acceptodds
Under review as a conference paper at ICLR 2027

Streaming-dLLM: Accelerating Diffusion LLMs via Suffix Pruning and Dynamic Decoding

Abstract

Diffusion Large Language Models (dLLMs) offer a compelling paradigm for natural language generation, leveraging parallel decoding and bidirectional attention to achieve superior global coherence compared to autoregressive models. While recent works have accelerated inference via KV cache reuse or heuristic decoding, they overlook the intrinsic inefficiencies within the block-wise diffusion process. Specifically, these methods suffer from spatial redundancy by modeling sparse suffix regions uniformly and temporal inefficiency by applying fixed denoising schedules across decoding. To address this, we propose Streaming-dLLM, a training-free framework that streamlines inference across both spatial and temporal dimensions. Spatially, we introduce approximated suffix pruning that leverages context modeling redundancy to compress suffix cost from linear to constant complexity, decoupling inference cost from sequence length. Temporally, we employ a dynamic confidence aware strategy that adaptively modulates denoising thresholds based on global mask state evolution, allowing the model to skip unnecessary iterations for converged tokens. Experiments on five text-only dLLMs and one diffusion multimodal large language model (dMLLM) across six benchmarks show that Streaming-dLLM matches or improves the accuracy of existing acceleration methods while achieving speedups of up to over these methods. For long sequence generation, it outperforms the strongest baseline by 2.5 accuracy points with higher throughput.

Primary area:
foundation or frontier models, including LLMs
Keywords:
Diffusion Large Language Models, Inference Acceleration

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_HS6aOpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
3
Soundness
2
Read review
Paper summary

The paper proposes a training-free acceleration framework for block-wise diffusion decoding. Approximated suffix pruning keeps only a sliding window of width ww after the current block plus the terminal position of the whole sequence, preserving order through RoPE position IDs. Dynamic confidence aware parallel decoding lowers the acceptance threshold with the mask ratio of the query region, τ(t)=τ0(1−α(1−rmask))\tau^{(t)}=\tau_0(1-\alpha(1-r_{\mathrm{mask}})). An early exit skips all remaining blocks once EOS is predicted. Across five text dLLMs and one multimodal dLLM on six benchmarks, average speedups over vanilla are 9.3-19.8x, and on GSM8K at length 2048 the method reports 2.5 accuracy points over the strongest baseline with 4.7x higher throughput.

Critical strengths

Empirical coverage is broad: three main backbones at two generation lengths across four benchmarks, plus two block-diffusion models and one multimodal model, with both throughput and per-sample latency. The method is training-free and easy to drop into existing stacks. Retaining the terminal position has clear empirical support.

Critical weaknesses

Once prior work is subtracted, almost nothing new remains, least of all the speedup. Tables 5 and 7 decompose the 28.0x over vanilla into 3.28x from prefix caching, 1.84x from early exit, 3.74x from threshold parallel acceptance and 1.24x from suffix pruning; the first three multiply to 22.6x and are all prior work, leaving the α\alpha schedule's 1.21x as this paper's own contribution, while the genuinely new part of ASP, the terminal anchor, retains one extra token and therefore costs rather than saves time. On the accuracy side, the largest jump is already present in the DPad row.

  1. Both claimed components overlap heavily with existing methods. DPad already provides a fixed-length suffix sliding window, a RoPE remapping that keeps retained tokens at their original absolute positions, and EOS early termination. Of DCAD's four ingredients, prefix KV reuse within a block and confidence-threshold acceptance with its fallback (τ0\tau_0 fixed at Fast-dLLM's default 0.9) are Fast-dLLM's and the query construction follows from ASP, leaving only the linear decrease of the threshold with mask ratio as new. Section 2.2 states that DPad "may lose positional information in the suffix region", but DPad remaps RoPE so each retained token keeps its original absolute position, so retained tokens lose no positional information. The one real difference is that DPad retains stochastically within the window (20-37.5% density) with no guarantee of keeping the final position of the whole sequence, which necessarily falls outside the window once the sequence is longer than it, whereas ASP keeps that position as a deterministic anchor. Early exit is introduced and ablated as one of three framework components without attribution to DPad, and EOS-based early stopping is common practice.
  2. The 3.7x over DPad cannot be explained by suffix cost. On LLaDA-1.5/GSM8K-512 the method deterministically retains 129 suffix tokens at w=128w=128, whereas DPad retains about 64 in expectation (window 256 at 25% density). The method keeps more yet reports 3.7x higher throughput (69.8 vs 19.0 TPS), so the speedup cannot come from a cheaper suffix. Decomposed through Table 5, the 3.67x factors into about 1.48x from fixed-threshold parallel acceptance, 1.21x from the α\alpha schedule, and about 2.05x left unexplained on the DPad side, the last measured against Table 5's "prefix cache + suffix pruning + exit" configuration at 38.9 TPS versus DPad's 19.0 TPS. That 2.05x can only come from implementation overhead (non-contiguous gather, per-block resampling, per-token RoPE remapping) or from the DPad row running without prefix caching, and the paper measures and discloses neither.
  3. The accuracy gain is misattributed and unexplained. The appendix claims 13.8-27.0 points over Fast-dLLM on MBPP, attributed to suppressing distant high-entropy masks, but in the LLaDA-Instruct MBPP-512 column Fast-dLLM scores 13.8, DPad 39.4 and the method 40.8, so 25.6 of the 27.0 points are already in DPad and this paper adds 1.4. Vanilla's 14.8 looks like non-termination or formatting failure, yet no termination, truncation or format-error rate is reported. The appendix elsewhere states that with appropriate hyperparameter configurations the method can recover or surpass vanilla accuracy.
  4. No isolation study, and asymmetric tuning. The Fixed and Ours columns of Table 16 are numerically identical to the main table, so no experiment changes only the acceptance rule with everything else fixed; the sole isolated figure for the α\alpha schedule is 1.21x in Figure 4(b). The paper's ww and α\alpha are set per model, benchmark and length while DPad runs at defaults. Table 20's three values coincide with the main table's GSM8K-512 cells, so the terminal anchor is tested at one length only, even though the appendix itself notes that retaining it makes behavior depend on overall generation length.
Questions for authors
  1. On LLaDA-1.5/GSM8K-512 the method retains 129 suffix tokens at w=128w=128 while DPad retains about 64 in expectation (window 256 at 25% density). The method keeps more yet is 3.7x faster, so where does that speedup come from? Please give step-level time breakdowns for both sides.
  2. For fairness, could you disclose your experimental configurations for Fast-dLLM and DPad?
  3. Tables 5 and 7 decompose the 28.0x over vanilla into 3.28x (prefix cache), 1.84x (early exit), 3.74x (threshold parallel acceptance) and 1.24x (suffix pruning), with the first three multiplying to 22.6x of prior work. Please confirm this decomposition and state what share of the total speedup comes from this paper's own additions. With prefix caching and EOS early termination given to all baselines, what do the main-table multiples become?
  4. Please decompose the 27.0 points on MBPP into early termination, suffix pruning and this paper's increment, report vanilla's termination, truncation and format-error rates in that cell, and repeat the trailing-position ablation separately at 512, 1024 and 2048.
7 comments
AuthorsOpenReview ↗

Response to Reviewer HS6a (Part 1/4)

Dear Reviewer HS6a,

We sincerely thank the reviewer for recognizing the breadth of our evaluation, the ease of integrating our training-free method into existing pipelines. Your feedback helped us make the manuscript clearer and more complete, and we have marked the corresponding changes in orange. We appreciate the constructive feedback and respond below.


Q1+Q2+W2: What were the Fast-dLLM and DPad settings? Why is Streaming-dLLM 3.7× faster on LLaDA-1.5/GSM8K-512 despite retaining 129 suffix tokens (w=128) vs. DPad’s ~64 (256×25%)?

  • Baseline implementations and configurations. We used the official Fast-dLLM and DPad code. The DPad implementation returned all-zero results for Dream on MBPP (an unresolved issue in its GitHub repository), so we ported the relevant code to the official Fast-dLLM codebase and adapted it for our evaluation. For Fast-dLLM, we used the best settings in its main results table, with prefix caching and fixed-threshold decoding (temperature 0, threshold 0.9). Table 9 of the DPad paper reports evaluation hyperparameters for only four datasets, each at a single generation length (GSM8K-256, MATH-256, HumanEval-512, MBPP-512). Moreover, DPad specifies k, a, and the window size separately according to the model, benchmark, and generation length. We therefore adjusted the window size for tasks with different generation lengths and generally reused DPad’s k and a values for similar datasets. Hyperparameter values changed during tuning are shown in bold in Table 1. We also ran a small hyperparameter comparison to illustrate the limitations of transferring the reported settings directly (Table 2).
  • Additional details on the DPad baseline. In Table 1 of our paper, we report DPad with fixed-threshold parallel decoding (Par) and early exit (Exit). For the long-generation experiments in Table 2 of our paper, we use prefix cache (PC) + Par + Exit, consistent with the configuration in the original DPad paper. To facilitate comparison, we also added DPad (PC + Par + Exit) as a baseline. On LLaDA-1.5, Streaming-dLLM achieves higher throughput while matching or slightly improving on DPad’s accuracy. On MBPP, it achieves 2.0× the throughput of DPad (PC+Par+Exit) (Table 3). On GSM8K with a generation limit of 2,048 tokens, it achieves 4.7× the throughput of that baseline and increases accuracy from 79.5% to 82.0% (Table 4).
  • Component-wise comparison of Streaming-dLLM and DPad. As shown in Table 5, under the same Par+PC+Exit configuration on GSM8K, Streaming-dLLM’s suffix approximation (ASP) achieves 1.23× and 2.36× DPad’s throughput at generation budgets of 512 and 2,048 tokens, respectively. Enabling dynamic-threshold decoding (DCAD) alone also yields higher throughput than DPad. With both ASP and DCAD enabled, Streaming-dLLM matches DPad’s accuracy and achieves 1.49× and 4.67× its throughput at the two budgets, respectively. The two paired comparisons show, respectively, the spatial gains from ASP over DPad under matched settings and the additional gains from DCAD over Par.
  • Suffix tokens retained by DPad and ASP. We measured the number of suffix tokens retained by each pruning strategy on LLaDA-1.5 (Table 1). Compared with DPad, Streaming-dLLM retains 31.88 fewer suffix tokens on average at a generation budget of 512 tokens and 185.42 fewer on GSM8K at a budget of 2,048 tokens. These results provide further evidence that our suffix approximation preserves positional information in the suffix more effectively than DPad while retaining fewer tokens.
  • The revision adds DPad (PC+Par+Exit) to main-text Table 1 and component ablations to Table 26 (Appendix B.12).

Table 1: Suffix retention statistics for the DPad and Streaming-dLLM pruning strategies on LLaDA-1.5. The parameters k, a, and window size are DPad hyperparameters.

kawindow sizeDPadStreaming-dLLM
GSM8K-5123.01.6512145.97129
GSM8K-20483.01.61024282.4297
Math-5123.02.3512206.34193
HumanEval-5123.01.6512145.5897
MBPP-5123.01.6512145.6197

Table 2: DPad Tuning Comparison.

kawindow sizeDPad Accuracy
GSM8K-5123.01.625679.9
GSM8K-5123.01.651280.9

Table 3: Additional baseline comparisons on LLaDA-1.5, Each entry reports Accuracy/Throughput.

HumanEvalGSM8KMBPPMATHAverage
+Dpad (PC+Par+Exit)39.1/16.680.9/47.038.2/30.734.6/45.448.2/34.9
+Streaming-dLLM40.2/26.781.2/69.838.4/61.435.1/62.448.7/55.1

Table 4: Additional baseline comparisons on LLaDA-1.5 for long-sequence tasks.

GSM8K-2048
+Dpad (PC+Par+Exit)79.5/14.5
+Streaming-dLLM82.0/67.6

Table 5: Component ablations on LLaDA-1.5. All rows enable Par+PC+Exit.

GSM8K-512GSM8K-2048
Baseline(Par+PC+Exit)80.4/31.880.7/8.4
Baseline+Dpad80.9/47.079.5/14.5
Baseline+ASP81.2/57.782.0/34.2
Baseline+DCAD81.0/56.481.8/22.5
Baseline+DCAD+ASP81.2/69.882.0/67.6
AuthorsOpenReview ↗

Response to Reviewer HS6a (Part 2/4)

Q3: Tables 5 and 7 decompose the 28.0x over vanilla into 3.28x (prefix cache), 1.84x (early exit), 3.74x (threshold parallel acceptance) and 1.24x (suffix pruning), with the first three multiplying to 22.6x of prior work. Please confirm this decomposition and state what share of the total speedup comes from this paper's own additions. With prefix caching and EOS early termination given to all baselines, what do the main-table multiples become?

  • In Table 5 of our paper, prefix cache (PC) achieves 3.28× throughput of Vanilla, and PC + Exit achieves 1.84× throughput of PC alone. PC + DCAD + Exit achieves 3.74× throughput of PC + Exit, and PC + DCAD + Exit + ASP achieves 1.24× throughput of PC + DCAD + Exit. PC + DCAD + Exit achieves 22.6× the throughput of Vanilla. Attributing the first three gains to prior work overlooks our contribution in the third comparison. DCAD is our proposed dynamic-threshold decoding strategy, denoted “Dyn.” in Table 5 of the original manuscript. In Table 1 of our paper, Fast-dLLM uses PC and fixed-threshold parallel decoding (Par). On GSM8K (5-shot) with LLaDA-1.5, Streaming-dLLM improves accuracy by 0.8 percentage points over Fast-dLLM and achieves 2.7× its throughput. These gains reflect the combined effects of our suffix pruning, dynamic-threshold decoding, and Exit.
  • Additional baseline with EOS early termination (Exit). In our response to Q1+Q2+W2, we added DPad (Par + PC + Exit). All baselines now use KV cache. We also adapted Fast-dLLM, a representative baseline, to use Exit. As shown in Table 6, Streaming-dLLM improves accuracy by 0.8 percentage points over Fast-dLLM (Exit) and achieves 2.2× its throughput. These gains reflect the contributions of our suffix pruning and dynamic-threshold decoding.
  • In the revised manuscript, we have added Fast-dLLM (Exit) to Table 1 in the main text.

Table 6: Additional baseline comparisons on LLaDA-1.5 in the main results table,Each entry reports Accuracy / Throughput.

HumanEvalGSM8KMBPPMATHAverage
+Fast-dLLM (Exit)39.8/14.180.4/31.838.1/21.335.1/40.348.4/26.9
+Streaming-dLLM40.2/26.781.2/69.838.4/61.435.1/62.448.7/55.1

Q4+W3: Please decompose the 27.0 points on MBPP into early termination, suffix pruning and this paper's increment, report vanilla's termination, truncation and format-error rates in that cell, and repeat the trailing-position ablation separately at 512, 1024 and 2048. The accuracy gain is misattributed and unexplained.

  • Clarification on the role of suffix pruning in the MBPP accuracy recovery. DPad and Streaming-dLLM’s ASP use different suffix pruning strategies, and both recover accuracy on MBPP. A plausible explanation is that suffix pruning reduces interference from distant mask tokens. We further evaluated different combinations of components on MBPP. As shown in Table 7, Exit alone yields no accuracy gain, indicating that the improvement comes from suffix pruning. In each MBPP demonstration, Python source code immediately follows [BEGIN], and [DONE] marks the end. The demonstrations thus provide the primary cue for the code-only output format. Our analysis of the generated outputs showed that when the full suffix was retained, the model generated additional material after the solution code, including explanations, tests, and repeated code, before emitting [DONE]. Under strict matching, this extra material causes compilation failures. With lenient matching, accuracy in the full-suffix setting rises from 13.8% to 36.8%. The ablations support our original attribution of the MBPP accuracy recovery to suffix pruning. The output analysis and the recovery under lenient matching are consistent with our hypothesis that distant mask tokens introduce interference and that pruning them helps the model follow the code-only format required by strict matching.
  • Additional experiments on the terminal anchor. We ablated the terminal anchor at different output lengths. As shown in Table 8, retaining the terminal anchor consistently improves performance across the tested lengths, providing further evidence of Streaming-dLLM’s effectiveness.
  • In the revised manuscript, we added Table 25 to Appendix B.11 to demonstrate that the terminal anchor design generalizes across output lengths. We also expanded the analysis of the source of the MBPP accuracy gains in Appendix B.1.

Table 7: MBPP-512 ablation on LLaDA-Instruct.

MBPP strict match resultMBPP flexible match result
Baseline (Par+PC)13.836.8
Baseline+Exit13.836.8
Baseline+DPad39.439.4
Baseline+ASP40.840.8
Baseline+DCAD13.836.8
Baseline+ASP+DCAD40.840.8

Table 8: Effect of retaining the trailing positional token at different generation length.

Trailing PositionAccuracy
GSM8K,len=512×79.6
GSM8K,len=512√81.2
GSM8K,len=1024×78.7
GSM8K,len=1024√81.2
GSM8K,len=2048×79.8
GSM8K,len=2048√82.0
AuthorsOpenReview ↗

Response to Reviewer HS6a (Part 3/4)

W1: Both claimed components overlap heavily with existing methods. DPad already provides a fixed-length suffix sliding window, a RoPE remapping that keeps retained tokens at their original absolute positions, and EOS early termination. Of DCAD's four ingredients, prefix KV reuse within a block and confidence-threshold acceptance with its fallback (fixed at Fast-dLLM's default 0.9) are Fast-dLLM's and the query construction follows from ASP, leaving only the linear decrease of the threshold with mask ratio as new. Section 2.2 states that DPad "may lose positional information in the suffix region", but DPad remaps RoPE so each retained token keeps its original absolute position, so retained tokens lose no positional information. The one real difference is that DPad retains stochastically within the window (20-37.5% density) with no guarantee of keeping the final position of the whole sequence, which necessarily falls outside the window once the sequence is longer than it, whereas ASP keeps that position as a deterministic anchor. Early exit is introduced and ablated as one of three framework components without attribution to DPad, and EOS-based early stopping is common practice.

  • We would like to clarify how DPad and Streaming-dLLM differ in the types of decoding redundancy they address and their applicability to pure block diffusion language models.

    • Streaming-dLLM reduces temporal and spatial redundancy in dLLMs through dynamic threshold decoding and approximated suffix pruning, respectively. DPad targets spatial redundancy alone. Temporally, dynamic threshold decoding uses the mask ratio as a proxy for uncertainty at the current decoding stage, allowing the threshold to adapt to the overall decoding state. This signal complements individual token confidence scores and enables higher throughput. Spatially, we observe that attention concentrates on nearby suffix positions and the end of the sequence, while intermediate suffix positions receive less attention and are still processed at every step. Based on this observation, ASP retains a contiguous window immediately after the current block and the terminal position of the sequence. DPad samples within a candidate window using a distance-decay rule, retaining some redundant intermediate positions while omitting the terminal position. Moreover, we measured how many suffix tokens each strategy actually retains on LLaDA-1.5 (Table 1). Streaming-dLLM retains 31.88 fewer suffix tokens on average in the experiments with a 512 token generation budget and 185.42 fewer on GSM8K with a 2,048 token budget.

    • DPad is limited to full-sequence dLLMs. Streaming-dLLM extends to the latest pure block-diffusion language models and achieves 2.6–5.4× speedups. Pure block diffusion is an increasingly important direction for dLLMs. These models support early exit and variable-length generation, and their masked generation region is small, typically just 32 tokens. DPad prunes the masked suffix beyond the current block in full-sequence diffusion language models. Pure block diffusion confines masking to the current block, leaving no cross-block suffix for DPad to prune. Streaming-dLLM can improve the performance of pure block diffusion models along both the spatial and temporal dimensions even with short suffixes. Spatially, we divide the current diffusion block into smaller sub-blocks, separate the sub-block being denoised from the subsequent masked sub-blocks, and apply approximated suffix pruning to those subsequent sub-blocks to reduce computation. Temporally, Streaming-dLLM adjusts the degree of decoding parallelism according to the proportion of masked tokens remaining in the block. Reducing the masked generation region further increases the benefit of the dynamic threshold. Table 3 of the original submission evaluates this extension on two pure block dLLMs. At comparable accuracy, Streaming-dLLM achieves 2.6–5.4× the throughput of the +Fast-dLLM baseline, which already uses prefix cache, fixed-threshold parallel decoding, and early stopping.

  • Streaming-dLLM’s contributions. As stated in the main text, we identify spatial and temporal redundancy inherent in dLLMs and design approximated suffix pruning (ASP) and dynamic confidence-aware parallel decoding (DCAD) to eliminate redundant computation along both dimensions. Early exit serves only as an auxiliary component. In this rebuttal, we also report results for additional baselines with EOS-based early exit enabled to provide a more complete comparison.

Reviewer_x2cEOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

Streaming-dLLM is a training-free framework for accelerating block-wise diffusion language decoding. It retains a nearby suffix window and the token at the original final position, it also adapts confidence thresholds to the remaining mask ratio, reuses prefix KV states, and terminates at EOS. Experiments cover five text models and one multimodal model across six tasks. On LLaDA-1.5/GSM8K with a 2048-token generation cap, the paper reports 82.0% accuracy and 67.6 tokens/s, compared with 79.5% and 14.5 tokens/s for DPad. The additional Fast-dLLM v2 comparison reports 82.0% and 22.1 tokens/s.

Critical strengths
  1. Substantial practical acceleration over already accelerated baselines. The evaluation covers several backbones and tasks. At the 2048-token generation cap on GSM8K, the reported throughput is approximately 4.7× that of DPad. Appendix Table 17 also includes Fast-dLLM v2, against which the improvement remains approximately 3.06× at the same reported 82.0% accuracy. These results support the practical value of the integrated implementation.
  2. The analysis includes useful component and stability checks. Table 5 evaluates all combinations of three components, and Table 20 investigates the trailing positional token. Appendices additionally report latency for the proposed method. This provides more informative evidence than a single aggregate speedup.
Critical weaknesses
  1. The incremental contribution over DPad is insufficiently established. DPad already includes a fixed suffix window, original RoPE positions, and EOS termination [1], contrary to aspects of Section 2.2. The specific additions are terminal-position retention and a mask-dependent threshold, whose separate benefits are not isolated against a matched DPad configuration in Table 5. Retaining the terminal position should be distinguished from preserving existing tokens’ positions.
  2. The accuracy throughput trade-off and long-output claims need clarification. Table 12 lacks a held-out tuning protocol, while Table 18 omits the configurations behind alternative results, including HumanEval-256 accuracy increasing from 39.0% to 44.5%; the standard setting trails vanilla by 4.9 points. Attention still reads all prefix KV states, so bounded suffix size does not imply length-independent total cost.

[1] Xinhua Chen, Sitao Huang, Cong Guo, Chiyue Wei, Yintao He, Jianyi Zhang, Hai “Hellen” Li, and Yiran Chen. DPad: Efficient Diffusion Language Models with Suffix Dropout. arXiv preprint arXiv:2508.14148, 2025.

Questions for authors
  1. Under identical DPad implementation, window, positional handling, prefix-cache, and EOS settings, what are the separate accuracy and throughput gains from the trailing positional token and dynamic threshold?
  2. What data and selection procedure were used to choose the hyperparameters in Table 12, and what complete configurations produce Table 18?
  3. If the 2048-token budget is exhausted without EOS, does the implementation truncate or append masks to continue? If continuation is supported, how are the terminal position and KV cache updated?
6 comments
AuthorsOpenReview ↗

Response to Reviewer x2cE (Part 1/3)

Dear Reviewer x2cE,

We sincerely thank the reviewer for recognizing our speedups over accelerated baselines and the value of our ablation and stability analyses. Your feedback help us make the manuscript clearer and more complete, and we have marked the corresponding changes in orange. We appreciate the constructive feedback and respond below.


W1+Q1: The incremental contribution over DPad is insufficiently established. DPad already includes a fixed suffix window, original RoPE positions, and EOS termination, contrary to aspects of Section 2.2. The specific additions are terminal-position retention and a mask-dependent threshold, whose separate benefits are not isolated against a matched DPad configuration in Table 5. Retaining the terminal position should be distinguished from preserving existing tokens’ positions. Under identical DPad implementation, window, positional handling, prefix-cache, and EOS settings, what are the separate accuracy and throughput gains from the trailing positional token and dynamic threshold?

  • Thank you for pointing out that Section 2.2 should describe DPad’s window settings, positional encoding, and EOS termination more precisely. We have updated the wording in the revised manuscript: For model pruning, DPad uses distance decay sampling within a candidate window. The sampling parameters and window width are selected for the dataset, backbone, and maximum generation length. It preserves retained tokens’ original RoPE positions and supports EOS termination. However, sampled subset may retain redundant intermediate positions and omit the terminal masked token that indicates the extent of the original suffix.
  • Additional DPad baseline comparison and Streaming-dLLM component ablations. In the initial submission, we reported DPad with fixed-threshold parallel decoding (Par) and early exit (Exit) in the main table and used prefix caching (PC)+Par+Exit in the long-generation experiments, consistent with the configurations reported in the original DPad paper. We added DPad (PC+Par+Exit) as a baseline to the main results table. On LLaDA-1.5, Streaming-dLLM achieves higher throughput at comparable or slightly higher accuracy. On MBPP, it achieves 2.0× the throughput of DPad (PC+Par+Exit) (Table 1). On GSM8K with a maximum generation length of 2,048 tokens, Streaming-dLLM achieves 4.7× the throughput of the fastest baseline, DPad (PC + Par + Exit), while improving accuracy from 79.5% to 82.0% (Table 2). Table 2 also shows that under the same configuration (Par + PC + Exit), Streaming-dLLM’s approximate suffix pruning (ASP) achieves 1.2× and 2.4× DPad’s throughput on GSM8K at generation budgets of 512 and 2,048 tokens, respectively. Dynamic threshold decoding (DCAD) alone also achieves higher throughput than DPad. Enabling both ASP and DCAD maintains comparable accuracy and achieves 1.5× and 4.7× DPad’s throughput, respectively. These two paired comparisons show the spatial gains of Streaming-dLLM’s suffix approximation over DPad under the same settings and the additional gains from dynamic thresholding relative to fixed-threshold decoding.
  • DPad and Streaming-dLLM’s ASP use distinct suffix pruning strategies that differ in both their token retention budgets and the positional distribution of the retained tokens. Their treatment of the terminal position is one of several differences. Please see the final section of our response for a detailed comparison of Streaming-dLLM and DPad.
  • In the revised manuscript, we updated the description of DPad in Section 2.2 and added the new DPad (PC + Par + Exit) baseline to the main results table. We also add Table 26 with further component analysis in the new Appendix Section B.12.

Table 1: Additional baseline comparisons on LLaDA-1.5 in the main results table,Each entry reports Accuracy / Throughput.

HumanEvalGSM8KMBPPMATHAverage
+ Dpad (PC+Par+Exit)39.1/16.680.9/47.038.2/30.734.6/45.448.2/34.9
+Streaming-dLLM40.2/26.781.2/69.838.4/61.435.1/62.448.7/55.1

Table 2: Component ablations on LLaDA-1.5. All configurations enable Prefix Cache (PC), Exit and fixed-threshold parallel decoding (Par). Each entry reports Accuracy / Throughput.

GSM8K-512GSM8K-2048
Baseline (Par+PC+Exit)80.4/31.880.7/8.4
Baseline+Dpad80.9/47.079.5/14.5
Baseline+ASP81.2/57.782.0/34.2
Baseline+DCAD81.0/56.481.8/22.5
Baseline+DCAD+ASP81.2/69.882.0/67.6
AuthorsOpenReview ↗

Response to Reviewer x2cE (Part 2/3)

W2+Q2: The accuracy throughput trade-off and long-output claims need clarification. Table 12 lacks a held-out tuning protocol, while Table 18 omits the configurations behind alternative results, including HumanEval-256 accuracy increasing from 39.0% to 44.5%; the standard setting trails vanilla by 4.9 points. Attention still reads all prefix KV states, so bounded suffix size does not imply length-independent total cost. What data and selection procedure were used to choose the hyperparameters in Table 12, and what complete configurations produce Table 18?

  • Tuning protocol. We selected the configurations in Table 12 based on the evaluation results for each of HumanEval, GSM8K, MBPP, and MATH. We used the same evaluation split for hyperparameter selection and for reporting results. We fixed the block size at 32 tokens and the initial confidence threshold at 0.9. Starting from a 96 token suffix window and a parallelism setting of α = 0.3, we made small adjustments to the window size and α, guided by the sensitivity analysis in Figure 4. We prioritized configurations with accuracy comparable to that of existing acceleration baselines and relatively high throughput. In the GSM8K experiments with LLaDA-1.5, accuracy gains began to level off at a window size of 128 tokens, and increasing the window further reduced throughput. Increasing α can improve throughput, with accuracy declining at overly aggressive settings.
  • Full configurations for Table 18. Relative to the corresponding settings in Table 12, each of the four alternative configurations in Table 18 increases the suffix window size by 32 tokens. The increases in α are 0.1 for MATH, 0.2 for HumanEval, and 0.3 for GSM8K. We will include these configuration details in the revised manuscript. On HumanEval-256, the original model achieved 43.9% accuracy, and Fast-dLLM achieved 37.2% with prefix KV cache and fixed-threshold parallel decoding. This gap may be partly attributable to the approximate reuse of prefix KV states within a block. With prefix KV caching also enabled, Streaming-dLLM can partially close the gap to the original model. Following the tuning criterion above, we reported configurations with accuracy comparable to that of existing acceleration baselines and relatively high throughput.
  • Computational cost for long outputs. Thank you for pointing out that attention still needs to read all prefix KV states. With the block size BB and suffix window size ww fixed, each denoising step processes at most w+1w + 1 suffix positions, and the query length is at most B+w+1B + w + 1. The prefix includes the prompt and the tokens generated so far, so the attention cost per step still depends on the prefix length. We will revise Section 3.3 to clarify that the reduction from linear to constant scaling applies to the number of suffix tokens processed per denoising step.
  • The default configuration provides a good starting point for tuning and transfers to new models and datasets, further demonstrating the practicality of Streaming-dLLM. Table 3 shows that under a 512 token budget, the shared default configuration and the tuned configurations achieve similar accuracy, with throughput varying across tasks. Table 4 shows the results of applying the same default configuration directly to IFEval, a new task. On LLaDA-Base (a new model) and LLaDA-1.5, accuracy was comparable to or higher than that of the original models, and throughput reached 6.7× and 15.7× that of the respective original models.
  • In the revised manuscript, we added Table 6 and an accompanying discussion of how well the default configuration transfers. We provide a more detailed analysis in the new Section B.8 of the appendix. In Appendix B.7, we added details on the hyperparameter selection criteria and the configurations used in Table 18. We also revised the wording about “decoupling inference cost from sequence length.”

Table 3: Comparison between tuned and unified hyperparameter configurations (generation length = 512). Each entry is reported as Accuracy/Throughput.

GSM8KMBPPHumaneval-
Dream-Base (Tuned)74.7/94.155.8/92.454.6/72.3
Dream-Base (Unified Default)74.7/82.955.0/110.854.3/76.7
LLaDA-1.5 (Tuned)81.2/69.838.4/61.440.2/26.7
LLaDA-1.5 (Unified Default)81.1/73.238.4/60.740.2/26.9

Table 4: Zero-shot transfer of the unified default configuration (suffix = 96, α = 0.3) to a new task (IFEval) and a new backbone (LLaDA-Base). Each entry reports Accuracy/Throughput.

IFEval-0shot
LLaDA-Base31.0/15.1
+Streaming-dLLM (suffix=96,α=0.3)30.9/101.9
LLaDA-1.560.7/3.1
+Streaming-dLLM (suffix=96,α=0.3)61.5/48.8

AuthorsOpenReview ↗

Response to Reviewer x2cE (Part 3/3)

Q3: If the 2048-token budget is exhausted without EOS, does the implementation truncate or append masks to continue? If continuation is supported, how are the terminal position and KV cache updated?

  • For both full-sequence dLLM (LLaDA, LLaDA-1.5, and Dream-Base) and pure block dLLM (Fast-dLLM v2 and openPangu), if the fixed 2,048-token generation budget is exhausted without an end-of-sequence (EOS) token, decoding stops at the limit and returns the generated output.

We would like to clarify how DPad and Streaming-dLLM differ in the types of decoding redundancy they address and their applicability to pure block diffusion language models.

  • Streaming-dLLM reduces temporal and spatial redundancy in dLLMs through dynamic threshold decoding and approximated suffix pruning, respectively. DPad targets spatial redundancy alone. Temporally, dynamic threshold decoding uses the mask ratio as a proxy for uncertainty at the current decoding stage, allowing the threshold to adapt to the overall decoding state. This signal complements individual token confidence scores and enables higher throughput. Spatially, we observe that attention concentrates on nearby suffix positions and the end of the sequence, while intermediate suffix positions receive less attention and are still processed at every step. Based on this observation, ASP retains a contiguous window immediately after the current block and the terminal position of the sequence. DPad samples within a candidate window using a distance-decay rule, retaining some redundant intermediate positions while omitting the terminal position. Moreover, we measured how many suffix tokens each strategy actually retains on LLaDA-1.5 (Table 5). Streaming-dLLM retains 31.88 fewer suffix tokens on average in the experiments with a 512 token generation budget and 185.42 fewer on GSM8K with a 2,048 token budget.
  • DPad is limited to full-sequence dLLMs, whereas Streaming-dLLM extends to the latest pure block-diffusion language models. Pure block diffusion is an increasingly important direction for dLLMs. These models support early exit and variable-length generation, and the entire masked generation region is small, typically just 32 tokens. DPad prunes the masked suffix beyond the current block in full-sequence dLLMs. Pure block diffusion confines masking to the current block, leaving no cross-block suffix for DPad to prune. Streaming-dLLM also can improve the efficiency of pure block-diffusion models in both the spatial and temporal dimensions even with short suffixes. Spatially, we divide the current diffusion block into 8-token subblocks, separate the subblock being denoised from the subsequent masked subblocks, and apply approximate suffix pruning to those subblocks to reduce computation. Temporally, Streaming-dLLM adjusts decoding parallelism according to the fraction of masked tokens remaining in the block. Reducing the masked generation region further increases the benefit of the dynamic threshold. Table 3 of the initial submission evaluates this extension on two pure block-diffusion models, Fast-dLLM v2 and OpenPangu-7B-Diffusion-Base. On GSM8K and MATH, Streaming-dLLM achieves 2.6–5.4× the throughput of the +Fast-dLLM baseline, which already uses prefix caching, fixed-threshold parallel decoding, and early exit, at comparable accuracy.

Table 5: Suffix retention statistics for the DPad and Streaming-dLLM pruning strategies on LLaDA-1.5. The parameters k, a, and window size are DPad hyperparameters. The DPad paper evaluated the four datasets only at specific generation lengths (GSM8K-256, MATH-256, HumanEval-512, and MBPP-512). For tasks with different generation lengths, we tune the window size accordingly and generally reuse DPad’s k and a values for similar datasets. Hyperparameter values changed during tuning are shown in bold.

kawindow sizeSuffix tokens retained by DPadSuffix tokens retained by Streaming-dLLM
GSM8K-5123.01.6512145.97129
GSM8K-20483.01.61024282.4297
Math-5123.02.3512206.34193
HumanEval-5123.01.6512145.5897
MBPP-5123.01.6512145.6197

Overall, thank you again for your constructive feedback. We hope our clarifications and new results address your concerns and would be grateful if you considered revisiting your assessment. We welcome any further suggestions.

With best regards!

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

Streaming-dLLM accelerates blockwise diffusion language models without retraining. It includes suffix pruning that preserves terminal positional information, mask-ratio-dependent confidence thresholds, and EOS-based early stopping. Validation covers five text models and one multimodal model across six benchmarks, supplemented by component ablations, latency measurements, and experiments.

Critical strengths

Overall, the paper introduces a compact suffix approximation that preserves both nearby masked tokens and the original sequence's final positional anchor, providing a concrete way to retain structural information during pruning. The practical gains seem substantial and appear across multiple backbones, mathematical reasoning, code generation, and multimodal tasks, supporting usefulness beyond one experimental setting.

  1. Suffix pruning preserves a compact representation of future context: A contribution is retaining both a fixed-size neighboring suffix window and the final token's original position, as shown in Figure 3, preserving positional structure while discarding distant masks.
  2. Dynamic decoding explicitly connects token acceptance to the remaining masked fraction: The adaptive threshold in Equation 10 provides a training-free mechanism for relaxing acceptance as decoding progresses, complementing the spatial pruning. In Table 5, the ablation experiments look good to demonstrate that increasing throughput from 8.2 tokens/s to 44.6 tokens/s while maintaining 81.0% GSM8K accuracy. When combining all components it reaches 69.8 tokens/s at 81.2%.
Critical weaknesses
  1. Reviewers feel confused when reading Figure 1 data and Table 17 data: In Table 17, Fast-dLLM v2 also achieves 82.0% accuracy at 22.1 TPS but it is neglected in Figure 1 that select Dpad 14.5 TPS as the baseline.
  2. Hyperparameters vary by benchmark, model, and generation budget, but tuning split and equivalent tuning effort for baselines are unspecified.
Questions for authors
  1. How were the suffix window and adaptation strength selected, and do the gains persist against equally tuned baselines with identical caching and stopping policies?
  2. What are the actual output-length distributions in the long-generation experiments, and how does performance change on examples that require outputs approaching 2048 tokens?
5 comments
AuthorsOpenReview ↗

Response to Reviewer JuhZ (Part 1/2)

Dear Reviewer JuhZ,

We sincerely appreciate your recognition of our compact suffix approximation, adaptive decoding, and the resulting throughput gains across backbones and tasks. We address your comments below and will incorporate all clarifications, additional results, and revisions from this discussion into the final manuscript. Changes made in response to your comments are highlighted in orange in the revised manuscript.


Q1+W2: Hyperparameters vary by benchmark, model, and generation budget, but tuning split and equivalent tuning effort for baselines are unspecified. How were the suffix window and adaptation strength selected, and do the gains persist against equally tuned baselines with identical caching and stopping policies?

  • Hyperparameter selection. We selected the configurations based on the evaluation results for each of benchmark. We used the same evaluation split for hyperparameter selection and for reporting results. We fixed the block size at 32 tokens and the initial confidence threshold at 0.9. Starting from a 96 token suffix window and a parallelism setting of α = 0.3, we made small adjustments to the window size and α, guided by the sensitivity analysis in Figure 4. We prioritized configurations with accuracy comparable to that of existing acceleration baselines and relatively high throughput. In the GSM8K experiments with LLaDA-1.5, accuracy gains began to level off at a window size of 128 tokens, and increasing the window further reduced throughput. Increasing α can improve throughput, with accuracy declining at overly aggressive settings.
  • The default configuration can transfers to new models and datasets, further demonstrating the practicality of Streaming-dLLM transfer. To test transferability, we applied this configuration without further tuning to a new benchmark (IFEval) in our evaluation, using both LLaDA-1.5 and a new model (LLaDA-Base). On LLaDA-Base, throughput rose from 15.1 to 101.9 tokens/s (6.7×), while accuracy remained essentially unchanged (31.0 to 30.9) (Table 1). Then, we sweep the suffix window from 1-4 decoding blocks. On IFEval, accuracy gains taper off around three blocks (96 tokens), while larger windows reduce throughput (Table 2). Increasing α from 0.2 to 0.4 relaxes the confidence threshold and improves throughput with little change in accuracy.
  • Baseline comparisons. To test whether the throughput gains persist when the baselines also use prefix cache (PC) and early exit (Exit), we added two baseline configurations to the main table. Fast-dLLM (Exit) adds early exit to Fast-dLLM’s existing combination of PC and fixed-threshold parallel decoding (Par), while retaining its published hyperparameters. In the initial submission, we reported DPad with Par + Exit in the main table and with PC + Par + Exit in the long-generation experiments, consistent with the configurations reported in the original DPad paper. We now also include DPad (PC + Par + Exit) in the main table with tuned hyperparameters. On LLaDA-1.5, Streaming-dLLM achieves higher throughput at comparable or slightly higher accuracy in the reported comparisons, reaching 2.0× the throughput of DPad (PC + Par + Exit) on MBPP (Table 3). On GSM8K with a 2,048-token generation budget, Streaming-dLLM achieves 4.7× the throughput of the fastest baseline, DPad (PC + Par + Exit), while improving accuracy from 79.5 to 82.0 (Table 4).
  • In the revised manuscript, we added Table 6 and accompanying discussion on the transfer of the default configuration, with further analysis in the new Appendix Section B.8. We also added Fast-dLLM (Exit) and DPad (PC + Par + Exit) to the main results table. In Appendix B.7, we added details on the hyperparameter selection criteria.

Table 1: Zero-shot transfer of the unified default configuration to a new task (IFEval) and a new backbone (LLaDA-Base). Each entry reports Accuracy/Throughput.

IFEval-0shot
LLaDA-Base31.0/15.1
+Streaming-dLLM (suffix=96,α=0.3)30.9/101.9
LLaDA-1.560.7/3.1
+Streaming-dLLM (suffix=96,α=0.3)61.5/48.8

Table 2: Sensitivity analysis of the default hyperparameters on IFEval using LLaDA-1.5.

IFEval-0shot
LLaDA-1.560.7/3.1
α Sensitivity
w=96,α=0.261.4/45.7
w=96,α=0.361.5/48.8
w=96,α=0.461.3/51.8
Suffix Window size w Sensitivity
w=32,α=0.352.9/57.4
w=64,α=0.357.8/52.9
w=96,α=0.361.5/48.8
w=128,α=0.361.9/43.8

Table 3: Additional baseline comparisons on LLaDA-1.5.

HumanEvalGSM8KMBPPMATHAverage
+Fast-dLLM (Exit)39.8/14.180.4/31.838.1/21.335.1/40.348.4/26.9
+Dpad (PC+Par+Exit)39.1/16.680.9/47.038.2/30.734.6/45.448.2/34.9
+Streaming-dLLM40.2/26.781.2/69.838.4/61.435.1/62.448.7/55.1

Table 4: Additional baseline comparisons on LLaDA-1.5 for long-sequence tasks.

GSM8K-2048
+Fast-dLLM (Exit)80.7/8.4
+Dpad (PC+Par+Exit)79.5/14.5
+Streaming-dLLM82.0/67.6
AuthorsOpenReview ↗

Response to Reviewer JuhZ (Part 2/2)


Q2: What are the actual output-length distributions in the long-generation experiments, and how does performance change on examples that require outputs approaching 2048 tokens?

  • We additionally measured the distribution of actual output lengths for LLaDA-1.5 on GSM8K (5-shot) with a generation budget of 2,048 tokens. As shown in Table 5, Streaming-dLLM had a median output length of 208 tokens across all 1,319 problems. Of these outputs, 1,304 were at most 512 tokens long and 14 exceeded 2,000 tokens, ranging from 2,046 to 2,048 tokens. On those same 14 problems, Streaming-dLLM and vanilla LLaDA-1.5 each answered 6 correctly (42.9%). Some Streaming-dLLM outputs repeated reasoning steps. This repetition may be related to the reuse of output-token KV states across denoising steps (CoTA [1] reported a similar phenomenon in dMLLM). A few vanilla model outputs reached the generation limit before completing their reasoning.
  • In the revised manuscript, We added these statistics to the Appendix. B.9 and added Figure 10 to show the distribution of output lengths in greater detail.

Table 5. Actual output length statistics.

DatasetMedian≤256≤512513–1,999≥2,000Accuracy (≥2,000 output)
GSM8K2081,00413041146/14

W1: Reviewers feel confused when reading Figure 1 data and Table 17 data: In Table 17, Fast-dLLM v2 also achieves 82.0% accuracy at 22.1 TPS but it is neglected in Figure 1 that select Dpad 14.5 TPS as the baseline.

  • In the original submission, Figure 1 compared acceleration methods evaluated on a common backbone, LLaDA-1.5. Because Fast-dLLM v2 (7B) uses a different backbone, it was not part of that comparison. Fast-dLLM v2 (7B) was developed by fine-tuning Qwen2.5-7B-Instruct on 1B tokens to convert it into a pure block diffusion language model (dLLM) that supports variable length generation. LLaDA-1.5 is a full sequence dLLM developed by applying preference optimization to LLaDA-8B-Instruct.
  • In the revised manuscript, We have clarified the descriptions of Figure 1 and Table 19 (Table 17 in initial version) and added Figure 9 to the appendix to include Fast-dLLM v2 in the comparison.

Overall, thank you again for taking the time to review our paper. We appreciate your constructive feedback and encouraging comments. We hope our clarifications and additional results address your concerns, and we welcome any further suggestions.

With best regards!

Reference

[1] Context tokens are anchors: Understanding the repetition curse in dmllms from an information flow perspective. ICLR, 2026.

AuthorsOpenReview ↗

Hope to Get Your Reply

Dear Reviewer JuhZ,

As the deadline approaches, we wanted to follow up briefly on our submission. We have carefully responded to each of your comments. During the rebuttal period, we revised the explanation accompanying Figure 1 and added Figure 9 to include Fast-dLLM v2 in the comparison. We also analyzed the output distributions and performance in our long-generation experiments and identified a unified default hyperparameter configuration that generalizes well across models and tasks and provides a good starting point for further tuning. We have incorporated these updates into the revised manuscript.

We are grateful for your careful reading and encouragement. Your feedback helped us clarify the presentation and expand the evaluation. We hope the updates address the concerns you raised.

Best wishes!

Comments on the paper

1 comment
AuthorsOpenReview ↗

Clarification of Factual Inaccuracies in Reviewer HS6a's Review

Dear AC, SACs, PCs, and Reviewers,

We would like to respectfully point out several significant factual inaccuracies in Reviewer HS6a's review that directly contradict the content of our manuscript and the cited prior work.


Factual Inaccuracy 1 — Misattribution of throughput gains in Table 5 (omitting DCAD's contribution). (Weakness summary + Q3) Reviewer HS6a attributes the entire 22.6×22.6\times throughput speedup to prior work. In Table 5 of our paper, 22.6×22.6\times represents the speedup of PC + DCAD + Exit over the vanilla model, where DCAD is our proposed dynamic-threshold decoding method. Attributing this entire factor to prior work overlooks DCAD's core contribution.


Factual Inaccuracy 2 — Incorrect baseline hyperparameter figures in comparing DPad and ASP. (W2 + Q1) In comparing DPad with Streaming-dLLM’s suffix approximation method (ASP), Reviewer HS6a states that DPad retains approximately 64 suffix tokens on LLaDA-1.5 / GSM8K-512. However, the hyperparameter configuration underlying that figure comes from GSM8K-256 (Table 9 of the DPad paper), whereas DPad uses distinct hyperparameter settings across different generation lengths. In the GSM8K-512 evaluation, DPad retains an average of 145.97 suffix tokens, compared with 129 for ASP.


Factual Inaccuracy 3 — Misreading the throughput breakdown in Table 5 (22.6×22.6\times to 28.0×28.0\times). (Weakness summary) Building on the misattributions above, the reviewer HS6a concludes: "Once prior work is subtracted, almost nothing new remains, least of all the speedup." Specifically, the reviewer excludes the 22.6×22.6\times speedup, attributes the increase from 22.6×22.6\times to 28.0×28.0\times to the dynamic nature of DCAD, and states that ASP increases the time cost. This directly contradicts Table 5: the increase from 22.6×22.6\times to 28.0×28.0\times is achieved specifically by adding ASP on top of PC + DCAD + Exit. Furthermore, beyond throughput gains over existing baselines, Streaming-dLLM extends to the latest pure block dLLMs, which is a capability that DPad lacks.


Factual Inaccuracy 4 — Misinterpretation of the MBPP accuracy gains. (W3 + Q4) Reviewer HS6a notes that both DPad and ASP yield gains on MBPP and uses this to argue that our attribution of the MBPP accuracy anomaly to interference from distant mask tokens in the full suffix region is incorrect, inferring instead that the anomaly stems solely from non-termination or formatting errors. Because both DPad and ASP are suffix-pruning methods, the fact that both alleviate this issue directly aligns with—rather than refutes—our hypothesis. The additional experiments provided in our rebuttal further corroborate this conclusion.


Factual Inaccuracy 5 — Claiming "no isolation study" despite the ablation in Table 5. (W4) Reviewer HS6a claims that no experiment isolates the effect of changing only the acceptance rule with all other conditions held fixed. In fact, Table 5 of our manuscript directly isolates the improvement of our DCAD over Fast-dLLM under identical settings.


Factual Inaccuracy 6 — Incorrect claim regarding DPad's hyperparameter tuning ("asymmetric tuning"). (W4) Reviewer HS6a states that the DPad baseline uses default values, alleging asymmetric tuning. As documented in Table 9 of the original DPad paper, DPad uses separate, non-default hyperparameter settings tailored to each model, benchmark, and generation length.


Overall, the primary criticisms raised by Reviewer HS6a stem from factual misunderstandings of the methods, experimental setups, and results clearly presented in our submission. While we welcome constructive feedback and diverse perspectives during peer review, a fair evaluation requires engagement with the actual content of the paper.

We will submit a detailed comment to the Area Chair detailing these discrepancies, and we respectfully request that these factual inaccuracies be taken into account when weighing Reviewer HS6a's assessment in the final decision.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.