acceptodds
Under review as a conference paper at ICLR 2027

Improving Parallel Decoding of Block Diffusion Language Models with Autoregressive Generation Order

Abstract

Diffusion language models enable flexible arbitrary-order generation, but existing token sampling methods are mostly designed for early masked diffusion models (MDMs). In this work, we study token sampling for recent block diffusion language models (BDLMs). We show empirically and analytically that these models are naturally more aligned with left-to-right decoding than MDMs. Beyond their AR-pretrained initialization, we attribute this alignment to the block diffusion training objective, which more frequently exposes the model to the left-to-right context used during AR decoding. Based on this observation, we propose Parallel Autoregressive Decoding (PARD), a simple training-free sampling method that preserves left-to-right unmasking structure while allowing parallel token commitment. Experiments show that PARD consistently improves the speed-quality trade-off over existing parallel samplers. Across three recent BDLMs and six benchmarks, it achieves the best average generation quality among parallel samplers. Compared with pure AR decoding, PARD achieves substantial speedups with only a small quality gap.

Primary area:
foundation or frontier models, including LLMs
Keywords:
diffusion language model, block diffusion language model

Reviews

Among:
Shape:
1234this paper, Top 15%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.00 over 3 reviews, above 85% of the others.
Reviewer_yAXjOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

Motivated by the observations that BDLMs are biased towards AR generation, which is both shown empirically and justified theoretically in the paper, the authors propose a token sampling method for block DLMs where the decoding order is biased towards autoregressive generation order. For a selection criteria of choice (e.g. confidence threshold) the proposed sampling technique, called PARD, selects consecutive tokens for which the selection is fulfilled until the first token that does not meet the criteria. Experiments show promising results for the proposed technique.

Critical strengths

The critical strength is the simplicity of the proposed technique, which does not seem to introduce any new hyperparameters, nor requires specific post-training, which make it very easily deployable in practice.

Critical weaknesses

Authors themselves make an observation that the proposed sampling technique has largest edge on coding benchmarks (l. 442). The entire motivation for the proposed method rests on the results obtained on the coding benchmark and MDMs are never benchmarked after section 3.1. Hence, I am not 100% convinced that AR order bias in BDLMs is as general as the paper claims. From Table 2, we see that AR outperforms other sequential methods most strongly on code, with code also being the domain where PARD has largest edge against other baselines. On the other benchmarks, PARAD does outperforming the rest in majority of cases, but the gains are very small in many cases (same for AR) and can potentially be attributed to noice effects?

Questions for authors
  • I would like to see whether and to which extent the AR bias is present on other domains than code (e.g. add other domains to section 3.1), this could better justify the more general claim the paper makes about AR bias in BDLMs
  • the paper is missing important baselines of learned unmasking policies like e.g. Jazbec et. al (ICML Oral) ,"Learning unmasking policies for diffusion language models". Also I wonder whether and how the observed AR bias can be used to improve these policies?
  • Table 2 average seem to be biased to coding domains, which are the ones on which proposed methods performs the best
  • Table 2 presents performance numbers with standard hyperparamters applied. Could authors also present throughout numbers at those hyperparameters for methods other than just AR and PARD (Table 3)? I would like to see numbers for e.g. confidence parallel and possibly some other parallel methods there as well (I would expect confidence parallel to have similar throughout given that it has similar excepted tokens pattern from Fig. 15).
  • regarding technical details: could authors also share which inference backend is being used for the evail as well as what batch size? I understand that the default batch size for DLM evals is usually 1, so I am assuming batch size here is also 1.
  • DMax (Chen, Zigeng, et al. "Dmax: Aggressive parallel decoding for dllms.") relies on a similar heuristic of keeping the decoded regions contiguous left-to-right based on a confidence threshold. I would appreciate a discussion on how the proposed method relates to that of DMax (in its motivation and mechanics)
4 comments

W1 & Q1. AR bias beyond coding benchmarks.

Thank you for the question. We conducted the AR bias analysis of Section 3.1 on GSM8K (math reasoning). We first compare the accuracy obtained from pure AR sampling with the arbitrary-order samplers (confidence/entropy/margin) across SDAR, Fast-dLLM v2, LLaDA, and Dream:

MethodFast-dLLM v2 7BSDAR 8BLLaDA 8BDream 7B
Confidence81.789.878.875.9
Entropy82.091.077.873.2
Margin81.389.578.571.7
AR83.791.178.175.4

For Fast-dLLM v2, its accuracies differ from Table 2 because we disable sub-block inference here to use the same block size of 32 as other models (discussed in Appendix C.1). We have also corrected Table 1 accordingly for Fast-dLLM v2; the conclusions remain the same.

The results are similar to those of HumanEval. Pure AR achieves the highest accuracy on both BDLMs. On both MDMs, pure AR falls slightly below Confidence, by 0.7 points on LLaDA and 0.5 points on Dream. In comparison, pure AR achieves higher accuracy than Confidence by 2.0 on Fast-dLLM v2 and 1.3 on SDAR.

As in Figure 1 of the submission, we record the within-block position unmasked at each step under Confidence. The average probability mass on the diagonal is 0.470.47 and 0.370.37 for Fast-dLLM v2 and SDAR, versus 0.270.27 and 0.300.30 for LLaDA and Dream. Within ±2\pm 2 positions of the diagonal, the mass is 0.810.81 and 0.740.74 for the BDLMs, versus 0.700.70 and 0.670.67 for the MDMs. The decoding orders of the BDLMs are again more concentrated along the diagonal, indicating stronger AR bias.

Then we compare the local and global AR-ness at different kk under the Confidence sampler for the four models:

Local AR-ness@kk on GSM8K

Modelk=1k=1k=2k=2k=3k=3k=4k=4k=5k=5k=6k=6k=7k=7k=8k=8
Fast-dLLM v2 7B0.6290.5370.4830.4470.4190.3970.3800.365
SDAR 8B0.5300.4260.3660.3270.2990.2770.2600.246
LLaDA 8B0.2720.1740.1330.1140.1030.0970.0910.089
Dream 7B0.3990.2740.2140.1780.1550.1370.1250.117

Global AR-ness@kk on GSM8K

Modelk=1k=1k=2k=2k=3k=3
Fast-dLLM v2 7B0.5900.8000.900
SDAR 8B0.5440.7520.865
LLaDA 8B0.3970.6370.783
Dream 7B0.4050.6170.748

Both BDLMs have substantially higher local and global AR-ness than both MDMs at every kk. This is consistent with Figures 1 and 2 in the submission. We have added these GSM8K results to Appendix D.5 of the revised paper and referred to them in Section 3.1.

W1. MDMs are never benchmarked after Section 3.1.

Thank you for raising this point. We conduct a preliminary evaluation to compare PARD with Confidence Parallel at τc=0.9\tau_c=0.9 on LLaDA 8B on GSM8K and HumanEval Base:

LLaDA 8BGSM8KHumanEval
Confidence Parallel77.643.9
PARD78.240.9

On GSM8K, PARD only slightly improves over Confidence Parallel, by 0.6 points. On HumanEval, PARD degrades performance by 3.0 points. Thus, PARD does not consistently outperform Confidence Parallel on LLaDA. In contrast, on all three BDLMs in Table 2, PARD outperforms Confidence Parallel on both benchmarks, by 1.5 to 6.7 points. This is consistent with our analysis in Section 3.1 that BDLMs have a stronger AR bias than MDMs. Hence, we focus our evaluation on BDLMs, which are also the more common choice among recent high-performing DLMs (as discussed in Section 2.1).

W2. Small gains on non-coding benchmarks and noise effects.

We agree that the gains are larger on coding tasks. Code generation is particularly sensitive to token ordering, as earlier tokens often determine subsequent code structure. Appendix D.3 illustrates this with an example.

However, we would like to emphasize that the benefit of PARD is not limited to code. The table below reports the average performance on the non-coding benchmarks (GSM8K, MATH, IFEval, and BBH):

MethodFast-dLLM v2 7BSDAR 8BLLaDA2.1-Mini 16B
KLASS64.370.085.9
EB-Sampler63.471.185.7
Confidence Parallel63.569.684.8
APD62.169.4--
Hierarchy62.368.384.9
PARD66.072.386.2
AR (sequential)66.673.086.4

As in Table 2, APD is not reported for LLaDA2.1-Mini because no small AR verifier is available.

PARD achieves the highest non-coding average among all parallel samplers on all three models. Compared with Confidence Parallel, the default parallel sampler of these BDLMs, PARD improves the non-coding average by 2.5, 2.7, and 1.4 points on Fast-dLLM v2, SDAR, and LLaDA2.1-Mini, respectively.

We used greedy decoding in the main experiments, so the generation process is deterministic. For a fixed model, sampler, and prompt, repeated runs produce essentially identical pass@1 scores. Thus, the reported differences cannot be attributed to stochastic token sampling. In addition, we evaluate every benchmark on its full test set rather than a randomly sampled subset. For BBH on LLaDA2.1-Mini, compute constraints limit us to 8 of the 27 tasks (discussed in Section 5.2), but each of these tasks is still evaluated in full. The reported scores are therefore not affected by noise from sampling test problems. Appendix D.1 (Figures 11–14) further shows that PARD achieves stronger speed-quality trade-offs on GSM8K compared with other parallel samplers. Consistent improvements across models, tasks, and thresholds support the benefit of PARD.

Q2. Comparison with learned unmasking policies (Jazbec et al.).

Thank you for the suggestion on comparison with learned unmasking policies. We compare PARD with the learned unmasking policy of Jazbec et al. on SDAR 8B. We adapt the official code to SDAR and follow the original training configuration. As in the original paper, we train a code policy on the KodCode-RL-10K dataset. The reward weight α\alpha controls the speed–quality trade-off of the policy. The original paper sweeps α∈{10,3,1,0.3,0}\alpha \in \{10, 3, 1, 0.3, 0\} rather than fixing a default. Since each α\alpha requires a separate training run (65.9 GPU-h), we train one policy and compare PARD against it at matched speed below. We use α=1.0\alpha=1.0, which the authors commonly use in their experiments that report a single policy rather than a Pareto front. We evaluate the trained unmasking policy on HumanEval and MBPP:

MethodTrainingHE pass@1HE tok/sMBPP pass@1MBPP tok/s
Learned policy (α=1.0\alpha=1.0)65.9 GPU-h61.695.664.590.6
PARD (0.9)None81.177.677.072.6
PARD (0.7)None72.698.873.890.7

Notably, the learned policy uses domain-specific training data, whereas PARD requires neither training nor task-specific data. Training the policy with GRPO takes 16.5 hours on 4 RTX A6000 GPUs (65.9 GPU-h). With the default confidence threshold of 0.9, PARD achieves substantially higher accuracy than the learned policy, while the learned policy is faster. With a lower threshold of 0.7, PARD is both faster and more accurate than the learned policy. It improves pass@1 by 11.0 and 9.3 points on HumanEval and MBPP, at equal or higher throughput. Thus, PARD outperforms the learned policy in speed and generation quality.

Regarding how the AR bias can improve learned policies, Jazbec et al. use expert steering to help exploration during GRPO training. They add rollouts from a confidence-based sampler to each group as expert trajectories. We suggest using PARD as the expert sampler for BDLMs. Since PARD exploits the AR bias of BDLMs, it may provide higher-reward trajectories that guide the learned policy toward better unmasking orders. We leave this direction for future work.

Q3. Table 2 average biased toward coding.

Thank you for pointing this out. We agree that the original Avg column in Table 2 could be misleading. We have updated Table 2 of the revised paper with a macro-average that weights the six benchmarks equally: we first average Base and Plus within HumanEval and MBPP, and then average the six benchmark scores.

PARD still achieves the highest macro-average among all parallel samplers on all three models. It improves over Confidence Parallel by 3.8, 3.2, and 2.0 points on Fast-dLLM v2, SDAR, and LLaDA2.1-Mini, respectively, compared with 4.5, 3.4, and 2.4 points under the original averages.

Q4. Throughput of other parallel samplers at Table 2 settings.

The table below reports the throughput (tokens/sec) of all parallel samplers under the hyperparameters of Table 2. Numbers in parentheses indicate speedup relative to sequential AR decoding.

Fast-dLLM v2 7B

MethodHumanEvalMBPPGSM8KMATHIFEvalBBH
AR85.882.796.169.664.877.6
KLASS90.4 (1.05×)84.7 (1.02×)95.8 (1.00×)82.5 (1.19×)76.1 (1.17×)78.0 (1.01×)
EB-Sampler131.8 (1.54×)113.4 (1.37×)183.7 (1.91×)151.6 (2.18×)111.9 (1.73×)119.5 (1.54×)
APD187.3 (2.18×)150.0 (1.81×)209.4 (2.18×)180.4 (2.59×)129.1 (1.99×)131.5 (1.69×)
Hierarchy146.1 (1.70×)131.9 (1.59×)188.3 (1.96×)166.1 (2.39×)120.4 (1.86×)124.0 (1.60×)
Confidence Parallel137.5 (1.60×)120.1 (1.45×)179.8 (1.87×)156.8 (2.25×)110.6 (1.71×)114.4 (1.47×)
PARD134.9 (1.57×)112.5 (1.36×)179.9 (1.87×)152.0 (2.18×)111.2 (1.72×)113.5 (1.46×)

SDAR 8B

MethodHumanEvalMBPPGSM8KMATHIFEvalBBH
AR46.542.446.240.732.040.1
KLASS50.3 (1.08×)49.5 (1.17×)57.1 (1.24×)60.7 (1.49×)42.9 (1.34×)40.8 (1.02×)
EB-Sampler70.1 (1.51×)64.5 (1.52×)99.2 (2.15×)114.1 (2.80×)50.3 (1.57×)54.8 (1.37×)
APD120.0 (2.58×)96.8 (2.28×)133.2 (2.88×)141.6 (3.48×)70.3 (2.20×)81.3 (2.03×)
Hierarchy87.0 (1.87×)88.6 (2.09×)124.3 (2.69×)137.6 (3.38×)68.5 (2.14×)66.0 (1.65×)
Confidence Parallel75.5 (1.62×)74.2 (1.75×)106.6 (2.31×)123.8 (3.04×)58.9 (1.84×)57.3 (1.43×)
PARD77.6 (1.67×)72.6 (1.71×)108.2 (2.34×)124.9 (3.07×)59.4 (1.86×)60.5 (1.51×)

LLaDA2.1-Mini

MethodHumanEvalMBPPGSM8KMATHIFEvalBBH
AR12.110.610.010.99.111.6
KLASS19.2 (1.59×)15.0 (1.42×)13.0 (1.30×)14.3 (1.31×)8.5 (0.93×)13.4 (1.16×)
EB-Sampler21.7 (1.79×)18.7 (1.76×)19.1 (1.91×)18.6 (1.71×)10.0 (1.10×)17.1 (1.47×)
Hierarchy45.3 (3.74×)38.1 (3.59×)26.9 (2.69×)31.5 (2.89×)11.5 (1.26×)27.7 (2.39×)
Confidence Parallel43.6 (3.60×)36.0 (3.40×)25.3 (2.53×)30.0 (2.75×)10.6 (1.16×)24.3 (2.09×)
PARD44.1 (3.64×)36.8 (3.47×)29.4 (2.94×)28.9 (2.65×)11.7 (1.29×)25.2 (2.17×)

As expected, Confidence Parallel and PARD achieve similar throughput on most models and tasks, consistent with their similar tokens/step patterns in Figure 15. PARD thus matches the speed of Confidence Parallel while improving its average accuracy by 3.8, 3.2, and 2.0 points on Fast-dLLM v2, SDAR, and LLaDA2.1-Mini, respectively (Table 2 in Section 5.2). KLASS runs slow due to its high precision operation (as discussed in Section 5.1), and EB-Sampler is similar to or slower than PARD.

Note that on Fast-dLLM v2 and SDAR, Hierarchy and APD are faster than PARD on HumanEval and MBPP, by 1.081.08–1.22×1.22\times and 1.331.33–1.55×1.55\times, respectively. However, on these tasks their base pass@1 is lower than that of PARD, by 3.5–12.2 points for Hierarchy and 4.5–10.4 points for APD (Table 2 in Section 5.2). We also showed that PARD achieves a better speed–quality trade-off than APD and Hierarchy in Section 5.1.

Q5. Inference backend and batch size.

We used HuggingFace Transformers (without model compilation) as the inference backend, on a single RTX A6000 GPU. Running inference with batch size 1 under-utilizes the A6000 GPU. Given the sufficient GPU memory, we therefore set the batch size to 4 to speed up our experiments. We used the same batch size for all methods for a fair comparison.

To confirm that PARD's speedup also holds at batch size 1, we measure the throughput (tok/s) of SDAR 8B at batch size 1 on HumanEval and GSM8K. Speedup over AR is shown in parentheses.

MethodHumanEvalGSM8K
AR16.215.8
Confidence Parallel37.0 (2.29$\times$)45.0 (2.85$\times$)
PARD37.6 (2.32$\times$)43.3 (2.75$\times$)

PARD's speedup over AR is larger at batch size 1 than at batch size 4 (Table 3): 2.32×2.32\times vs. 1.67×1.67\times on HumanEval, and 2.75×2.75\times vs. 2.34×2.34\times on GSM8K. The speedup decreases as the batch size grows because the sequences in a batch move to the next block together. All sequences need to wait for the slowest sequence in each block. The speedups we report at batch size 4 are therefore conservative relative to the standard batch size of 1. The throughput of PARD is also comparable to that of Confidence Parallel at both batch sizes.

Q6. Relation to DMax.

Thank you for raising this related work. We have cited and discussed DMax in Sections 2.2 and 4 of the revised paper.

Mechanics. PARD-Confidence uses the same prefix selection rule as lines 5–8 of DMax's Algorithm 1. Both select the longest contiguous prefix of masked positions whose confidence exceeds the threshold, and fall back to the leftmost masked position if the prefix is empty. However, the two decoding procedures differ in three ways. First, in DMax the selected prefix positions are not finalized as tokens. Each one is fed back as a soft embedding, i.e., an interpolation between the embedding of its predicted token and the mask embedding, weighted by confidence, and its predicted token can change at every later step. In PARD, the selected positions are unmasked as discrete tokens. Second, DMax requires On-Policy Uniform Training (OPUT) to handle soft embeddings. When Algorithm 1 is run on the original LLaDA-2.0-mini without OPUT, GSM8K accuracy drops to 0.0% (DMax, Table 3). PARD operates in the discrete token space, so it applies directly to off-the-shelf BDLMs without training. Third, DMax uses confidence both as the selection score and as the interpolation weight. PARD is criterion-agnostic and also applies the prefix constraint to Margin and Entropy samplers.

Motivation. DMax and PARD adopt the prefix selection rule for different reasons. In DMax, the prefix selection rule is introduced heuristically to mitigate error accumulation during aggressive parallel decoding. PARD follows from our analysis of the stronger AR bias of BDLMs compared with MDMs (Section 3, Propositions 1–2). A prefix constraint is the simplest way to follow this bias while keeping parallel decoding.

Reviewer_bKFrOpenReview ↗
Rating
4(+1)
Confidence
4
Informativeness
3
Presentation
3
Soundness
3(+1)
Read review
Paper summary

The authors analyze unmasking strategies for diffusion language models (DLMs) and block diffusion language models. They empirically point out that these models already have a strong autoregressive bias under confidence-based (including entropy-based and margin-based) decoding, and demonstrate empirically and theoretically that this behavior is aggravated in modern block diffusion language models. They propose to adapt these confidence-based decoding strategies to always unmask a contiguous set of tokens from the current unmasked frontier, stopping at the first token whose confidence does not pass a threshold. They demonstrate across a wide range of models and benchmarks that this significantly improves the performance of diffusion language models, providing a reality-check to the arbitrary-order decoding aspirations that has long motivated the DLM community.

Critical strengths

The paper is very clearly written, with a cohesive and important message. The empirical evidence is extensive, spanning many SoTA diffusion language models and a long list of benchmarks. While being somewhat of a negative result to the DLM community, it provides a much-needed reality check to many of the claims made in current and past DLM papers.

Critical weaknesses

While the analysis on the limitations of DLMs is strong, the proposed solution only partially mitigates them. If unmasking is performed causally, then block DLMs do not possess any inherent capabilities beyond full autoregressive models, except for speed improvements. But speed improvements can also be achieved via speculative decoding, which seems much more effective. For example, DFlash with Qwen3-8B on an NVIDIA H100 GPU achieves over 900 tokens/second on MATH-500, without any quality degradation. This compares unfavorably to the 100-200 tok/s achieved here with DLMs (and at a significant drop in quality). The paper should either compare with modern speculative decoding methods, or at least discuss them and transparently state that the limitations identified by the paper call DLMs into question and that future DLM research may have to directly compete with modern speculative decoding approaches.

PARD is a bad acronym, since it overlaps with a well-known speculative decoding framework called PARD (with follow-up PARD-2).

Questions for authors

Questions:

  1. You write in line 125: "For the MDMs, we employ the semi-AR decoding strategy". What do you mean? Just block-diffusion unmasking?
  2. On line 291, you write x^sb,j\hat x_s^{b,j} = arg max ...$. Is the method really constrained to greedy decoding? Maybe you could keep this more general.
  3. The main text is missing the hardware used to obtain the throughput numbers (token/s). Also, what is the used backend/library? Is the model compiled?
  4. The throughput numbers are roughly one order of magnitude lower than what can be achieved with modern speculative decoding frameworks such ad DFlash, even at batch size 1 settings. Can you compare against DFlash on an equally-sized AR model?

Comments:

  1. The caption of Figure 1 would benefit from explicitly describing that confidence-based unmasking is used to decide which positions to unmask and that the probability distribution / heat map is taken over the dataset. Mentioning in the caption the key lesson that BDLM are more diagonal than MDLM and visually splitting BDLMs and MDMs would also be beneficial.
5 comments

Q1. You write in line 125: "For the MDMs, we employ the semi-AR decoding strategy". What do you mean? Just block-diffusion unmasking?

Yes. By "semi-AR decoding," we mean that at inference time, the output sequence is divided into contiguous blocks and unmasked block by block from left to right. Decoding proceeds to the next block only after the current block is fully unmasked. We use "semi-AR" to distinguish this inference procedure for full-sequence MDMs from the block-diffusion training of BDLMs.

Q2. Generalizing beyond greedy decoding.

Thank you for pointing this out. Yes, PARD supports both greedy and stochastic decoding. For each selected position j∈Usj \in U_s, we can do stochastic decoding by sampling x^sb,j\hat{x}_s^{b,j} from the predictive distribution pθ(x0b,j=v∣x0<b,xsb)p_\theta(x_0^{b,j} = v \mid x_0^{\lt b}, x_s^b), v∈Vv \in \mathcal{V}. For greedy decoding, we take x^sb,j=arg⁡max⁡vpθ\hat{x}^{b,j}_s = \arg\max_v p_\theta, as in our main experiments. Under stochastic decoding, entropy and margin scores continue to be computed directly from pθp_\theta, while the confidence score uses the predictive probability pθ(x0b,j=x^sb,j∣x0<b,xsb)p_\theta(x_0^{b,j} = \hat{x}_s^{b,j} \mid x_0^{\lt b}, x_s^b) of the sampled token (following standard stochastic confidence unmasking in LLaDA). We have revised the notation in Section 4 of the revised paper to cover both greedy and stochastic decoding.

Q3. Hardware, backend, and compilation for throughput.

We measured decoding throughput on a single NVIDIA RTX A6000 GPU using HuggingFace Transformers as the decoding backend (without model compilation). We used a batch size of 4 to measure throughput for all methods. We have added these details to Section 5 and Appendix C.3 of the revised paper.

W1. & Q4. Relationship to modern speculative decoding (e.g., DFlash) and implications for DLMs.

Thank you for this important point. We agree that current BDLMs still trail lossless speculative decoding of AR models in raw throughput, but an empirical benchmark against DFlash is outside the scope of our work.

  • Scope difference. Our paper studies training-free sampling for off-the-shelf BDLMs. We propose PARD and compare it with other BDLM samplers, demonstrating improvements in the speed-quality trade-off. Our main comparison asks how to decode an existing BDLM more efficiently. In contrast, DFlash accelerates AR models using a drafter trained separately for each target model. Thus, the comparison with DFlash would evaluate different inference approaches.

  • Training-free vs. trained drafter. DFlash requires training a separate drafter for each target model on ~800K prompts, which incurs substantial computational cost. In contrast, PARD requires neither training nor an additional model and can be applied directly to existing BDLM checkpoints.

  • Speculative decoding can also accelerate DLMs. Speculative decoding is not limited to AR models. Recent work has extended it to DLMs. For example, SimSD (arXiv:2606.02544) uses SDAR 1.7B as a drafter for SDAR 8B. MRP (arXiv:2605.18817) trains a lightweight module to predict next-step logit corrections from a frozen SDAR model's hidden states. Thus, comparing a BDLM decoded by PARD with an AR model accelerated by a trained drafter is not a like-for-like comparison.

We have added a new Limitations section in Appendix E of the revised paper to discuss the relationship to speculative decoding.

W2. Naming conflict with PARD.

Thank you for pointing out this naming conflict. We have renamed our method to PAUD (Parallel AUtoregressive Decoding) in the revised paper, including all figures. To stay consistent with our responses to the other reviewers, we keep the name PARD in this rebuttal.

C1. Clarifying Figure 1.

Thank you for this suggestion. We have updated Figure 1 in the revised paper. The caption now states that Confidence is used to select the unmasking positions, and that the heatmaps are computed over the HumanEval dataset. The figure also visually separates the BDLM and MDM panels, and the caption points out that the BDLM heatmaps are more concentrated along the diagonal.

Thank you for your response.

You have adequately answered most of my questions. I still believe a comparison with D-Flash is missing to transparently assess the state of accelerated inference with block diffusion language modeling. I believe this to be of essential importance to avoid misleading researchers on the capabilities of the current BDLM landscape, especially since PARD drops any arbitrary-order inference capabilities.

As a result, I intend to maintain my score.

Thank you for your reply. Although this comparison is outside the scope of our work for the three reasons discussed in our last reply, we agree that a comparison with DFlash can transparently assess the current inference speed of DLMs.

We compare SDAR 8B with Qwen3-8B using z-lab/Qwen3-8B-DFlash-b16 as the drafter. As references, we also include the AR decoding of Qwen3-8B and SDAR 8B (one output token per step). The throughputs (tok/s) are measured on A6000 with batch size 1. We first compare them using each model's HuggingFace Transformers implementation:

MethodSpeculative DecodingHumanEval (tok/s)GSM8K (tok/s)
Qwen3-8B + DFlash✓159.6134.2
Qwen3-8B×27.427.4
SDAR 8B AR×16.215.8
SDAR 8B Confidence Parallel×37.045.0
SDAR 8B PARD×37.643.3

PARD is 1.4× and 1.6× faster than Qwen3-8B AR on HumanEval and GSM8K. This is smaller than its speedup over SDAR AR (2.3× and 2.7×), because SDAR is slower than Qwen3-8B even under the same AR order. This is because each SDAR forward processes the whole 32-token block instead of one token, and once a block is fully decoded, SDAR runs one more forward to cache the committed block. DFlash accelerates Qwen3-8B AR by 5.8× on HumanEval and 4.9× on GSM8K, and is about 4.2× and 3.1× faster than SDAR with PARD.

We also run all models with the SGLang backend:

MethodSpeculative DecodingHumanEval (tok/s)GSM8K (tok/s)
Qwen3-8B + DFlash✓229.5194.8
Qwen3-8B×42.042.1
SDAR 8B AR×32.631.9
SDAR 8B Confidence Parallel×72.289.3
SDAR 8B PARD×72.088.3

Under SGLang, the AR speed gap between Qwen3-8B and SDAR narrows. With SGLang, PARD is 1.7× and 2.1× faster than Qwen3-8B AR on HumanEval and GSM8K, and 2.2× and 2.8× faster than SDAR AR. DFlash is about 3.2× and 2.2× faster than PARD.

We also compare their accuracies:

MethodSpeculative DecodingHumanEvalGSM8K
Qwen3-8B + DFlash✓85.489.2
Qwen3-8B×85.489.2
SDAR 8B AR×79.391.1
SDAR 8B Confidence Parallel×75.089.8
SDAR 8B PARD×81.191.4

Qwen3-8B achieves higher accuracy on HumanEval (85.4 vs. 81.1), while SDAR with PARD performs better on GSM8K (91.4 vs. 89.2). Neither model shows a consistent advantage in generation quality.

These results show that PARD already makes SDAR 8B faster than the equally sized AR model Qwen3-8B, by 1.4–2.1× across both backends, with comparable accuracy. However, DFlash reaches higher throughput, which shows that BDLM inference still has room to become faster. PARD is training-free, so it can be combined with future advances in BDLM training and inference engines. Beyond the practical speedup, our work identifies and explains the strong left-to-right decoding preference of current BDLMs, which provides a diagnostic foundation for future BDLM development. We have added this comparison into Appendix E of the revised paper.

Reviewer_7n92OpenReview ↗
Rating
2(+1)
Confidence
4
Informativeness
1
Presentation
3
Soundness
3
Read review
Paper summary

This paper studies token generation order in block diffusion language models. It reports that recent BDLMs exhibit stronger left-to-right decoding preferences than full-sequence masked diffusion models, and analyzes differences in their training-context distributions. Motivated by these observations, the authors propose Parallel Autoregressive Decoding (PARD), a training-free sampler that accepts the longest leftmost prefix satisfying a confidence, margin, or entropy criterion, falling back to the leftmost masked position when the accepted prefix is empty. Experiments on Fast-dLLM v2, SDAR, and LLaDA2.1-Mini evaluate generation quality and decoding efficiency across six benchmarks.

Critical strengths

The paper provides a useful empirical study of decoding order in recent BDLMs. It compares sequential and parallel samplers, includes threshold sweeps with both throughput and function-evaluation measurements, and evaluates across multiple model families. The reported results suggest that enforcing a prefix structure can improve generation quality while retaining parallelism. The comparison against fixed-size leftmost decoding also helps establish that adaptive acceptance matters, rather than simply generating several leftmost tokens at every step. PARD is straightforward to implement and requires neither additional training nor an auxiliary verifier. This makes its demonstrated effectiveness on existing checkpoints practically useful, independently of whether the underlying selection rule is novel.

Critical weaknesses

The novelty of the central decoding rule is insufficiently established. DMax (arXiv:2604.08302, Algorithm 1) already selects the longest contiguous high-confidence prefix and falls back to the leftmost masked token when that prefix is empty. These are also the defining position-selection operations of PARD-Confidence. Although the complete methods differ—DMax includes additional training and soft-state updates, whereas PARD is a training-free sampler—these differences do not establish novelty for the shared selection rule itself. The manuscript neither cites DMax nor explains what algorithmic contribution PARD adds beyond applying this rule to existing BDLMs and extending it to margin and entropy scores. The empirical improvements are useful, but their significance should be assessed with this overlap explicitly acknowledged.

The motivating observation also needs clearer differentiation from prior work. Section 3.1 cites Ni et al., The Flexibility Trap (arXiv:2601.15165), primarily for the benefits of autoregressive RL rollouts. However, that work also directly compares autoregressive and arbitrary-order inference on existing diffusion models, reports stronger pass@k scaling under autoregressive order, and examines how confidence-based ordering bypasses uncertain reasoning-critical tokens. Consequently, the broad observation that restricting generation order can benefit diffusion-model reasoning is not new. PARD’s potentially distinct contribution is the stronger AR preference observed in BDLMs, its implications for greedy decoding, and the training-context analysis. These narrower contributions deserve explicit positioning against the existing findings; the current framing leaves their incremental significance insufficiently established.

The broader relationship to causal parallel decoding is also underdeveloped. CLLMs and Jacobi Forcing explore parallel generation through Jacobi iterations under causal attention, while CARD and WeDLM combine causal attention with confidence-driven parallel decoding. WeDLM additionally motivates left-to-right commitment through prefix cacheability. These methods are not equivalent to PARD: they differ in training requirements, attention structure, and acceptance or convergence criteria, and some incorporate explicit candidate verification. Nevertheless, they provide important context for assessing the contribution of combining parallel prediction with ordered prefix advancement. The manuscript should distinguish its training-free prefix-selection heuristic from these established approaches, rather than leaving the reader to infer that distinction.

Questions for authors
  1. How does PARD-Confidence’s position-selection rule differ from DMax’s longest high-confidence prefix and leftmost-token fallback? Please distinguish differences in the selection rule from differences in training and state updates. If the rules are equivalent, what is the principal new methodological contribution of PARD beyond its application to existing BDLMs and its margin/entropy variants?
  2. How do the findings in Section 3 differ from the inference-time ordering analysis in Sections 3.1–3.2 of The Flexibility Trap, beyond the change from pass@k reasoning coverage to greedy pass@1 evaluation? In particular, which results establish a specifically BDLM-related insight, rather than extending an already observed benefit of autoregressive ordering to additional models?
  3. How should PARD be positioned relative to CLLMs, Jacobi Forcing, CARD, and WeDLM? Please clarify which contribution is specific to PARD beyond combining parallel prediction with left-to-right advancement, distinguishing its unchanged bidirectional intra-block attention and confidence-based commitment from causal-attention methods, fixed-point convergence, and candidate verification. A focused discussion would be sufficient where differences in training or architecture make direct experimental comparisons inappropriate.
6 comments

W1 & Q1. PARD-Confidence's position-selection rule overlaps with DMax (arXiv:2604.08302, Algorithm 1). How do the two methods differ, and what does PARD contribute beyond this rule?

Thank you for pointing us to DMax and recognizing the empirical improvements of our method.

Status of DMax. We were not aware of DMax when preparing our submission, as it was then a recent arXiv preprint (posted in April 2026). DMax was accepted to NeurIPS 2026 on September 24, 2026, one day before the ICLR 2027 submission deadline, so at submission time it was effectively an arXiv preprint. The ICLR 2027 Reviewer Guide states that authors "may be excused for not knowing about papers not published in peer-reviewed conference proceedings or journals, which includes papers exclusively available on arXiv," and that "the lack of such comparisons cannot be a basis for rejection." Nevertheless, we agree that discussing DMax provides valuable context, and we have added the discussion to the revised paper.

Procedural and conceptual differences between DMax and PARD. PARD-Confidence uses the same prefix selection rule as lines 5–8 of DMax's Algorithm 1. However, DMax's decoding procedure and the role of this rule in it differ from PARD in four important ways.

  1. Revisable soft embeddings vs. committed discrete tokens. In DMax, the selected prefix positions are not finalized as tokens. Each one is fed back as a soft embedding, i.e., an interpolation between the predicted token embedding and the mask embedding, weighted by confidence. The predicted token can change at every later step. In PARD, the selected positions are unmasked as discrete tokens.

  2. Retraining-dependent vs. training-free. In DMax, Algorithm 1 requires On-Policy Uniform Training (OPUT), because standard DLMs are trained on binary mask-to-token transitions and cannot accept soft embeddings. When Algorithm 1 is run on the original LLaDA-2.0-mini without OPUT, GSM8K accuracy drops to 0.0% at all tested thresholds (DMax, Table 3). In PARD, no training is needed as it directly operates in discrete token space. PARD applies directly to off-the-shelf BDLMs (e.g., Fast-dLLM v2, SDAR, LLaDA2.1-Mini) without modifying model weights.

  3. Confidence-coupled vs. criterion-agnostic. In DMax, predictive confidence serves as both the selection score and the interpolation weight between token and mask embedding. So its Algorithm 1 is tied to the predictive confidence. In contrast, PARD is formulated as a general criterion-agnostic method that adapts arbitrary-order parallel scoring rules, including Confidence, Margin, and Entropy. Notably, Section 5.1 shows that applying PARD to Margin yields the largest empirical gain among all criteria, making an otherwise weak criterion much more effective (Figure 3).

  4. Core motivation and theoretical grounding. In DMax, the prefix selection rule is introduced heuristically to mitigate error accumulation during aggressive parallel decoding. In PARD, the prefix selection rule follows from our analysis. Through empirical AR-ness analyses (Section 3.1) and theoretical comparisons on training-context distribution (Propositions 1, 2), we show that the block diffusion objective exposes models to exact AR mask patterns with higher probability than MDMs and achieves higher expected AR pattern alignment. These analyses identify the strong AR bias of BDLMs. We therefore propose PARD as the simplest method to enable AR order while keeping parallel decoding at inference time.

Contributions. In summary, beyond this selection rule, we make two contributions. First, we identify why existing BDLMs benefit from prefix-constrained decoding (Section 3). Second, we show empirically that the prefix constraint alone, without DMax's training or soft embedding refinement, substantially improves the speed–quality trade-off of off-the-shelf BDLMs (Section 5). We have cited and discussed DMax in Sections 2.2 and 4 of the revised paper.

W2. Citing the inference-time findings of Ni et al.

Thank you for pointing this out. We agree that Ni et al. have compared AR and arbitrary-order inference on MDMs. We have revised Section 3.1 to credit their inference-time findings of stronger pass@kk scaling for AR order and how confidence-based ordering bypasses uncertain reasoning-critical tokens.

Q2. Differences from Ni et al. and BDLM-specific findings.

The difference between our analysis and that of Ni et al. is not only a change of metric. Single-sample quality (pass@1) is the important setting for practical inference. Pass@kk costs kk generations and requires an oracle to select the correct response, which a deployed model does not have. Ni et al.'s goal is to improve reasoning capability of MDM with RL, so they use pass@kk to analyze the explorative reasoning potential of the model. In contrast, our work focuses on training-free sampling that improves the speed-quality trade-off of BDLMs at inference time. We therefore use single-sample accuracy (pass@1) to evaluate the quality of a sampler.

BDLM-specific insights. Three results in our Section 3 analysis are specific to BDLMs. First, Figures 1–2 show that, under the same arbitrary-order sampler (Confidence), the decoding orders of BDLMs exhibit a stronger AR bias than those of MDMs. Second, the pass@1 findings of Ni et al. on MDMs do not imply our pass@1 findings on BDLMs. They suggest that AR order brings little or no benefit on MDMs when k=1k=1, while our results show that AR order improves pass@1 on BDLMs. This difference leads to different uses of AR order. Ni et al. find that AR order helps mainly at large kk, i.e., it improves exploration, so they use it in RL rollouts. We find that AR order already helps BDLMs at k=1k=1, so it can be exploited directly at inference time by a training-free sampler, without RL or any other training. Third, we explain theoretically the difference between BDLMs and MDMs by comparing their training-context distributions (Propositions 1–2).

W3. Relationship to causal parallel decoding.

Thank you for pointing us to these works. We agree that prior methods already combine parallel prediction with left-to-right advancement.

CLLMs and Jacobi Forcing post-train AR models to speed up Jacobi decoding, which iterates until it reproduces the greedy AR output. In contrast, PARD needs no training. It commits tokens of an existing BDLM from a single forward pass, without convergence checks or candidate verification.

CARD and WeDLM train diffusion language models with causal attention, and their commit rules allow a later position to be filled before an earlier one. In contrast, PARD keeps the bidirectional intra-block attention of existing BDLMs. It also requires a contiguous prefix from the leftmost mask, which keeps the decoding contexts close to those seen in block diffusion training (Section 3.2).

We have added the discussion of these related works to Section 2.2 of the revised paper.

W3. WeDLM's prefix-cacheability motivation.

Unlike WeDLM, PARD is not motivated by prefix cacheability. Prefix cacheability requires causal attention, where a committed left-to-right prefix can be cached immediately. In BDLMs, attention within a block is bidirectional, so committed tokens are not cache-valid until the whole block is decoded, regardless of the decoding order. In contrast, PARD's left-to-right rule is motivated by the BDLM's AR bias (Section 3). One may ask why not switch a trained BDLM to causal attention to gain prefix cacheability without training. We test this on SDAR 8B, using PARD as the sampler. Changing the attention to causal hurts performance:

AttentionHumanEvalGSM8K
Block-bidirectional (original)81.191.4
Change to causal75.687.0

Switching to causal attention deviates from the attention pattern used in training. Prefix cacheability would therefore require retraining the model, as in WeDLM.

Q3. Positioning relative to CLLMs, Jacobi Forcing, CARD, and WeDLM.

Positioning. CLLMs and Jacobi Forcing accelerate AR models losslessly, as their output equals greedy AR decoding. CARD and WeDLM are closest to PARD, as they also commit multiple tokens per step based on confidence. However, all four methods train a new or adapted model with causal attention. PARD instead addresses a different setting: improve training-free sampling of existing BDLMs, which use bidirectional intra-block attention.

Contributions specific to PARD. The causal parallel decoding methods above build left-to-right advancement into the model, by training AR models or causal-attention diffusion models. In contrast, we show that existing BDLMs, trained with bidirectional intra-block attention, already prefer left-to-right decoding, and we explain why in Section 3. PARD exploits this AR bias at inference time only, so it is training-free and plug-and-play. We have added the discussion in Section 2.2 of the revised paper.

I thank the authors for the detailed rebuttal and the additional experiments (the causal-attention ablation, the GSM8K AR-bias analysis, and the comparison with learned unmasking policies). These address several of my concerns. I want to be clear that my rating is not a soundness judgment: the experiments are carefully executed, and the empirical improvements on the tested checkpoints are real. My concern is that the central contribution is insufficiently novel, and the central insight insufficiently general, to meet the acceptance bar. I therefore maintain my rating.

  1. The methodological delta is negligible. The authors acknowledge that PAUD-Confidence uses the same position-selection rule as DMax (Algorithm 1, lines 5-8). I understand the ICLR policy on arXiv-only prior work, and I am not rejecting the paper for failing to cite or compare with DMax at submission time. The exemption, however, concerns missing comparisons; it does not resolve the novelty question once the overlap is acknowledged. Given that the selection rule is identical, the remaining delta is (i) applying it training-free to off-the-shelf BDLMs, (ii) the margin/entropy variants, and (iii) the BDLM AR-bias analysis. Item (i) is a natural simplification of DMax (discrete commitment in place of soft-embedding feedback), and (ii) is a routine extension of the scoring rule.

More importantly, the prescription itself - restricting confidence-based decoding to autoregressive/prefix order to avoid the pathologies of arbitrary-order commitment - is not new. The underlying problem (early commitment of high-confidence, low-entropy tokens degrades subsequent generation quality and diversity) and the order-restriction remedy are already established by Ni et al. (2026) and the broader line of work on diversity loss in confidence-based decoding. The revised manuscript cites Ni et al. only for the pass@k exploration finding and does not engage with this mechanism: the words "collapse" and "diversity" do not appear anywhere in the paper. A method that repackages an established remedy, without positioning against the problem it is known to solve, leaves its methodological contribution unclear.

  1. The central insight does not generalize. The paper presents the strong AR bias as a property of BDLMs, explained by their training-context distribution (Propositions 1-2). But this is a property of the specific, under-adapted checkpoints studied, not of block diffusion as a paradigm. Clover-35B-A3B-Preview (DiffuSpace, 2026), a recently released BDLM with substantial post-training, uses out-of-order confidence-based decoding within blocks (the Fast-dLLM rule, not a leftmost prefix) and reports strong agentic quality and high throughput. While these numbers are vendor-reported and the training differs, they are direct evidence that a well-trained BDLM does not require the AR-order prescription. The propositions are elementary, and the cross-model attribution is confounded by differences in initialization and adaptation budget (Section 3.2 acknowledges this but provides no controlled experiment). The insight is therefore narrow in scope.

  2. The "reality-check" significance is not the paper's own claim. Part of the positive assessment from other reviewers rests on reading the paper as a reality check on DLM speed claims relative to AR speculative decoding. The original submission does not articulate this message. The revised Appendix E now acknowledges that "current DLMs do not show a clear advantage over strong AR models with speculative decoding," but frames it defensively as motivation for the work rather than as a finding. A paper should be assessed on the contribution it actually claims and substantiates, not on a significance that reviewers infer.

In summary, the empirical study is sound and the sampler works on the tested checkpoints, but the selection rule is inherited from DMax, the order-restriction prescription is an established remedy for a known problem the paper does not discuss, and the AR-bias insight does not generalize beyond the specific checkpoints studied. On novelty and significance grounds, I maintain my rating of 1.

Comments on the paper

4 comments

I would like to engage with Reviewer bKFr's W1 (the comparison with DFlash). I agree with the qualitative message of that concern, but I believe the quantitative comparison is confounded, and one factual detail should be corrected.

  1. Measurement settings on both sides. A small factual correction: DFlash's experiments are not run on H100. The main results use H200 GPUs with a Transformers backend, and the serving experiments (DFlash, Section 5.3) use a single B200 with SGLang (FlashAttention-4 and Spec-v2 scheduling overlap), reporting 1175 tok/s for Qwen3-8B on MATH-500 at concurrency 1, versus 230 tok/s for the AR baseline. In contrast, the present submission measures throughput on a single RTX A6000 with HuggingFace Transformers, without model compilation, at batch size 4 (Section 5 and Appendix C.1). The ">900 vs. 100-200 tok/s" comparison therefore mixes at least three variables: (i) hardware generation (Ampere vs. Hopper/Blackwell, with very different memory bandwidth); (ii) inference stack (uncompiled HF Transformers vs. a serving engine optimized end-to-end for speculative decoding); and (iii) method setting (a training-free sampler on off-the-shelf BDLMs vs. a drafter trained per target model on ~800K prompts with lossless verification), not to mention batch size 4 vs. concurrency 1. An order-of-magnitude gap cannot be attributed to the decoding paradigm on this basis.

  2. BDLM throughput is not intrinsically capped at 100-200 tok/s. As an existence proof, Clover-35B-A3B-Preview (DiffuSpace, Oct 2026), a 35B MoE block-diffusion model with ~3B active parameters, reports 680 tok/s on a single GB300 (BF16, batch=1, internal agentic workload) and 515 tok/s on H200 agentic tasks, compared in the same report with 365 tok/s for a same-size Qwen model accelerated by DFlash2. Its official SGLang implementation supports both block-diffusion denoising and self-speculative decoding. Caveats: these numbers are vendor-reported at batch=1 without third-party reproduction, and part of the gain comes from the MoE design (block-level expert reuse amortizes weight reads), so this is not a controlled counter-comparison. It nonetheless shows that with an optimized serving stack and a suitable architecture, BDLM decoding can reach the several-hundred tok/s range; the 100-200 tok/s measured here reflects the unoptimized stack as much as the paradigm.

  3. What I agree with. Lossless speculative decoding of strong AR models currently dominates standalone BDLM decoding in the speed-quality trade-off, and the revised Appendix E acknowledges this. I would only sharpen the conclusion: DFlash itself uses a BDLM drafter, so the evidence speaks against standalone BDLM decoding rather than against block diffusion as a component. Its parallelism is currently best harvested inside a speculative pipeline, or with serving-optimized architectures such as Clover.

Suggestion to the authors: Appendix E would be stronger if it noted that cross-paper throughput comparisons are confounded by hardware and backend, and that a like-for-like measurement (same GPU, same serving stack) is needed to quantify the gap to AR speculative decoding. The current text acknowledges the gap but does not analyze how much of the reported numbers is attributable to the measurement setup.

  1. Indeed, there are several confounders (which are transparent now since they reported on the used hardware in the rebuttal). However, this is exactly why I asked them to clarify and compare empirically against D-Flash. The burden of proof is on the authors to reproduce credible baselines with their hardware and setup. Regarding the H100 vs H200 vs B200 discussion, there have since been several reproductions of D-Flash, including inference on H100s, to which I referred.

  2. I never insinuated that BDLM throughput is "intrinsically capped at 100-200 tok/s", that would be nonsense. Again, burden of proof is on the authors to provide a transparent and fair comparison. Vendor-reported numbers are not scientifically sound, given the conflict of interest. There are many ways to set up the experiment to "hack the numbers".

  3. Calling the DFlash drafter a BDLM is a bit of an overstretch. There is no diffusion objective in training due to its single-step nature. Yes, it uses bidirectional attention and masks, but so does BERT etc. While the architectural choices in D-Flash are undoubtedly inspired by BDLMs, calling it a BDLM is inaccurate. In particular, the question raised in the paper under discussion (PARD/PAUD) is inapplicable to D-Flash since everything is unmasked at once.

Thank you both for the discussion.

For completeness, we have now added a direct empirical comparison with DFlash in our reply to Reviewer bKFr's thread, as well as in Appendix E of the revised paper.

Specifically, we compare SDAR 8B (with PARD and other decoding strategies) against Qwen3-8B with DFlash on the same RTX A6000 GPU, using both HuggingFace Transformers and SGLang backends, at batch size 1. We also report generation accuracy.

The results show that DFlash achieves higher throughput than standalone BDLM inference with PARD, indicating that there is still room for improvement. Nevertheless, PARD already makes SDAR 8B 1.7–2.1× faster than Qwen3-8B AR under SGLang, with competitive accuracy.

We also agree with Reviewer bKFr that calling the DFlash drafter a BDLM is not entirely precise. DFlash is not trained with a block diffusion objective that averages over different diffusion timesteps. Instead, it is trained to predict the whole masked block in a single step, so the question of decoding order within a block does not arise. Moreover, the DFlash drafter relies on context features from the target AR model and cannot generate independently as a standalone BDLM.

Thus, while DFlash demonstrates the effectiveness of parallel block prediction within speculative decoding, our work focuses on improving standalone BDLM inference and provides a diagnostic foundation for future BDLM development.

Dear reviewers,

The author-reviewer discussion period is ending soon. If you have not done so yet, please read the authors’ rebuttal and reply to it. Tell the authors whether it addresses your concerns, and update your score if your assessment has changed.

The ICLR 2027 Reviewer Guidelines list « discussing with the authors to clarify and improve the paper » as one of the core duties you accepted as a reviewer. With the high number of submissions this year, your input matters more than ever. Recall that reviewers that do not fulfill their duties may also face penalties on their own submissions.

Thank you for your time and help.

Best regards,


Your Area Chair

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.