- Rating
- 3(+1)
- Confidence
- 3
- Informativeness
- 2
- Presentation
- 3
- Soundness
- 2
Read review
Paper summary
This paper proposes MaskHorizon, a simple training-free framework for open-ended diffusion language model decoding that grows the workspace block by block while keeping a local reserve of future masked positions as context. The method avoids relying on model-derived signals such as EOS or confidence for workspace growth and achieves lower decoding cost across multiple tasks and two DLM backbones.
However, the main evaluation uses only 100 examples for several benchmarks, which is too limited to support reliable accuracy conclusions, especially when many reported quality differences are only a few percentage points.
Critical strengths
(1) The idea of this work is interesting. It does not require signals such as EOS or confidence to adjust the inference length of DLMs. Instead, it performs inference block by block in a way similar to AR generation, while keeping a certain amount of consecutive masks as context.
(2) The proposed method is simple and clear.
Critical weaknesses
(1) The evaluation uses only 100 samples, which is too few. Considering that this is a training-free inference method, and the proposed method itself accelerates inference, the evaluation should not require excessive computational resources. However, the main evaluation only uses 100 samples from the test sets (e.g., the full GSM8K test set contains 1,319 samples and MBPP contains 500 samples). This makes the evaluation results less stable than evaluation on the full datasets and insufficient to support reliable accuracy conclusions.
(2) The metrics used to measure speedup are not sufficiently reasonable. The authors point out that different methods process different numbers of tokens at each step, and therefore NFE is not a suitable metric, but NFE is still reported in the main results. In addition, the authors introduce the CFP metric. However, the computational cost of the model is not directly proportional to the number of tokens, which is related to the complexity of attention. Therefore, CFP is still not a sufficiently reasonable metric for measuring computational cost.
Questions for authors
(1) The hyper-parameter initial length N0 uses different values for different benchmarks. How were these task-specific defaults determined?
13 comments
CFP therefore complements NFE as a measure of cumulative token processing, rather than a complete measure of computation. To address your concern about quadratic attention cost, we additionally report estimated FLOPs that include this cost, alongside actual decoding time. In the revised manuscript, we will present these two measures in the main text and retain the CFP analysis and results in the appendix as supplementary evidence.
We estimate decoding FLOPs from the recorded per-step input lengths, including the prompt, and the model architecture. The estimate covers Q/K/V/O projections, the MLP, the vocabulary output projection, and both attention matrix multiplications (QKᵀ and AV), thereby accounting for quadratic attention cost. One multiply-add counts as two FLOPs. Table 2 reports mean TFLOPs per sample; ratios compare the displayed mean computational costs with MaskHorizon, not measured speedup; valid sample counts may differ.
As shown in Table 2, after including the quadratic term, the estimated computational cost reductions across the 20 baseline comparisons using the displayed means range from 16.94% to 72.41%, indicating that the computational savings of the listed configurations do not arise from ignoring the quadratic cost of attention.
Table 2
| Model | Dataset | AdaBlock-dLLM | DAEDAL | Fast-dLLM | MaskHorizon |
|---|---|---|---|---|---|
| LLaDA | GSM8K | 1679.78(3.62×) | 724.61(1.56×) | 1324.21(2.86×) | 463.51(1.00×) |
| MATH500 | 2307.42(1.71×) | 2172.01(1.61×) | 2689.90(2.00×) | 1345.65(1.00×) | |
| MBPP | 2205.94(3.35×) | 1709.22(2.60×) | 2082.96(3.17×) | 657.62(1.00×) | |
| HumanEval | 3506.24(2.94×) | 3826.26(3.20×) | 3316.19(2.78×) | 1193.93(1.00×) | |
| Dream | GSM8K | 1229.02(2.71×) | 545.64(1.20×) | — | 453.21(1.00×) |
| MATH500 | 2329.34(3.02×) | 1164.13(1.51×) | — | 770.21(1.00×) | |
| MBPP | 1797.56(2.47×) | 938.32(1.29×) | — | 726.33(1.00×) | |
| HumanEval | 2095.73(2.77×) | 1134.82(1.50×) | — | 756.68(1.00×) |
We also provide actual decoding time to reflect runtime factors not covered by the theoretical estimates above. Table 3 reports the mean decoding time per sample in seconds, excluding model loading, warm-up, and scoring. Times come from historical runs across different periods and concurrent workloads and are descriptive, not controlled speedup measurements.
Table 3
| Model | Dataset | AdaBlock-dLLM | DAEDAL | Fast-dLLM | MaskHorizon |
|---|---|---|---|---|---|
| LLaDA | GSM8K | 15.92 | 7.01 | 12.46 | 4.51 |
| MATH500 | 21.49 | 20.18 | 25.23 | 12.49 | |
| MBPP | 20.89 | 15.76 | 19.70 | 6.33 | |
| HumanEval | 32.77 | 34.59 | 30.94 | 11.01 | |
| Dream | GSM8K | 10.56 | 5.47 | — | 4.44 |
| MATH500 | 19.98 | 11.43 | — | 7.26 | |
| MBPP | 15.35 | 9.17 | — | 6.95 | |
| HumanEval | 17.75 | 10.89 | — | 7.16 |
Q1
The hyper-parameter initial length N0 uses different values for different benchmarks. How were these task-specific defaults determined?
R3:
Thank you for your question. The initial lengths used for GSM8K, MATH500, MBPP, and HumanEval in the original manuscript are task-level empirical settings, rather than optimal values determined through a unified search.
We compared N₀=192, 224, and 256 on the full evaluation sets of all four tasks using LLaDA, fixing all other parameters and evaluation procedures. We report every tested setting and retain the original defaults without reselecting them. Table 4 reports "Score (%) / NFE." The results assess applicability within these tasks, rather than establish a universally optimal initial value.
The N₀ sensitivity results in Table 4 show that the GSM8K default of 192 maintains comparable quality with lower NFE; the MATH500 default of 256 achieves a higher score within the tested range, but at a higher cost; MBPP achieves its highest score at 224, whereas the default of 192 has a lower cost; and increasing HumanEval from the default of 224 to 256 passes only one additional problem while requiring more calls. We therefore regard the defaults as task-level empirical starting points.
Table 4
| Dataset | N₀=192 | N₀=224 | N₀=256 |
|---|---|---|---|
| GSM8K | 83.32 / 79.83 | 83.09 / 86.64 | 83.24 / 92.55 |
| MATH500 | 34.00 / 150.87 | 36.20 / 165.26 | 39.60 / 173.18 |
| MBPP | 37.88 / 103.61 | 39.88 / 111.42 | 38.88 / 115.35 |
| HumanEval | 40.85 / 136.78 | 43.29 / 144.90 | 43.90 / 148.85 |
References
[1] Jing Xiong, Qiujiang Chen, Fanghua Ye, Zhongwei Wan, Chuanyang Zheng, Chenyang Zhao, Hui Shen, Alexander Hanbo Li, Chaofan Tao, Haochen Tan, Haoli Bai, Lifeng Shang, Lingpeng Kong, and Ngai Wong. ATTS: Asynchronous Test-Time Scaling via Conformal Prediction. In International Conference on Learning Representations (ICLR), 2026.
[2] Juntao Zhao, Wenhao Lu, Sheng Wang, Lingpeng Kong, and Chuan Wu. QSpec: Speculative Decoding with Complementary Quantization Schemes. In Proceedings of the 2025 Conference on Empirical Methods in Natural Language Processing (EMNLP), pp. 4779–4795, 2025. DOI: 10.18653/v1/2025.emnlp-main.240.
[3] Zipeng Gao, Zhi Zheng, Qingrong Xia, Junda Lin, Ziwei Zhao, Tong Xu, Zhefeng Wang, and Enhong Chen. Unlocking Parallelism in Autoregressive Language Models via Speculative Decoding with Progressive Tree Drafting. In Conference on Language Modeling (COLM), 2026.
[4] Andre He, Sean Welleck, and Daniel Fried. Reasoning with Latent Tokens in Diffusion Language Models. In Advances in Neural Information Processing Systems (NeurIPS), 2026.
[5] Aiwei Liu et al. WeDLM: Reconciling Diffusion Language Models with Standard Causal Attention for Fast Inference. In International Conference on Machine Learning (ICML), 2026.
Thank you for carefully reviewing our paper and providing specific suggestions for improvement. We particularly appreciate your recognition of the value of our approach: rather than relying on model signals such as EOS probability or confidence to determine when the workspace grows, it advances block by block with decoding progress while retaining contiguous future MASK tokens as context. We also appreciate your positive assessment of the simplicity and clarity of the method.
W1
R1:
We agree that 100 examples are insufficient to establish small accuracy differences. The original evaluation used a shared 100-example subset to compare workspace and commitment strategies for efficiency and mechanistic analysis. Similar settings appear in ATTS (MATH100) [1], QSpec (100 examples per task across six tasks) [2], and PTD (100 examples each from GSM8K and MBPP) [3]. These precedents do not replace larger-scale quality evaluation, which we therefore added in response to your suggestion.
We added expanded evaluations of LLaDA and Dream on GSM8K (1,319 examples), MATH500 (500 examples), and MBPP (up to 500 examples), rather than reporting results based only on the original subset of 100 examples. Below, we present generation quality and the number of model calls; time and estimated computational costs are listed separately in W2.
Table 1 shows that MaskHorizon achieves slightly higher scores than the listed baselines on all three tasks with LLaDA. NFE is lower than all listed baselines on GSM8K and MBPP; on MATH500, it is lower than DAEDAL and Fast-dLLM but higher than AdaBlock-dLLM. These full-scale results extend the quality–cost evaluation beyond the original 100-example subsets.
Quality and Decoding Statistics
Table 1
W2
R2:
Thank you for raising this concern. We agree that neither NFE nor CFP alone fully captures computational cost. We retain NFE for comparison with prior work, but use CFP to account for an important distinction that NFE overlooks: each forward call may process a different number of tokens. This accounting follows related evaluation practices in SIDM [4], which accumulates the sequence lengths processed during inference, and WeDLM [5], which counts token instances processed across decoding forwards, including recomputation. While WeDLM excludes initial prefill, our CFP = ∑ᵢNᵢ includes the prompt in each forward input length Nᵢ.