluation and the discussion of related methods.


[W1/Q1] Configuration selection on separate development data.

Thank you for highlighting this concern. HumanEval served as our initial development benchmark for selecting the coupling weights. To test whether the gains depend on this choice, we repeated configuration selection using separate development data, without using test results for selection.

Using the MBPP training and validation splits (464 problems), we selected the coupling parameters, unary criterion, readout layer, kernel width, and per-block step cap. We selected the configuration with the highest development accuracy, breaking ties by fewer forward passes. Most settings remained unchanged: κ\kappa increased from 2 to 4, with the repulsion weight adjusted to preserve the coupling budget, and the per-block step cap increased from 16 to 32. We kept the commitment threshold at the paper’s τ=0.85\tau=0.85.

We evaluated this single configuration across all four test sets. Each entry reports (problems solved / NFE).

BenchmarkRAPD (paper)RAPD (dev-selected)Fast-dLLM
HumanEval (164)66 / 56.363 / 65.866 / 78.0
MBPP (500)205 / 49.4210 / 51.0211 / 69.1
GSM8K (1,319)1,026 / 55.91,041 / 55.21,034 / 78.5
MATH (5,000)1,578 / 69.51,622 / 71.21,649 / 97.1

The development-selected configuration uses 16–30% fewer forward passes than Fast-dLLM, with no significant accuracy difference on any benchmark. These results show that RAPD’s efficiency gains persist when configuration selection uses separate development data.


[W2/Q2a] Direct comparison with Mean-Field Parallel Decoding.

Thank you for identifying MFPD [1] as the closest related method. We reimplemented its Algorithm 1 because the official repository was unavailable, using its published LLaDA configuration: block size 20, τ\tau=0.85, and no KV cache. Using the corresponding scoring protocols, our reproduction is within 2.3 percentage points of its reported accuracy on all four tasks. For the comparisons below, both methods use the same scoring procedure.

At the default settings, the results are as follows.

BenchmarkRAPDMFPD
HumanEval (164)66 / 56.363 / 61.2
MBPP (500)205 / 49.4206 / 53.0
GSM8K (1,319)1,026 / 55.91,008 / 65.1
MATH (5,000)1,578 / 69.51,620 / 77.5

RAPD uses 7–14% fewer forward passes across all four benchmarks, with no significant accuracy difference. Measurements on the same machine (RTX 4090) also show lower runtime: 9.5 vs. 11.1 seconds per GSM8K problem, a reduction of approximately 14%.


[W2/Q2b] Contribution beyond MFPD.

We appreciate the opportunity to clarify this distinction. Both methods use a binary energy and mean-field updates. RAPD’s central contribution is to jointly model attraction and repulsion within this formulation.

The main differences are:

  • Signed interactions. MFPD’s interactions only suppress commitment. RAPD also promotes commitment through attraction, allowing positions below the confidence threshold to join the commit set. Repulsion helps preserve accuracy as this set expands.
  • Interaction representation. MFPD compares full predictive distributions through JSD. RAPD combines positional locality with similarity between output-embedding decision directions, retaining only mutually selected neighbors for repulsion.
  • Coupling control. RAPD bounds the total interaction strength per position. Its contraction analysis provides a checkable condition for convergence to a unique mean-field fixed point.

The controlled ablation demonstrates why both interaction signs matter. On HumanEval, attraction reduces NFE from 64.4 to 51.9. Adding repulsion increases the number of solved problems from 55 to 66, while requiring 56.3 NFE. These results show that attraction reduces decoding steps, while repulsion recovers accuracy and preserves the efficiency gain.

Thank you for bringing this closely related work to our attention. We will cite MFPD and add the direct comparison and these methodological distinctions to the revised manuscript.

[Q2c] Decision-direction similarity and co-commitment conflicts.

Thank you for suggesting this analysis. We examined 765,933 pairs from threshold-decoding traces, using all 164 HumanEval problems and 200 problems each from MBPP and GSM8K. For each pair (i,j)(i,j), we committed only ii and reran the model, using a change in jj’s top prediction as a conflict proxy. The second position was either co-committed or still masked with confidence of at least 0.5.

On held-out problems, a logistic predictor using confidence, margin, and distance achieves an AUROC of approximately 0.915. Adding decision-direction similarity increases this by approximately 0.0003. Among pairs actually committed together, predictions change in only 370 of 472,453 pairs (0.08%), and similarity adds no measurable predictive information. The incremental predictive value may appear small in this isolated setup.

However, the controlled ablation shows where repulsion helps: when attraction expands the commit set. Adding repulsion to attraction significantly improves accuracy from 55 to 66 on HumanEval and from 191 to 205 on MBPP, while retaining approximately 13% fewer forward passes than no coupling. On HumanEval, adjacent same-token co-commitments also decrease from 822 to 536.

We will clarify the distinction between the predictive analysis on threshold-decoding traces and the practical effect of repulsion during attraction-driven parallel decoding.


[W2/Q3] Transfer to a newer backbone with its native decoder.

Thank you for this valuable suggestion. We evaluated SDAR-8B-Chat-b32, a native block-diffusion model that reuses the keys and values of completed blocks. Both configurations use τ\tau=0.90 and differ only in the commitment rule.

We followed the authors’ evaluation code and settings: OpenCompass, fp16, greedy decoding, and KV caching enabled. RAPD uses the coupling weights selected on LLaDA without additional tuning for SDAR. MBPP uses the sanitized 257-problem set in the released evaluation code.

BenchmarkSolved: SDAR default / RAPDNFE: SDAR default / RAPDTime per problem: SDAR default / RAPD
HumanEval (164)120 / 12036.3 / 31.81.82 / 1.66 s
MBPP (257)152 / 14635.5 / 30.61.75 / 1.60 s
GSM8K (1,319)1,169 / 1,16269.3 / 59.83.45 / 3.17 s
MATH (5,000)3,382 / 3,262165.0 / 144.88.18 / 7.60 s

RAPD reduces NFE by 12.2–13.8% and per-problem runtime by 7–9%. These results demonstrate that the forward-pass and wall-clock gains transfer to a native block-diffusion backbone.


[W2] Longer generation lengths and larger blocks.

We thank the reviewer for their careful reading. The feedback helped us improve the decoding settings for different generation shapes.

Increasing generation length also substantially increases the computational budget. On HumanEval, extending the length from 256 to 512 gives the three baselines 4.9–6.1 percentage points of additional accuracy but requires 2.03–2.11× as many forward passes. We therefore investigated whether RAPD can recover accuracy at this longer length while retaining forward-pass savings.

We tested the configuration discovered in Q1.

MethodHumanEval (164)GSM8K (1,319)
RAPD (paper settings)66 / 117.21,027 / 70.0
RAPD (dev-selected threshold and cap)77 / 1391,025 / 81
Fast-dLLM74 / 164.91,031 / 95.1
LocalLeap75 / 124.31,026 / 72.8
DAWN72 / 130.01,009 / 70.4

The HumanEval score increases significantly from 66 to 77. The new setting also requires approximately 15–16% fewer forward passes than Fast-dLLM.

At block sizes 64 and 128, the same development-set procedure selects τ=0.90\tau=0.90 and a step cap equal to the block size. On HumanEval and GSM8K, these settings also show no significant accuracy difference from any baseline and retain 17–20% fewer forward passes than Fast-dLLM.

These results indicate that the commitment threshold and per-block step cap should be selected with the generation length and block size in mind. The threshold controls commitment aggressiveness, while the cap determines the decoding budget within each block. We will clarify this distinction in the revision.

[Q4] Relation to revision-based decoding and dependency-aware selectors.

We thank the reviewer for suggesting these comparisons, which help clarify RAPD’s relationship to existing decoding methods.

WINO [2] and ReMix [3]. WINO verifies and revokes earlier discrete drafts, whereas ReMix refines continuous input embeddings before discretization. RAPD instead refines commitment probabilities using fixed model outputs to determine which positions to finalize together. These approaches address complementary aspects of decoding.

LocalLeap [5]. LocalLeap promotes commitment around high-confidence anchors by relaxing the threshold for neighboring positions, but includes no repulsive interaction between candidate positions. RAPD combines locality-based attraction with explicit repulsion to control conflicting co-commitments within the same energy formulation.

DAWN [4]. DAWN uses attention-derived dependency graphs, incurring costs for attention-map processing, graph construction, and greedy scheduling. RAPD uses locality and decision-direction similarity within a signed energy, jointly promoting and suppressing commitment without attention-map extraction or dependency-graph construction.

MFPD [1]. We agree that MFPD is the closest empirical comparator, as both methods formulate binary commitment decisions through a pairwise energy and apply iterative refinement. MFPD uses inhibitory interactions, whereas RAPD jointly models attraction and repulsion. RAPD also bounds the total coupling per position and provides a sufficient condition for convergence to a unique mean-field fixed point. The direct comparison is presented in Q2a.

We thank the reviewer again for these suggestions. We will cite all five works and incorporate this discussion, together with the direct MFPD comparison, into the revised manuscript.

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

RAPD is a training-free method for accelerating diffusion language models by deciding which token positions to finalize together. It combines individual token confidence with attraction between nearby positions and repulsion between positions making similar token decisions. These signals define an energy function, and a few inexpensive refinement rounds determine the final commitment set without additional model forward passes. The central idea is that attraction enables more aggressive parallel decoding, while repulsion helps prevent the resulting prediction conflicts. Experiments on LLaDA-8B-Instruct and Dream-7B-Instruct show fewer decoding steps and higher throughput, with some accuracy trade-offs.

Critical strengths

Meaningful measured efficiency gains across two backbones. On LLaDA HumanEval, RAPD matches the reported Fast-dLLM parallel baseline’s 40.24% accuracy with 27.8% fewer forward passes, increasing throughput from 50.05 to 72.06 tokens/second. Applying the same interaction parameters to Dream provides evidence of transfer beyond one model.

Critical weaknesses

The evaluation does not establish efficiency in a caching-enabled inference system. The Fast-dLLM baseline is labeled “+Parallel,” its KV-cache configuration is unspecified, and the paper reports no RAPD experiments combined with approximate KV caching or on a native block diffusion backbone. Consequently, the reported speedups do not establish competitiveness with the complete Fast-dLLM pipeline. Evaluating all commitment policies with identical caching is particularly important because cheaper model passes could make RAPD’s selection overhead more consequential. To achieve inference speed, kv-cache approximation is necessary, but kv-cache approximation sometimes significantly changes the effectiveness of the decoding algorithms. In my experience, kv-cache approximation sometimes is incompatible with some decoding algorithms.

Questions for authors

Do the reported experiments enable approximate KV caching, such as Fast-dLLM’s KV cache or DualCache, for RAPD and the baselines? The tables label Fast-dLLM as “+Parallel,” but the cache configuration is unspecified. If caching is disabled, please compare RAPD and the baseline decoding policies under the same approximate KV-cache implementation, reporting end-to-end throughput and accuracy.

3 comments

We thank the reviewer for the careful review and the valuable comments.


[W1, Q1] Notation of the Fast-dLLM baseline and the cache setting.

Thank you for pointing out this ambiguity. We would like to clarify that all methods in the submission, including RAPD, were evaluated without a KV cache.

We followed the notation of the Fast-dLLM paper, which uses "+Parallel" for its parallel decoding without a KV cache and "+Cache+Parallel (Fast-dLLM)" for its full method with the cache. Accordingly, the row "+ Parallel (Fast-dLLM)" in our tables corresponds to the former (threshold 0.9, without a cache). In the revised version, we will explicitly state the cache setting in every table and rename this row "Fast-dLLM (parallel, no cache)".


[W2, Q1] Comparison under identical KV caching.

Thank you for raising this important point. To address this concern, we conducted additional experiments in which RAPD and Fast-dLLM (FD) were evaluated with identical PrefixCache and DualCache on LLaDA-8B-Instruct, using the published threshold of 0.9 for Fast-dLLM. "Solved" denotes the number of correctly answered problems (all 5000 problems for MATH), and TPS is measured end to end on a single RTX 4090 GPU.

TaskCacheSolved: RAPD 0.85 / RAPD 0.90 / FDNFE: RAPD 0.85 / RAPD 0.90 / FDTPS: RAPD 0.85 / RAPD 0.90 / FD
HumanEval (164)prefix64 / 70 / 6959.4 / 70.5 / 84.5102.1 / 88.9 / 76.0
dual56 / 58 / 5863.6 / 77.9 / 90.773.8 / 61.5 / 55.1
MBPP (500)prefix188 / 203 / 21150.3 / 59.9 / 69.664.2 / 56.9 / 50.4
dual174 / 189 / 19351.0 / 59.3 / 69.351.4 / 45.0 / 40.3
GSM8K (1319)prefix1011 / 1034 / 102557.9 / 68.2 / 81.270.8 / 63.9 / 54.7
dual1013 / 1035 / 104261.8 / 72.2 / 85.358.6 / 52.5 / 46.1
MATH (5000)prefix1551 / 1615 / 164971.6 / 86.6 / 100.876.2 / 65.4 / 57.4
dual1518 / 1615 / 163576.0 / 92.8 / 107.159.6 / 50.3 / 45.0

At τ = 0.90, RAPD requires 13–17% fewer forward passes and achieves 12–17% higher end-to-end throughput than Fast-dLLM, with no significant accuracy difference in any of the eight settings.

This comparison directly evaluates RAPD against the complete Fast-dLLM pipeline, with both approximate KV caching and parallel decoding enabled under the same cache implementation.


[W3] Evaluation on a native block-diffusion backbone.

Thank you for the valuable suggestion. Following the reviewer's suggestion, we evaluated SDAR-8B-Chat-b32, a block-diffusion model that reuses the keys and values of completed blocks, under the evaluation settings released by its authors. The two configurations differ only in the commit rule, and RAPD uses the coupling weights selected on LLaDA without additional tuning.

TaskSolved: SDAR default / RAPDNFE: SDAR default / RAPDTime per problem: SDAR default / RAPD
HumanEval (164)120 / 12036.3 / 31.81.82 / 1.66 s
MBPP (sanitized, 257)152 / 14635.5 / 30.61.75 / 1.60 s
GSM8K (1319)1169 / 116269.3 / 59.83.45 / 3.17 s
MATH (5000)3382 / 3262165.0 / 144.88.18 / 7.60 s

Without additional tuning of the coupling parameters, RAPD reduces NFE by 12.2–13.8% and per-problem runtime by 7–9% across the four benchmarks. These results demonstrate that RAPD’s efficiency gains transfer to a native block-diffusion backbone with KV caching enabled.

[W4] Selection overhead under reduced forward-pass cost.

Thank you for this insightful comment. To quantify this effect, we separately measured the model invocation time and the remaining time of the decoding loop, including RAPD’s selection, at τ = 0.85 on HumanEval and 200 GSM8K problems. The remaining time is 3.1–3.4 ms per step regardless of caching. With PrefixCache, the model invocation time decreases to 36–55 ms per step, which increases the share of the remaining time from 2–5% without the cache to 5.5–8%, and RAPD requires an additional 0.4–1.5 ms per forward pass compared with Fast-dLLM.

RAPD remains faster with caching despite its selection overhead. Since RAPD requires fewer decoding steps, it remains faster under caching; the per-problem runtime of Fast-dLLM is 1.26–1.35× that of RAPD at τ = 0.85 and 1.12–1.18× at τ = 0.90.


[W5] Effect of approximate caching on each decoding method.

We appreciate the reviewer sharing this insight from their experience. To examine this effect, we compared each method with and without the cache on a per-problem basis: RAPD at τ = 0.85 (the setting in the paper) and τ = 0.90, and Fast-dLLM (FD) at τ = 0.9. Each cell reports the number of problems solved without → with the cache.

TaskCacheRAPD τ = 0.85RAPD τ = 0.90FD
HumanEval (164)prefix66 → 6466 → 7066 → 69
dual66 → 5666 → 5866 → 58
MBPP (500)prefix205 → 188214 → 203211 → 211
dual205 → 174214 → 189211 → 193
GSM8K (1319)prefix1026 → 10111037 → 10341034 → 1025
dual1026 → 10131037 → 10351034 → 1042
MATH (5000)prefix1578 → 15511623 → 16151649 → 1649
dual1578 → 15181623 → 16151649 → 1635

Among problems that both methods solve without caching, 74–100% of RAPD’s cache-induced failures are also cache-induced failures of Fast-dLLM. These results indicate that the observed cache sensitivity is not unique to RAPD.

We thank the reviewer again for the constructive comments and suggestions, which have helped us strengthen the paper. We would be happy to answer any further questions.

Dear Reviewer bdmU,

Thank you again for your time and constructive feedback.

Your suggestions have substantially improved our paper. In response, we made several additions, most notably comparing RAPD and Fast-dLLM under identical KV caching and evaluating RAPD on a native block-diffusion model. The revised manuscript has been uploaded, with all changes highlighted in red.

As the discussion period ends soon, we would sincerely appreciate your feedback on whether our responses have addressed your concerns. We would be glad to answer any remaining questions.

Best regards,

The Authors

Comments on the paper

1 comment

We thank all reviewers for their constructive feedback. We summarize our responses to the shared concerns below; all revisions are highlighted in red in the revised manuscript.

Inference with KV caching. Under identical PrefixCache and DualCache, RAPD at τ = 0.90 requires 13–17% fewer forward passes and achieves 12–17% higher end-to-end throughput than the full Fast-dLLM pipeline, with no significant accuracy difference (Section 4.2 and Appendix J). On the native block-diffusion model SDAR-8B-Chat, RAPD reduces NFE by 12.2–13.8% without retuning (Appendix K).

Configuration selection on development data. A configuration reselected on held-out MBPP development data requires 16–30% fewer forward passes than Fast-dLLM, with no significant accuracy difference (Appendix I), and the same procedure recovers accuracy at longer generations (Appendix H).

Relation to MFPD and the role of repulsion. We added a discussion of MFPD and related methods (Section 2.3) and a direct comparison, in which RAPD requires 7–14% fewer forward passes than MFPD (Appendix L). Controlled ablations show that repulsion recovers the accuracy lost under attraction-only decoding (Section 5.1 and Appendices G and M).

We hope these revisions address the reviewers’ concerns and would be happy to discuss any remaining questions.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.