acceptodds
Under review as a conference paper at ICLR 2027

Learning When to Commit: Schedule Distillation for Fast Diffusion Language Models

Abstract

Diffusion large language models (dLLMs) can generate multiple tokens in parallel, but often require many decoding rounds to complete a response. Efficient generation depends not only on which tokens are easy to predict, but also on how committing to them helps predict the remaining positions. A highly predictable token may provide little useful context, while revealing a less predictable token can help several other positions become predictable sooner. We propose \bf \em dSpark, a new schedule distillation-based framework for learning when to commit tokens for faster diffusion language generation. To achieve this, we construct unmasking schedules that prioritize tokens whose reveal makes other positions easier to predict, reducing uncertainty earlier in decoding. We then distill these schedules into the model to enable faster parallel sampling, with a KL-divergence regularizer to help preserve generation quality. We extensively evaluate our approach across mathematical reasoning and code generation tasks. On GSM8K, our method achieves remarkable \bf 10.12 and \bf 5.12 speedups over the various baselines on LLaDA-8B-Instruct and Dream-7B-Instruct, respectively, while maintaining superior task performance.

Primary area:
generative models
Keywords:
Diffusion LLMs, Distillation, Parallel decoding, Fast sampling

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 2 reviews, above 27% of the others.
Reviewer_t3PkOpenReview ↗
Rating
1(−1)
Confidence
4(+1)
Informativeness
1
Presentation
1
Soundness
2
Read review
Paper summary

The paper proposes dSpark, a framework for accelerating diffusion language models through offline schedule construction and subsequent fine-tuning. Given a correct reference response, a frozen base model constructs a reveal schedule by promoting plausible reference tokens whose revelation reduces the surprisal of other masked reference tokens. The student is then trained on intermediate states from these schedules using masked-token cross-entropy and confidence sharpening, together with selective KL regularization on failed trajectories. At inference, the fine-tuned model uses confidence-based parallel decoding without the offline search. Experiments on LLaDA and Dream across mathematical reasoning and code generation benchmarks report improved accuracy–parallelism trade-offs, with ablations examining the promotion strategy and regularization.

Critical strengths

The paper provides empirical evidence that its training recipe improves decoding efficiency. In particular, Table 11 compares the proposed promotion criterion with natural scheduling, random plausible promotion, and highest-surprisal promotion. With KL regularization, the proposed criterion achieves 10.1 tokens per forward at 78.9% accuracy on LLaDA GSM8K, compared with 7.5 tokens per forward at 78.7% accuracy for random plausible promotion. This supports the practical value of the proposed training-state construction beyond simply promoting arbitrary plausible tokens. I regard this controlled empirical evidence as the paper’s main strength.

Critical weaknesses
  1. The connection between finding a better reveal order and learning a better commitment strategy is insufficiently established. Section 4.2 constructs schedules for a fixed reference response using a frozen teacher. However, Section 4.3 trains the student to predict all eligible masked reference tokens from the resulting intermediate states, with additional confidence sharpening. The objective does not explicitly supervise which positions should be committed next, and I did not find a direct analysis showing that the student learns the useful commitment behavior identified by the offline search. Table 11 demonstrates that the proposed training-state construction improves final performance, but it does not establish whether this improvement comes from better commitment decisions, more effective supervised contexts, or changes in confidence. Furthermore, fine-tuning changes the model’s conditional predictions, and its generated responses may also change, leaving the relevance of the teacher’s optimized schedules to the student’s actual generation insufficiently explained. Exact reproduction of the original order is unnecessary, but the paper needs clearer evidence connecting its scheduling principle to the student’s inference behavior. This gap weakens the justification of the central “schedule distillation” claim.

  2. The conceptual contribution is limited relative to the number of components and the additional trajectory-construction effort. The method combines candidate filtering, greedy promotion search, training on selected trajectory states, confidence sharpening, and selective KL regularization. Although the individual components have understandable purposes, I find their combination insufficiently supported by a clear, unified explanation. The existing ablations establish the effectiveness of the assembled recipe more convincingly than they establish a distinct new insight. In particular, the paper does not sufficiently justify whether the additional computation spent constructing optimized trajectories teaches useful behavior that simpler training could not obtain at a comparable total cost. The offline construction cost is also insufficiently characterized. Overall, I view the contribution as a carefully assembled empirical recipe with limited conceptual advance, which substantially reduces its significance in my assessment.

Questions for authors
  1. Clarification of the claimed learning mechanism. What evidence shows that training on the optimized trajectory states causes the student to make more useful commitment decisions during its own generation? In particular, do the student’s actual early commitments have greater downstream utility than those of students trained with natural or random promotion schedules? Please also clarify how the interpretation of “schedule distillation” accounts for changes in the student’s conditional predictions and generated responses after fine-tuning.

  2. Clarification of the ablation controls. In Table 11, how are states sampled from trajectories of different lengths, and are the comparisons matched for training updates, sampled-state counts, and total supervised tokens? Is there an analysis separating the effect of optimized-state construction from confidence sharpening? These details would help determine whether the results support learning a better commitment strategy or a broader benefit from changing the distribution and weighting of supervised examples.

6 comments
AuthorsOpenReview ↗

Response to Reviewer t3Pk [1/4]

Thank you for your constructive and valuable comments, which are definitely helpful for us to improve the quality of the paper. We have uploaded a revised version of the manuscript to accommodate all your mentioned suggestions. Below, we further address each of your concerns as follows:

W1.1: The connection between finding a better reveal order and learning a better commitment strategy is insufficiently established. Section 4.2 constructs schedules for a fixed reference response using a frozen teacher. However, Section 4.3 trains the student to predict all eligible masked reference tokens from the resulting intermediate states, with additional confidence sharpening. The objective does not explicitly supervise which positions should be committed next, and I did not find a direct analysis showing that the student learns the useful commitment behavior identified by the offline search. Table 11 demonstrates that the proposed training-state construction improves final performance, but it does not establish whether this improvement comes from better commitment decisions, more effective supervised contexts, or changes in confidence.

Q1.1: Clarification of the claimed learning mechanism. What evidence shows that training on the optimized trajectory states causes the student to make more useful commitment decisions during its own generation? In particular, do the student’s actual early commitments have greater downstream utility than those of students trained with natural or random promotion schedules?

Thanks for the comment. We clarify that the decoder commits the tokens for which the model has the highest confidence, so the model's confidence under the current context determines each commitment decision. Our training states shape exactly this confidence. A promoted token is a prediction target only in the early states, before the schedule commits it, so the student learns to predict it confidently at that early point. To verify the effect directly, we compared dSpark with a control student trained with the same losses (sharpening included) but on natural schedules. Both students generate freely on the GSM8K test set, and the frozen base model evaluates every token they commit.

Table 1. Commitment of high-utility tokens during free generation on the GSM8K test set. Both students are trained with the same losses, sharpening included, and differ only in the training schedule. ↑ means higher is better.

StudentEarly commits ↑Utility of early commits ↑Schedule gain recovered ↑
Natural-schedule33.7%1.12$\times$11.9%
dSpark56.3%1.41$\times$36.6%

Early commits are tokens that a student commits while the base model would still keep them masked. Utility measures how much these tokens help predict the other masked positions (Eq. 6), relative to tokens left masked, so 1$\times$ means no difference. Schedule gain recovered measures how far the decoding order moves from the base model's order (0%) toward the offline schedule (100%).

dSpark makes clearly better commitment decisions than the natural-schedule student. It commits more tokens early, the tokens it commits early are more useful to the remaining positions, and its decoding order moves about three times as far toward the offline schedule. The two students share the same losses, sharpening included, so this improvement comes from the optimized schedule and not from higher confidence. The student therefore learns the commitment behavior identified by the offline search and applies it during its own generation. This analysis is now in Section 5.3 of the revised paper (Table 4).

AuthorsOpenReview ↗

Response to Reviewer t3Pk [2/4]

W1.2: Furthermore, fine-tuning changes the model’s conditional predictions, and its generated responses may also change, leaving the relevance of the teacher’s optimized schedules to the student’s actual generation insufficiently explained. Exact reproduction of the original order is unnecessary, but the paper needs clearer evidence connecting its scheduling principle to the student’s inference behavior. This gap weakens the justification of the central “schedule distillation” claim.

Q1.2: Please also clarify how the interpretation of “schedule distillation” accounts for changes in the student’s conditional predictions and generated responses after fine-tuning.

Thanks for raising this point. We clarify that the analysis in Table 1 was conducted in exactly this setting. The fine-tuned student decodes its own responses to unseen test problems without access to any schedule, and the offline schedule is computed on those same responses. If the teacher's schedules were no longer relevant after fine-tuning, dSpark should behave like the control student trained on natural schedules. Instead, its early commitments are more helpful to the remaining positions, and its decoding order moves about three times as far toward the offline schedule.

We believe the connection is preserved for two reasons. First, the LoRA adaptation and the KL term keep the student close to the teacher. Second, the usefulness of a promoted token is largely determined by the content of the response. For example, an intermediate result that later steps depend on stays useful to commit early, and fine-tuning is not expected to alter such dependencies. Schedule distillation therefore transfers the scheduling principle, which is to commit high-utility tokens early, and does not require the student to reproduce the teacher's exact order. The revised Section 5.3 now states this.

W2.1: The conceptual contribution is limited relative to the number of components and the additional trajectory-construction effort. The method combines candidate filtering, greedy promotion search, training on selected trajectory states, confidence sharpening, and selective KL regularization. Although the individual components have understandable purposes, I find their combination insufficiently supported by a clear, unified explanation. The existing ablations establish the effectiveness of the assembled recipe more convincingly than they establish a distinct new insight.

Thanks for the comments. We clarify that dSpark follows from a single observation. Decoding speed depends on the order in which tokens are revealed, and the most confident token is often not the most useful one to reveal (Fig. 1). We formalize this with the uncertainty area (Eq. 4), and each component serves this one objective. The promotion rule (Eq. 6) is a greedy step that reduces the area, and the candidate filter restricts it to tokens the model itself finds plausible. Training on the resulting states transfers the order to the student. Sharpening is needed because the decoder commits by confidence, so a learned order can only be expressed through confidence, and the KL term confines this confidence gain to correct predictions. The method therefore consists of one schedule and two loss terms on top of cross-entropy.

The ablations also isolate the insight from the recipe. In Table 11 of the submission (now Table 15), the data, losses and training budget are fixed and only the reveal order changes. TPF increases step by step from the natural schedule to random promotion and to highest-surprisal promotion, and it is highest with our criterion. The gain therefore comes from the utility principle itself, and not from early reveals in general or from the other components.

AuthorsOpenReview ↗

Response to Reviewer t3Pk [3/4]

W2.2: In particular, the paper does not sufficiently justify whether the additional computation spent constructing optimized trajectories teaches useful behavior that simpler training could not obtain at a comparable total cost. The offline construction cost is also insufficiently characterized. Overall, I view the contribution as a carefully assembled empirical recipe with limited conceptual advance, which substantially reduces its significance in my assessment.

Thanks for this comment. Following your suggestion, we have added a cost breakdown and an equal-cost comparison to Appendix A.7 of the revised paper (Table 11). On LLaDA, generating the responses takes 200 A100 GPU-hours, a cost shared by any method that distills from model trajectories. Constructing the optimized schedules adds 13 GPU-hours, about 6% of the total. It requires 20.4 teacher forward passes per full response on average, since all candidates of a round (4.3 on average) are scored in one batched pass. Training takes 10 GPU-hours for 6 epochs, and inference incurs no additional cost. To compare at equal cost, we gave the construction budget to the simpler alternative and trained the natural-schedule student for 14 epochs, which brings it to the same total cost as dSpark (Table 2).

Table 2. Equal-cost comparison on GSM8K with LLaDA. Both students spend 23 A100 GPU-hours after response generation.

StudentCost after response generationTPFAcc. (%)
Natural schedule, 14 epochs23 GPU-hours8.877.01
dSpark, construction and 6 epochs23 GPU-hours10.1278.94

At the same total cost, the natural-schedule student remains behind dSpark in both parallelism and accuracy. The optimized schedules therefore teach a behavior that additional training on natural schedules does not provide. This is consistent with Table 1, where dSpark moves much further toward the offline schedule than the natural-schedule student.

Q2.1: Clarification of the ablation controls. In Table 11, how are states sampled from trajectories of different lengths, and are the comparisons matched for training updates, sampled-state counts, and total supervised tokens?

Thanks for the question. Yes, the comparisons are matched. Training updates and sampled-state counts are identical across rows, and supervised tokens agree to within 0.1 token per state. All rows of Table 11 of the submission (now Table 15) use the same sampling procedure. Each response contributes exactly one state per epoch, regardless of the length of its trajectory. We sample a block uniformly and then a round uniformly within that block, and the state is redrawn at every epoch. Longer trajectories therefore receive neither more states nor more weight. All rows also use the same 52k correct responses (and, with KL, the same 13k failed responses), the same 6 epochs and the same batch size of 64, so the number of training updates and the number of sampled states are identical. The number of supervised tokens per state is determined by the schedule and in practice it is nearly the same. On average a state from a correct response supervises 49.2 tokens under the natural schedule and 49.3 under ours, and a state from a failed response supervises the same 127.8 tokens in every row. The loss is averaged over the supervised tokens of each batch, so every update carries the same weight in all rows.

Reviewer_e8irOpenReview ↗
Rating
3
Confidence
4
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes dSpark, a schedule-distillation framework for accelerating masked diffusion language models. Using verified correct responses, the method constructs offline unmasking trajectories by promoting tokens whose revelation reduces reference-token surprisal at other masked positions. The model is then trained on these trajectories using cross-entropy and confidence sharpening, with KL regularization on failed trajectories. Experiments on LLaDA-8B-Instruct and Dream-7B-Instruct demonstrate improved accuracy–parallelism trade-offs on mathematical reasoning and code generation.

Critical strengths

The distinction between token predictability and downstream utility is meaningful. A token that is easy to predict need not provide useful context for other positions. The paper investigates this distinction through reference-guided interventions and schedule construction. The experimental evaluation includes useful comparisons and ablations.

Critical weaknesses
  1. The connection between the optimized offline schedules and the student’s actual decoding order is insufficiently demonstrated. Figures 1–2 analyze interventions that reveal known reference tokens and schedules constructed with access to a correct response. These results support the usefulness of the offline schedules, but do not establish that the trained student autonomously commits high-utility tokens earlier during generation. Since inference still uses confidence-based commitment, the observed gains could reflect confidence sharpening or improved prediction on selected training states, rather than acquisition of the intended scheduling behavior. Direct analysis of commitment order before and after training would strengthen the central claim.
  2. The source of the parallelism improvement needs clearer attribution. Appendix C states that TPF excludes end-of-sequence padding, which addresses direct inflation of the token count. However, this does not establish how much of the reduction in forward passes comes from faster generation of substantive response tokens versus padding completion or termination behavior. Although the paper reports wall-clock gains on LLaDA, a breakdown of effective response lengths and decoding work before and after substantive completion would better connect the acceleration to the proposed semantic scheduling mechanism.
  3. Applicability to models trained with block-diffusion objectives remains unverified. Using a block size of 32 and multi-block decoding on LLaDA/Dream does not establish effectiveness on backbones trained with block diffusion, such as LLaDA2.1-mini [1] or DreamReasoner [2]. This distinction matters because the proposed utility criterion depends on how revealing one position affects other positions, and that interaction is constrained differently under block-causal dependencies. A controlled evaluation on at least one such backbone would clarify the method’s scope.
  4. The discussion omits closely related progressive-unmasking training work. PUMA [3] and Trims[4] also constructs training examples by progressively revealing ground-truth tokens and training on intermediate masked states. The paper should clarify its contribution relative to these work.

Reference

[1] LLaDA2.1: Speeding Up Text Diffusion via Token Editing. https://arxiv.org/abs/2602.08676

[2] DreamReasoner-8B: Block-Size Curriculum Learning for Diffusion Reasoning Models. https://arxiv.org/abs/2606.19257

[3] Stop Training for the Worst: Progressive Unmasking Accelerates Masked Diffusion Training. https://arxiv.org/abs/2602.10314

[4] TRIMS: Trajectory-Ranked Instruction Masked Supervision for Diffusion Language Models. http://arxiv.org/abs/2604.00666

Questions for authors
  1. Can the authors measure whether high-utility positions are committed earlier and correctly by the distilled student during free generation, compared with the original model?
  2. Please clarify whether multi-block decoding is reflected in the training trajectories. It appears that multi-block decoding may be introduced only at inference, while distillation uses trajectories constructed under a different unmasking scheme. If so, this could create a mismatch between the training states and the partially decoded multi-block contexts encountered at inference. Are trajectory construction and inference governed by the same block-activation and commitment rules? If not, how is this mismatch addressed? Does training retain the backbones’ bidirectional attention, or introduce block-causal attention?
4 comments
AuthorsOpenReview ↗

Response to Reviewer e8ir [1/4]

Thank you for the thoughtful and constructive comments, which are definitely helpful for us to improve the quality of the paper. We have uploaded a revised version of our paper incorporating all concrete revisions following your suggestions. Below, we further address each of your comments.

W1: The connection between the optimized offline schedules and the student’s actual decoding order is insufficiently demonstrated. Figures 1–2 analyze interventions that reveal known reference tokens and schedules constructed with access to a correct response. These results support the usefulness of the offline schedules, but do not establish that the trained student autonomously commits high-utility tokens earlier during generation. Since inference still uses confidence-based commitment, the observed gains could reflect confidence sharpening or improved prediction on selected training states, rather than acquisition of the intended scheduling behavior. Direct analysis of commitment order before and after training would strengthen the central claim.

Q1: Can the authors measure whether high-utility positions are committed earlier and correctly by the distilled student during free generation, compared with the original model?

Thanks for raising this point. Following your suggestion, we ran this analysis during free generation and added it to Section 5.3 of the revised paper (Table 4). We compare dSpark with a natural-schedule student trained with the same losses, so the two differ only in the training schedule. Both decode the GSM8K test set normally, without any reference. For each committed token, the frozen base model tells us whether it would have committed that token at that point, and how useful the token is to the remaining positions (Eq. 6). The base model thus serves as the reference before training.

Table 1. Commitment of high-utility tokens during free generation on the GSM8K test set. Both students are trained with the same losses and differ only in the training schedule. ↑ means higher is better.

StudentEarly commits ↑Utility of early commits ↑Schedule gain recovered ↑
Natural-schedule33.7%1.12$\times$11.9%
dSpark56.3%1.41$\times$36.6%

Early commits are tokens that the student commits while the base model would have kept them masked at that step. Utility is how much these tokens help predict the other masked positions, compared with the tokens the student chose not to commit. A value of 1$\times$ means no difference, and a higher value means the student commits the more helpful tokens first. The last column shows how close each model's decoding order gets to the offline schedule, where 0% is the base model's order and 100% is the offline schedule itself.

The natural-schedule student already commits some tokens early, which we attribute to sharpening. dSpark goes clearly beyond this on all three measures. It commits early more often, picks more useful tokens when it does, and moves about three times as far toward the offline schedule. Since the losses are identical, this gap is due to the schedule and not to confidence sharpening.

AuthorsOpenReview ↗

Response to Reviewer e8ir [2/4]

W2: The source of the parallelism improvement needs clearer attribution. Appendix C states that TPF excludes end-of-sequence padding, which addresses direct inflation of the token count. However, this does not establish how much of the reduction in forward passes comes from faster generation of substantive response tokens versus padding completion or termination behavior. Although the paper reports wall-clock gains on LLaDA, a breakdown of effective response lengths and decoding work before and after substantive completion would better connect the acceleration to the proposed semantic scheduling mechanism.

Thanks for the helpful suggestion. We clarify that the gain comes from the response tokens themselves, not from padding or termination. To show this directly, we ran the suggested breakdown on GSM8K for both backbones (Tables 2 and 3). For each model, we report the response length without padding, the forward passes used until the last response token is committed, and the passes used after that.

Table 2. Decoding work before and after the response is complete, on GSM8K with LLaDA-8B-Instruct. Response tokens exclude padding. TPF (response only) is response tokens divided by passes before completion.

ModelResponse tokensPasses before completionPasses after completionTPF (response only)
LLaDA247.9247.550.011.00
dParallel212.429.010.947.32
d3LLM244.926.990.069.07
dSpark249.624.660.0010.12

Table 3. The same breakdown on GSM8K with Dream-7B-Instruct.

ModelResponse tokensPasses before completionPasses after completionTPF (response only)
Dream101.1101.080.011.00
dParallel110.536.090.013.06
d3LLM114.222.980.004.97
dSpark103.720.230.035.13

dSpark spends almost no forward passes after the response is complete, and neither do the base models. It produces responses of about the same length as the base model in about 10$\times$ fewer passes on LLaDA and 5$\times$ fewer on Dream, and it needs the fewest passes of all methods on both backbones. Dream is the clearer case, since its responses fill only about 40% of the 256-token budget, yet finishing costs almost no extra passes.

AuthorsOpenReview ↗

Response to Reviewer e8ir [3/4]

W3: Applicability to models trained with block-diffusion objectives remains unverified. Using a block size of 32 and multi-block decoding on LLaDA/Dream does not establish effectiveness on backbones trained with block diffusion, such as LLaDA2.1-mini [1] or DreamReasoner [2]. This distinction matters because the proposed utility criterion depends on how revealing one position affects other positions, and that interaction is constrained differently under block-causal dependencies. A controlled evaluation on at least one such backbone would clarify the method’s scope.

Thanks for this suggestion. We clarify that dSpark is effective on backbones trained with block diffusion. We evaluated it on both models the reviewer mentions, LLaDA2.1-mini [1] and DreamReasoner-8B [2], and it raises TPF in all 12 settings (Table 4). These results are now in Section 5.3 of the revised paper (Table 5), with the setup in Appendix A.5.

We kept each backbone's block-causal attention when constructing the schedules, during training, and at decoding, so the utility criterion (Eq. 6) is computed under the same dependencies the model uses at inference. We trained with the same recipe on 29,488 verified responses for DreamReasoner-8B and 30,210 for LLaDA2.1-mini. The base model and dSpark use the same decoder, which is each backbone's own confidence-threshold decoder with block size 32 and up to 1,024 tokens. For LLaDA2.1-mini we report both of its settings, quality mode (threshold 0.7, editing threshold 0.5) and speed mode (threshold 0.5, no editing).

Table 4. dSpark on backbones trained with block diffusion. Each cell shows accuracy (%) / TPF. Each "+ dSpark" row uses the same decoder as the base row above it and shows, in parentheses, its TPF gain over that row.

ModelGSM8KMATH-500HumanEvalMBPP
DreamReasoner-8B88.4 / 2.0072.4 / 1.9373.8 / 1.5260.4 / 1.68
+ dSpark88.4 / 3.65 (+83%)71.8 / 2.92 (+51%)76.8 / 2.89 (+90%)59.6 / 2.51 (+49%)
LLaDA2.1-mini, quality mode92.8 / 4.2563.0 / 4.9084.8 / 7.5367.8 / 4.99
+ dSpark91.9 / 5.84 (+37%)64.0 / 5.90 (+20%)83.9 / 8.94 (+19%)65.0 / 6.07 (+22%)
LLaDA2.1-mini, speed mode91.9 / 4.9963.2 / 5.8884.1 / 8.3065.8 / 5.92
+ dSpark91.8 / 7.05 (+41%)63.8 / 7.46 (+27%)82.7 / 10.51 (+27%)66.4 / 7.46 (+26%)

The TPF gain is consistent across backbones, benchmarks and decoding modes, at 68% on average on DreamReasoner-8B and 27% on LLaDA2.1-mini. Accuracy stays close to the base model in most settings and improves in several. On LLaDA2.1-mini, dSpark in quality mode is even faster than the base model in its own speed mode on all four benchmarks at similar accuracy. These gains come on top of backbones that already decode several tokens per pass. Schedule distillation is therefore complementary to block-diffusion training and to token editing.

W4: The discussion omits closely related progressive-unmasking training work. PUMA [3] and Trims[4] also constructs training examples by progressively revealing ground-truth tokens and training on intermediate masked states. The paper should clarify its contribution relative to these work.

Thanks for pointing this out. Following your suggestion, we have added both to Related Work (Section 2). PUMA [3], TRIMS [4] and dSpark all train on structured masked states instead of random masks. They differ in which order defines those states, and this order is our contribution.

PUMA aims at faster training. It trains on the masking patterns the model meets at inference, by revealing ground-truth tokens in the order the model itself would unmask them, and it keeps this order unchanged. dSpark changes the order, promoting tokens that make the remaining positions easier to predict, so the model learns to finish in fewer steps.

TRIMS reveals hard tokens first, where difficulty is a token's own likelihood under an autoregressive teacher. dSpark ranks a token by how much revealing it helps the other positions, measured with the diffusion model itself. Table 11 of the submission (now Table 15) supports this choice. Under the same loss, our utility-based order gives the highest TPF, ahead of a hardest-first order (highest surprisal under the diffusion model) and well ahead of the natural order.

AuthorsOpenReview ↗

Response to Reviewer e8ir [4/4]

Q2.1: Please clarify whether multi-block decoding is reflected in the training trajectories. It appears that multi-block decoding may be introduced only at inference, while distillation uses trajectories constructed under a different unmasking scheme. If so, this could create a mismatch between the training states and the partially decoded multi-block contexts encountered at inference.

Thanks for the question. Yes, multi-block decoding is reflected in training, through the loss. Trajectories are constructed one block at a time, but during training, once the current block is partly decoded, the loss also covers the masked positions of the next block. The student therefore learns to predict two blocks in parallel, as in multi-block decoding at inference.

Q2.2: Are trajectory construction and inference governed by the same block-activation and commitment rules? If not, how is this mismatch addressed?

We clarify that construction and inference use the same block size and differ in two ways, both handled in training.

Block activation. During trajectory construction, each block is fully decoded before the next block is opened, whereas inference opens the next block while the current one is only partly decoded. To cover this, the training loss also includes the next block once the current block is partly decoded (see Q2.1), so the student is trained for exactly this case.

Commitment rule. Construction has access to the reference response and, in each round, commits the positions the frozen model already predicts correctly, plus one promoted token. Inference has no reference and instead commits tokens by confidence. Cross-entropy and sharpening raise confidence on correct tokens so that they meet the inference confidence criterion.

Q2.3: Does training retain the backbones’ bidirectional attention, or introduce block-causal attention?

We clarify that training retains each backbone's original attention, and we never change it. LLaDA and Dream are trained and decoded with full bidirectional attention, and we do not introduce block-causal attention for them at any stage. The block-diffusion backbones in W3 keep their native block-causal attention at every stage. We state this in Appendix A.5.


References

[1] LLaDA2.1: Speeding Up Text Diffusion via Token Editing. https://arxiv.org/abs/2602.08676

[2] DreamReasoner-8B: Block-Size Curriculum Learning for Diffusion Reasoning Models. https://arxiv.org/abs/2606.19257

[3] Stop Training for the Worst: Progressive Unmasking Accelerates Masked Diffusion Training. https://arxiv.org/abs/2602.10314

[4] TRIMS: Trajectory-Ranked Instruction Masked Supervision for Diffusion Language Models. https://arxiv.org/abs/2604.00666

Comments on the paper

3 comments
AuthorsOpenReview ↗

Response to Reviewer SJyz [1/3]

Thank you for your thoughtful and constructive comments, which are definitely helpful for us to improve the quality of the paper. We have uploaded a revised version of our paper that incorporates all concrete revisions based on your suggestions. Below, we further provide detailed responses to address each of your concerns:

W1: Offline schedule construction evaluates candidates with separate teacher forwards, but its total cost is not reported. This leaves the training/inference compute tradeoff and the amount of deployment needed to amortize construction unclear.

Q1: What are the teacher-forward count and GPU-hours for schedule construction, and how does that cost compare with the reported inference savings?

Thanks for the question. Following your suggestion, we have added the cost breakdown to Appendix A.7. Construction is a one-time offline step. All candidates of a round (4.3 on average) are scored in a single batched forward pass, so a response requires 20.4 teacher forward passes on average. For LLaDA, the pipeline costs 200 A100 GPU-hours for response generation, which any method that distills from model trajectories also requires, 13 GPU-hours for schedule construction and 10 GPU-hours for training. Construction therefore accounts for about 6% of the total. Spending the same 13 GPU-hours on longer natural-schedule training instead (14 epochs instead of 6) gives 8.8 TPF at 77.01% accuracy on GSM8K, below the 10.12 TPF and 78.94% of dSpark (Table 11 of the revised paper).

At inference dSpark adds no cost, since it uses no search and no additional module, and it reduces the forward passes per GSM8K response from 247.6 to 24.7. In wall-clock terms it generates 249 tokens per second against 15 for the base model, which saves about 15.5 seconds per response on one A100. The construction cost is thus recovered after roughly 3k generated responses, and the entire pipeline, including response generation and training, after roughly 52k.

AuthorsOpenReview ↗

Response to Reviewer SJyz [2/3]

W2: The contribution of scheduling is not fully separated from the decoder configuration. Appendix C gives a 0.90-0.95 threshold range and overlapping-block rules without a complete per-method configuration, making the fairness of operating-point comparisons difficult to assess.

Thanks for the comment. We clarify that every method in the main results (Table 1 of the paper) is evaluated with its own decoder and with the settings released by its authors, since each model is trained for its decoder. Table 1 below gives the complete configuration, which we have added to Appendix C (Table 13). All methods share the block size (32), the generation length (256), the prompts and the evaluation harness.

Table 1. Decoder and settings of each method in the main results. τ is a confidence threshold and η an entropy threshold.

MethodDecoderCommit ruleLLaDADream
Fast-dLLMBlock-wise decoding with KV cacheConfidence ≥ ττ = 0.9τ = 0.9
D2FPipelined block decoding, next block added at a% and fully active at 95%Confidence > ττ = 0.9, a = 70 (GSM8K), 10 (MATH, HumanEval), 90 (MBPP)τ = 0.9, a = 70 (GSM8K), 10 (MATH, HumanEval), 90 (MBPP)
dParallelBlock-wise decodingEntropy ≤ ηη = 0.5 (0.45 on MBPP)η = 0.45 (0.5 on HumanEval and MBPP)
d3LLMMulti-block decoding, next block added at 10% and fully active at 95% (after completion on HumanEval with LLaDA)Entropy ≤ ηη = 0.5 (0.4 on MBPP)η = 0.4 (0.5 on HumanEval)
dSparkMulti-block decoding, next block added at 50% and fully active at 50%Confidence ≥ ττ = 0.90 (GSM8K, MATH, MBPP), 0.95 (HumanEval)τ = 0.90 (GSM8K, MATH, MBPP), 0.95 (HumanEval)

To separate the method from the decoder, we also evaluated the distilled models under one shared decoder. Table 2 uses the entropy-based decoder and settings of d3LLM for every model, and Table 3 uses the decoder and settings of dSpark for every model. The row of the method that owns the decoder is taken from Table 1 of the paper.

Table 2. All distilled models with the decoder and settings of d3LLM. Each cell shows accuracy (%) / TPF, and the best value per backbone and benchmark is in bold.

BackboneModelGSM8KMATHMBPPHumanEval
LLaDAdParallel75.8 / 8.6030.8 / 4.9241.4 / 3.1740.2 / 5.28
LLaDAd3LLM74.2 / 9.0530.3 / 5.7539.7 / 4.5639.6 / 5.95
LLaDAdSpark78.8 / 9.5132.3 / 5.8941.4 / 4.9539.0 / 6.15
DreamdParallel82.9 / 4.0538.3 / 3.2156.2 / 2.6854.3 / 2.93
Dreamd3LLM80.7 / 4.9738.6 / 4.0755.2 / 2.8957.1 / 3.19
DreamdSpark83.5 / 5.2039.0 / 3.9556.2 / 3.3161.0 / 3.83

Table 3. All distilled models with the decoder and settings of dSpark. Same format as Table 2.

BackboneModelGSM8KMATHMBPPHumanEval
LLaDAdParallel75.8 / 8.1731.2 / 5.4336.6 / 3.8539.0 / 5.21
LLaDAd3LLM72.6 / 9.7429.8 / 5.6836.2 / 4.7938.4 / 6.15
LLaDAdSpark78.9 / 10.1232.0 / 6.0340.0 / 5.5239.6 / 6.73
DreamdParallel82.1 / 3.9937.6 / 3.3655.8 / 2.5257.3 / 2.92
Dreamd3LLM80.9 / 5.0138.1 / 3.9555.2 / 3.1556.1 / 3.38
DreamdSpark83.6 / 5.1238.9 / 4.3156.0 / 3.7757.3 / 4.35

The ranking does not depend on the decoder. Across the 16 settings (two decoders, two backbones, four benchmarks), dSpark has the highest TPF in 15 and the highest or tied accuracy in 15. Both comparisons are in Appendix D.4 of the revised paper (Table 17).

AuthorsOpenReview ↗

Response to Reviewer SJyz [3/3]

W3: Confidence sharpening is a separate ingredient in Equation 7, yet its effect is not isolated in the main ablations. A direct comparison with sharpening removed would clarify how much improvement comes from informative ordering versus increased commitment confidence.

Thanks for the suggestion. Following it, we ran the additional comparison and extended Table 2 of the paper with the sharpening term removed, which gives a full factorial ablation on LLaDA (Table 4). The revised Table 2 and Section 5.3 include the new rows.

Table 4. Full factorial ablation of the optimized schedule, KL regularization and sharpening on LLaDA. The last four rows (with sharpening) are Table 2 of the submission, and the first four rows are new in the revised Table 2.

ScheduleKLSharp.GSM8K TPFGSM8K Acc.GSM8K AUPHumanEval TPFHumanEval Acc.HumanEval AUP
6.175.2401.84.338.474.5
✓9.575.0613.47.436.697.8
✓5.377.5388.63.139.064.3
✓✓8.477.9623.45.939.096.4
✓7.674.9463.45.237.982.2
✓✓11.474.2718.08.335.1105.8
✓✓6.378.0458.43.740.274.3
✓✓✓10.178.9761.76.739.6106.2

Informative ordering accounts for the larger part of the improvement. Under KL regularization and with sharpening removed, the optimized schedule still raises TPF substantially on both GSM8K and HumanEval at similar accuracy. Sharpening alone contributes much less, with a gain on natural schedules that is several times smaller on both benchmarks. The two gains combine in the full model, which gives the best accuracy-parallelism trade-off (AUP) on both benchmarks.

W4: Figure 4 connects two operating points per method. Consequently, its interpolated curves and AUP scores provide limited evidence for a measured accuracy-parallelism frontier across decoding thresholds.

Thanks for the comment. We clarify that we measured the frontier directly by sweeping the decoding threshold, starting from one token per forward pass. To keep the comparison controlled, all models share one decoder within a table. Table 5 uses the confidence decoder of dSpark and Table 6 the entropy decoder of d3LLM, on GSM8K and MATH with LLaDA. The revised paper plots these measured points without interpolation in Figure 6 (Appendix A.6).

Table 5. Threshold sweep with the confidence decoder of dSpark (threshold τ) on LLaDA. Each cell shows accuracy (%) / TPF.

BenchmarkModelTPF = 1τ = 0.99τ = 0.95τ = 0.9
GSM8KdParallel76.9 / 1.0076.2 / 2.3176.4 / 5.4275.8 / 8.17
GSM8Kd3LLM75.6 / 1.0075.3 / 2.8774.7 / 6.1172.6 / 9.74
GSM8KdSpark80.1 / 1.0080.0 / 2.8179.4 / 6.9878.9 / 10.12
MATHdParallel32.4 / 1.0032.7 / 1.7631.8 / 3.6531.2 / 5.43
MATHd3LLM32.8 / 1.0032.3 / 1.8931.2 / 4.0129.8 / 5.68
MATHdSpark32.9 / 1.0032.7 / 1.8532.0 / 4.2832.0 / 6.03

Table 6. Threshold sweep with the entropy decoder of d3LLM (threshold η) on LLaDA. Each cell shows accuracy (%) / TPF.

BenchmarkModelTPF = 1η = 0.1η = 0.3η = 0.5
GSM8KdParallel76.4 / 1.0076.2 / 3.1475.9 / 6.2375.8 / 8.60
GSM8Kd3LLM75.5 / 1.0075.2 / 3.8974.7 / 7.0174.2 / 9.05
GSM8KdSpark80.4 / 1.0079.8 / 3.9079.1 / 7.3878.8 / 9.51
MATHdParallel32.0 / 1.0031.6 / 2.0131.5 / 3.9330.8 / 4.92
MATHd3LLM32.4 / 1.0032.2 / 2.4531.2 / 4.2830.3 / 5.75
MATHdSpark32.7 / 1.0032.5 / 2.5132.7 / 4.5932.3 / 5.89

Across the whole sweep and under both decoders, dSpark has the highest or tied accuracy at every threshold as parallelism increases, which confirms the frontier shown in Figure 4 of the paper with measured points.

Q2: With a shared decoder and matched training budget, how much benefit remains from optimized schedules alone? Please include the result without entropy sharpening and the exact operating-point settings.

Thanks for the question. With the decoder, the data and the training budget fixed, the optimized schedule alone, without sharpening, raises TPF by about 1.6$\times$ on GSM8K and 1.9$\times$ on HumanEval at similar accuracy (KL rows of Table 4 without sharpening). Giving the natural-schedule student the construction budget as extra training does not close the gap to dSpark in either TPF or accuracy (see W1). Under a shared decoder, dSpark also has the highest TPF among the distilled models in 15 of 16 settings (Tables 2 and 3). The exact operating-point settings of every method are listed in Table 1. In the revised paper, these results are in Table 2 and in Appendix A.7, C and D.4.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.