acceptodds
Under review as a conference paper at ICLR 2027

VisWave: Revisable Wavefront Decoding for Diffusion Vision-Language Models

Abstract

Diffusion multimodal large language models (Diffusion MLLMs) leverage bidirectional context modeling and intra-block multi-token denoising to provide an efficient text-generation path distinct from autoregressive decoding. However, long-form generation remains constrained by serial dependencies across blocks, causing inference efficiency to degrade as the output grows. Existing acceleration methods mainly reduce intra-block denoising cost and do not exploit computation that can overlap between adjacent blocks. The key obstacle is that an early-started successor depends on an unstable predecessor context and may become invalid as that context is updated. We find that this effect is usually local: most successor draft tokens remain reusable, while only a small number require re-verification.Based on this observation, we introduce VisWave, a training-free decoder for diffusion vision-language models. VisWave overlaps the decoding of adjacent blocks through revisable wavefront execution and uses virtual-remask token-to-token verification (VR-T2T) to re-evaluate affected draft tokens under the latest textual and visual context. Supported tokens are retained, while the remaining positions are remasked and returned to the native denoising process. This local selective revision makes inter-block parallelism reliable. Across seven natural workloads, VisWave accelerates BARD-VL-4B by , with only a 0.27 percentage-point decrease in macro accuracy across six accuracy tasks; at a fixed generation length of 192 tokens, it reaches speedup. The efficiency gains remain stable from 2B to 8B and transfer to SDAR-VL.

Primary area:
foundation or frontier models, including LLMs
Keywords:
diffusion vision-language models, parallel decoding, inference acceleration, long-form multimodal generation, token verification

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

This paper proposes VisWave, a training-free decoder for block-diffusion vision-language models. It overlaps adjacent blocks through revisable wavefront decoding and introduces VR-T2T verification to retain or revise tokens generated under stale contexts selectively. The method reports a 1.57× average speedup on BARD-VL-4B with a 0.27pp decrease in macro accuracy, and demonstrates transfer across model scales and architectures.

Critical strengths

S1. Effective training-free acceleration. The paper addresses the serial dependency between blocks and achieves consistent wall-clock speedups without additional training.

S2. Broad evaluation and detailed analysis. Experiments cover multiple model scales, two diffusion VLM architectures, and open-ended generation. The paper also provides statistical analyses and verification ablations.

Critical weaknesses

W1. Insufficient comparison with optimized native decoding.

The default configuration uses four denoising steps for a block of four tokens (Table C1), resulting in nearly one forward evaluation per generated token (Table 1 and Figure 1a). Thus, the reported acceleration is measured against a decoder with nearly autoregressive-like sequential depth. Without evaluating alternative native configurations, such as fewer denoising steps or larger blocks, it remains unclear whether VisWave offers a better quality–latency trade-off than simpler decoding schedules.

W2. Limited practical generality of the evaluation protocol.

The main benchmarks use an Evidence-192 + Answer-8 protocol, which encourages longer generations than standard direct-answer VQA. Since VisWave primarily benefits long sequences, this setting may overstate its practical advantages. For instance, ChartQA achieves only 1.226× speedup with an 8.0% NFE reduction. A comparison with standard direct-answer decoding would better establish its practical usefulness.

W3. Limited novelty over existing block-diffusion decoding methods and unclear multimodal-specific contribution.

The core idea of VisWave is overlapping adjacent blocks and revising tokens generated under stale contexts, which closely resembles existing approaches for text-only block-diffusion LLMs, particularly FlowBlock (Tian et al., 2026). While the authors introduce VR-T2T to address the lack of native token-editing capabilities, the underlying verification mechanism appears largely modality-agnostic.

Furthermore, Section 5.3 only demonstrates that visual inputs influence verification decisions, which does not establish what challenges are unique to multimodal diffusion or why a distinct multimodal method is necessary. The absence of a direct comparison with a FlowBlock-inspired baseline makes the algorithmic novelty difficult to assess.

The additional complexity also appears insufficiently justified: Table 5 shows no accuracy improvement from adding VR-T2T (65.60 → 65.60), despite its substantial token-level verification gains. A clearer distinction from prior text-only methods and stronger evidence for the necessity of the proposed components would strengthen the contribution.

W4. Quality preservation and verification reliability require further analysis.

Despite the small average accuracy degradation, several tasks show notable drops, including MMStar (−1.87pp), SDAR-VL ChartQA (−4.24pp), and DOCCI CIDEr (11.24 → 7.67). Token agreement with native decoding is also substantially lower for BARD-VL-4B than for its 2B/8B variants.

Moreover, the visible-token verifier achieves an AUROC below chance (0.418), and the reported 87.7% draft reuse rate is based on only 219 tokens. These observations warrant further analysis of verification reliability and quality degradation.

References Tian, B., Liu, H., Zhong, X., Duan, Z., Luo, Z., Jin, H., Wang, Z., & Liao, X. (2026). FlowBlock: Wavefront-Parallel Decoding for Self-Correcting Diffusion Language Models. arXiv:2607.17652.

Questions for authors
  1. How does VisWave compare with native decoding using alternative block sizes and denoising steps? Can the authors provide a quality–latency Pareto comparison?
  2. What are the key algorithmic differences between VisWave and existing text-only wavefront decoding approaches such as FlowBlock? Which challenges specifically arise from multimodal conditioning?
  3. Can the authors provide a simpler wavefront verification baseline and clarify the contribution of VR-T2T in Table 5?
3 comments
AuthorsOpenReview ↗

Response to Reviewer 3 (Part 1): Native Decoding Configurations and Quality–Latency Comparisons

Thank you for your careful review and constructive feedback. We have added experiments and analyses to address your concerns.

Q1. Quality–latency comparisons with alternative native configurations

Response: We supplemented our evaluation with quality–latency comparisons across different block sizes and numbers of denoising steps under the same experimental configuration. The results are presented in Table R3-1.

Compared with native decoding under different configurations, VisWave maintains competitive overall quality–latency performance.

On QA tasks, VisWave's task scores are close to those of Native under the same configuration. It achieves speedups on all four datasets at 4/4 and on three datasets at 4/2. Under the 8/4 and 16/8 configurations, the model generates only approximately 9–16 tokens on average, leaving limited computation for cross-block overlap and making the additional overhead difficult to amortize.

On DOCCI long-form description generation, VisWave further reduces latency in all four same-configuration comparisons, achieving speedups of 1.23–1.55×. In cross-configuration comparisons, VisWave K4/T2 is faster than Native K8/T4 while also achieving higher CIDEr and METEOR. These results show that, for long-description tasks, fewer steps and larger blocks can accelerate native decoding, while revisable wavefront execution provides additional acceleration for long-form text generation.

R3-1(a). Full QA evaluations with static K/T

(K/T: block length/denoising steps per block; RWQA: RealWorldQA.)

K/TMethodRWQAAI2DChartQAMME
QualityLatencyNFETokensQualityLatencyNFETokensQualityLatencyNFETokensQualityLatencyNFETokens
4/4Native72.422.61169.9358.0682.062.93288.5572.9272.880.84122.4918.982288.292.47275.1862.15
4/4VisWave72.421.701 (1.535×)40.0253.9781.571.710 (1.715×)48.3866.5472.120.706 (1.190×)17.9718.682251.291.472 (1.679×)42.6555.62
4/2Native72.160.84014.0120.1184.070.3798.4512.8772.440.49911.8713.362304.020.43611.7117.16
4/2VisWave71.900.736 (1.141×)10.7518.9684.260.356 (1.064×)7.5412.2472.200.513 (0.973×)12.1513.432280.930.385 (1.133×)9.8016.96
8/4Native73.860.72010.9315.8684.100.3648.0411.5072.440.50712.1512.002253.390.2776.869.44
8/4VisWave73.590.728 (0.989×)10.6415.5684.000.393 (0.927×)8.6511.6772.440.541 (0.938×)13.0411.992258.240.309 (0.899×)7.639.33
16/8Native68.370.71110.3910.4082.930.43610.339.6168.760.62115.7211.342272.870.37910.089.07
16/8VisWave68.500.756 (0.940×)11.3210.3683.260.479 (0.910×)11.259.5768.680.667 (0.930×)16.7011.362277.910.427 (0.887×)11.069.06

(b) DOCCI: confirmation on 300 images

K/TMethodCIDEr ↑METEOR ↑Latency (s) ↓NFE ↓Tokens
4/4Native7.50417.3818.110250.34200.25
4/4VisWave6.70717.0865.223 (1.553×)151.21191.71
4/2Native5.98614.0685.651171.89228.73
4/2VisWave7.55914.4923.884 (1.455×)108.21220.35
8/4Native4.75213.7594.505132.36211.28
8/4VisWave7.83713.9053.529 (1.277×)96.15208.99
16/8Native7.81012.7433.44799.18175.95
16/8VisWave7.19212.7172.805 (1.229×)73.55154.08
AuthorsOpenReview ↗

Response to Reviewer 3 (Part 2): Algorithmic Differences, Verification Baselines, and Evaluation Protocol

Q2. Algorithmic differences and multimodal conditioning

Response: VisWave and FlowBlock both shorten serial waiting by overlapping computation across adjacent blocks. The main differences lie in the draft revision mechanism and its portability. FlowBlock relies on the backbone's native visible-token editing capability. VisWave's VR-T2T instead retains the logical draft, temporarily masks candidate positions in the model input, and jointly re-predicts them using the latest context. Top-1 agreement and a support threshold determine whether a candidate is retained or remasked. Thus, VR-T2T does not require native visible-token editing, allowing wavefront execution to extend to target models that lack reliable visible-editing signals. On the BARD-VL checkpoint used in this paper, the state-matched audit shows that masked re-prediction increases verification AUROC from 0.418 for visible-token scoring to 0.942; see Eqs. 8–10 and Table F1.

Our work focuses on visual evidence generation and detailed descriptions. The core challenge is to construct reliable revision signals for visual descriptions generated early while the textual context continues to evolve. A fixed image does not imply that the draft's visual support remains unchanged: predecessor updates may change the object, attribute, or relation to which a successor token refers. Therefore, predictive support under the old context cannot directly establish validity in the new context. VR-T2T masks and re-predicts candidates under the joint conditioning of the latest text and the original image, reassesses their support, and then decides whether to retain or regenerate them; see Eqs. 9–10.

Q3. Simpler verification baselines and VR-T2T's contribution

Response: We added visible-token verification as a simple baseline and used wavefront decoding without revision as a reference. We compared different revision strategies under identical wavefront scheduling and fused-preview settings. VR-T2T verifies candidates through masked re-prediction, retains drafts that remain supported by the latest multimodal context, and returns rejected positions to native denoising. Full VisWave additionally incorporates verified prefetch. On the same 100 DOCCI images, VR-T2T achieves METEOR 17.25, exceeding wavefront without revision and visible-token verification by 2.02 and 7.76 points, respectively; see Table R3-3.

Table R3-2. Comparison of wavefront revision strategies on DOCCI

MethodCIDEr↑METEOR↑Latency(s)↓NFE↓TokensRemasks
BARD-VL-4B11.2417.998.200253.51202.800.00
Wavefront w/o revision4.1615.233.801106.62205.240.00
Visible-T2T6.079.494.703134.45246.240.00
VR-T2T7.6717.255.037141.86183.1636.18
VisWave7.6717.244.960142.30183.0036.10

W2. Evidence protocol and practical VQA benefits

Response: VisWave primarily targets scenarios that continuously generate multiple text blocks. We use Evidence-192 + Answer-8 to separate visual evidence generation from final-answer generation and examine cross-block decoding efficiency under a common generation budget. We use VQA benchmarks because their reference answers allow us to test whether changing the decoding process preserves final task performance. This evaluation therefore measures both generation efficiency under long-evidence workloads and question-answering accuracy.We additionally use DOCCI to directly evaluate the complete generated descriptions. The corresponding quality–latency results are reported in the 300-image comparisons in our response to Q1.

Thank you for the detailed rebuttal and additional experiments. I appreciate the authors' efforts to address my concerns, particularly the new quality–latency comparisons and verification ablations.

The additional VR-T2T ablations provide useful evidence for the effectiveness of the proposed verification mechanism. However, my primary concerns regarding the practical advantage over optimized native decoding and the algorithmic novelty remain.

First, the new results in Table R3-1(a) suggest that the acceleration benefits are strongly dependent on the decoding configuration. While VisWave achieves meaningful speedups under K4/T4, simply adjusting the native block size and denoising steps yields more favorable quality–latency trade-offs on QA benchmarks. For example, Native K8/T4 achieves higher accuracy and substantially lower latency than VisWave K4/T4 on RealWorldQA. Furthermore, VisWave is consistently slower than native decoding under K8/T4 and K16/T8 across all four QA benchmarks. These results raise concerns about the practical advantages of VisWave under more efficient native decoding configurations.

Second, while the authors clarify the differences between VR-T2T and FlowBlock, the multimodal-specific algorithmic contribution remains insufficiently established. The proposed masked re-prediction strategy appears broadly applicable to text-only diffusion models as well, and the rebuttal does not provide direct evidence demonstrating its necessity specifically for multimodal decoding.

I acknowledge that the DOCCI results demonstrate consistent latency reductions for long-form generation, and that the additional ablations strengthen the motivation for VR-T2T. Nevertheless, the above concerns remain significant.

I will maintain my original rating, as the rebuttal partially addresses my concerns but does not fully resolve the questions regarding practical advantages and algorithmic novelty.

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

VISWAVE is a training-free decoder that overlaps adjacent text blocks of a frozen block-diffusion VLM. A successor is spawned once the predecessor's reveal ratio r reaches rho (Eq. 5), and a block commits only when fully revealed with no pending validations (Eq. 6). VR-T2T (Sec. 3.2) masks each stale, low-support successor token while keeping its draft in logical state, re-predicts all candidates in one forward under the latest multimodal context, and keeps a draft only if top-1 agrees and p_vr >= tau_v (Eq. 10). On BARD-VL-4B, under an evidence-first protocol (up to 192 evidence tokens, then an 8-token answer; App. A), it reports a 1.57x mean natural speedup at a 0.27-point six-task macro accuracy drop (Table 1), 2.07x at a fixed 192-token length (Fig. 4), 2B/8B scaling, and SDAR-VL transfer.

Critical strengths
  1. VR-T2T is cleanly validated at the token level. The paper tests directly whether the frozen backbone can judge its own visible drafts, and it cannot: across the 4,744-event audit, visible-token scoring reaches 0.418 AUROC, while masking each candidate and re-predicting it reaches 0.942 (Table F1). This is a simple, training-free remedy for self-conditioning, and the same audit backs the paper's argument that a literal FlowBlock port lacks a reliable revision operator on this backbone (L765-766). It is the most original part of the paper.

  2. The speed gain is attributed to the right mechanism. With both decoders forced to the same 192-token length, NFE falls from 241.00 to 109.78 (54.45%) while profiled QK pairs rise 2.63%, so the gain comes from fewer sequential forwards rather than attention sparsity (App. D.2, L859-863). VISWAVE's paired bootstrap intervals all stay above 1x (L351-352; Table D2), and the speed gain carries over to the 2B and 8B checkpoints (Table 2), to SDAR-VL (Table 3) and to DOCCI descriptions (1.515x, Table 4a). For a scheduling method, this is the right measurement.

  3. Quality changes are tested per dataset, not only as a macro average. Table E1 gives a paired interval and a McNemar test for all seven datasets, and the paper names MMStar as the one significant loss (-1.87, p=0.0451; L881-883) instead of leaning on the -0.27 macro. That makes each per-dataset quality claim checkable, which is the right standard for a decoder that claims to preserve quality.

Critical weaknesses
  1. The headline speedup is measured in a long-output regime that the paper's own protocol creates. The evidence-first prompt "elicits task-relevant long-form content" (L316), so the native model emits 62.1 tokens on average for MME yes/no questions (Table 1), and no Table 1 metric scores that text. BARD-VL's released evaluation code caps generation at 16 new tokens on MME, ChartQA, RealWorldQA and AI2D. The paper ties the gain to output length, and ChartQA, at 19.0 native tokens, gets the smallest speedup (1.226x; L333-335). No run uses the backbone's own prompts or caps, so the speedup at that setting is unknown.

  2. No baseline is built to cut the number of forwards, so the 1.57x is relative only to a native schedule with one denoising step per token. VISWAVE's speed comes from fewer sequential forwards (L863), while the native decoder runs 4 denoising steps per 4-token block (Table C1; L313-314). RedVTP and D3ToM instead cut per-forward cost (L328-329) and run at 0.985x and 0.598x as ported. The paper itself lists confidence-aware decoding as an existing axis (L40, L746), and BARD-VL's released inference defaults to it, committing every masked token above a threshold in one forward. A comparison against it at matched accuracy would show what VISWAVE adds.

  3. The ablation does not show that VR-T2T, the audited mechanism, measurably changes the end-to-end result. In the cumulative Table 5, adding VR-T2T leaves accuracy at exactly 65.60, the same as the "+ fused preview" row, and moves speed only from 1.80x to 1.83x. The table names no dataset and reports no intervals, and its native 66.60 matches no Table 1 cell, so it cannot be placed against the main results. Naming the subset and adding paired intervals per row cost no new runs; repeating the VR-T2T step on two or three Table 1 datasets would show whether the Table F1 gain survives end to end.

Questions for authors
  1. Can you report accuracy and latency for both decoders under BARD-VL's released evaluation prompts and generation caps? This decides whether the speedup holds outside your protocol. (Weakness 1)
  2. How does VISWAVE compare at matched accuracy with BARD-VL's own confidence-threshold decoding, and do the two compose? S2D2 on SDAR-VL would answer the same question. (Weakness 2)
  3. Which dataset and how many examples does Table 5 use? Can you add paired intervals per row and repeat the "+ VR-T2T" step on two or three Table 1 datasets? (Weakness 3)
  4. DOCCI is the only task whose long output is itself scored, and there qual_test CIDEr falls from 11.24 to 7.67 with 9.76% shorter outputs (Table 4b). Can you report length-matched CIDEr and METEOR with intervals?
  5. Clarification: which BARD-VL checkpoint (block size, noise type) is used? Chen et al. (2026) add mixed-noise training to improve token revision, which bears on L765-766 and Table F1.
3 comments
AuthorsOpenReview ↗

Response to Reviewer 2 (Part 1): Released Evaluation Settings and Confidence-Threshold Comparisons

Thank you for your careful review and constructive feedback. We have added experiments and analyses to address your concerns.

Q1. Evaluation under BARD-VL's released prompts and generation caps

Response: We completed full evaluations on four tasks using BARD-VL's released prompts and generation caps; see Table R2-1. Under this setting, VisWave's task quality is generally close to native decoding: accuracy changes across the three accuracy-based tasks range from −0.36 to +0.52 percentage points, while the MME score increases slightly. Mean latency increases slightly on three tasks.

VisWave is motivated by reducing inter-block serial waiting in long-text workloads, such as visual evidence generation and detailed descriptions. Its gains come from overlapping computation across adjacent text blocks. Under the official setting, short answers average only 4–6 tokens, leaving very little cross-block computation to overlap. The additional scheduling and verification overhead is therefore difficult to amortize, and speedups are not universal. Short-answer VQA is not the primary acceleration target of our method.

Table R2-1. Quality and efficiency under the official short-answer setting

(B/V : BARD-VL/VisWave)

DatasetQuality B/V (Δ) ↑Latency B/V (s) ↓ (speedup)Tokens B/V
RealWorldQA72.55 / 73.07 (+0.52)0.295 / 0.330 (0.894×)4.00 / 4.00
AI2D83.52 / 83.58 (+0.07)0.199 / 0.234 (0.851×)4.00 / 4.00
ChartQA83.84 / 83.48 (−0.36)0.278 / 0.260 (1.070×)5.98 / 5.98
MME2316.22 / 2329.74 (+13.52)0.183 / 0.218 (0.841×)4.00 / 4.00

Q2. Native confidence-threshold decoding and its combination with VisWave

Response: We added full evaluations on five datasets to compare the original VisWave configuration with BARD-VL's native confidence-threshold decoding and to evaluate their combination; see Table R2-2. At operating points with similar accuracy, VisWave further reduces latency on some tasks. Moreover, at the same confidence threshold, combining the two yields additional speedups on most tasks. These results show that VisWave can complement native confidence-threshold decoding.

Table R2-2. Native confidence-threshold decoding and its combination with VisWave

(a) Original VisWave versus native confidence-threshold decoding

Conf.DatasetQuality B/V (Δ) ↑Latency B/V (s) ↓Speedup↑NFE B/V ↓Tokens B/V
0.6RealWorldQA71.24 / 72.42 (+1.18)2.043 / 1.6681.225×53.98 / 40.0257.53 / 53.97
0.6AI2D81.87 / 81.57 (−0.29)2.175 / 1.6651.306×66.51 / 48.3872.42 / 66.54
0.6ChartQA71.76 / 72.12 (+0.36)0.680 / 0.7040.966×17.54 / 17.9718.78 / 18.68
0.6MMMU-Val52.73 / 54.35 (+1.62)2.699 / 2.1231.271×102.32 / 76.73115.60 / 108.75
0.6MMMU-Pro32.89 / 33.35 (+0.46)2.614 / 2.0451.278×99.63 / 74.08112.53 / 105.10
0.9RealWorldQA72.29 / 72.42 (+0.13)2.323 / 1.6671.394×62.44 / 40.0258.01 / 53.97
0.9AI2D82.03 / 81.57 (−0.45)2.577 / 1.7261.493×76.96 / 48.3872.92 / 66.54
0.9ChartQA72.80 / 72.12 (−0.68)0.748 / 0.6991.069×19.58 / 17.9718.95 / 18.68
0.9MMMU-Val52.94 / 54.35 (+1.42)3.139 / 2.1231.479×119.81 / 76.73117.27 / 108.75
0.9MMMU-Pro32.54 / 33.35 (+0.81)2.993 / 2.0451.464×114.91 / 74.08112.94 / 105.10

(b) Combining VisWave with native confidence-threshold decoding

Conf.DatasetQuality B/V (Δ) ↑Latency B/V (s) ↓Speedup↑NFE B/V ↓Tokens B/V
0.4RealWorldQA71.50 / 71.24 (−0.26)1.728 / 1.3921.241×42.50 / 30.8152.50 / 50.34
0.4AI2D82.12 / 81.99 (−0.13)1.760 / 1.3641.290×53.12 / 38.9665.78 / 63.67
0.4ChartQA71.08 / 70.68 (−0.40)0.625 / 0.6241.002×15.99 / 15.6718.71 / 18.83
0.4MMMU-Val54.05 / 51.82 (-2.23)2.221 / 1.6911.313×84.94 / 61.51108.15 / 104.66
0.4MMMU-Pro33.82 / 33.99 (+0.17)2.164 / 1.6261.331×81.68 / 58.06104.75 / 98.45
0.6RealWorldQA71.24 / 72.68 (+1.44)2.095 / 1.5621.341×53.98 / 36.1057.53 / 53.54
0.6AI2D81.87 / 81.80 (−0.06)2.181 / 1.5421.414×66.51 / 44.4172.42 / 66.85
0.6ChartQA71.76 / 71.16 (−0.60)0.676 / 0.6511.038×17.54 / 16.4818.78 / 18.58
0.6MMMU-Val52.73 / 52.73 (+0.00)2.699 / 1.9371.393×102.32 / 69.72115.60 / 109.47
0.6MMMU-Pro32.89 / 33.47 (+0.58)2.614 / 1.8661.401×99.63 / 67.25112.53 / 105.92
0.9RealWorldQA72.29 / 73.99 (+1.70)2.352 / 1.6651.412×62.44 / 38.8158.01 / 54.13
0.9AI2D82.03 / 81.51 (−0.52)2.496 / 1.6101.551×76.96 / 46.5872.92 / 66.85
0.9ChartQA72.80 / 72.16 (−0.64)0.738 / 0.6761.092×19.58 / 17.3318.95 / 18.73
0.9MMMU-Val52.94 / 53.54 (+0.61)3.139 / 2.0511.530×119.81 / 73.98117.27 / 109.86
0.9MMMU-Pro32.54 / 33.06 (+0.52)2.993 / 1.9661.523×114.91 / 70.95112.94 / 105.61
AuthorsOpenReview ↗

Response to Reviewer 2 (Part 2): VR-T2T Ablations, Length-Matched Quality, and Checkpoint Clarification

Q3. Table 5 and the end-to-end contribution of VR-T2T

Response: The original Table 5 used three multiple-choice visual question answering datasets, MMStar, AI2D, and MMMU Val, focusing on the impact of cumulative components on the average accuracy and execution efficiency of the final short answer macro, without directly evaluating the quality of the generated text. Therefore, the presentation of the revision effect of VR-T2T is not direct enough. According to your suggestion, adding VR-T2T to the controlled controls did not show clear additional accuracy or delayed benefits. We will adjust Table 5 in the revised manuscript to reflect the revision strategy comparison on DOCCI, which better reflects the role of the module. In the newly added experiment, the METEOR of VR-T2T is 17.25, which is 2.02 and 7.76 points higher than the unmodified wavefront and Visible T2T, respectively, as shown in Supplementary Table R2-3.

Table R2-3. Comparison of wavefront revision strategies on DOCCI

MethodCIDEr ↑METEOR ↑Latency (s) ↓NFE ↓TokensRemasks
BARD-VL-4B11.2417.998.200253.51202.800.00
Wavefront w/o revision4.1615.233.801106.62205.240.00
Visible-T2T6.079.494.703134.45246.240.00
VR-T2T7.6717.255.037141.86183.1636.18
VisWave7.6717.244.960142.30183.0036.10

Q4. Length-matched CIDEr and METEOR on DOCCI

Response: We added a length-controlled experiment with 256 fixed generation positions on the same 100 DOCCI qual_test images, using consistent EOS handling, and report CIDEr, METEOR, and their paired 95% confidence intervals; see Table R2-4. A small METEOR gap remains after controlling for length. The table also includes the natural-termination results for comparison. We will update Table 4 in the revised manuscript.

Table R2-4. DOCCI quality under natural and fixed-length generation

GenerationMetric ↑BARD-VL-4BVisWaveΔScore [95% CI]
Natural EOS (cap 384)CIDEr11.2427.665−3.577 [−11.968, +4.094]
Natural EOS (cap 384)METEOR17.99417.245−0.749 [−1.352, −0.136]
Fixed 256 positionsCIDEr1.9372.569+0.632 [−3.203, +3.497]
Fixed 256 positionsMETEOR18.92318.206−0.718 [−1.293, −0.171]

Q5. BARD-VL checkpoint and noise type

Response: The original experiments use Bard-VL-B4-Mask-4B-Instruct, with a default block size of 4 and four denoising steps per block; see Appendix C. The official configuration for this checkpoint specifies prior_dist='Mask' and block_size=4. Our experiments use the Mask checkpoint, not the mixed-noise variant. All verification results in Table F1 are based on this checkpoint. We will add this clarification to the revised manuscript.

Thanks for the authors' responses. Given that my major concerns have been addressed, I would like to raise my score to 3.

Reviewer_e8g9OpenReview ↗
Rating
3(+1)
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

VisWave proposes a training-free decoding method for diffusion vision-language models that accelerates generation by overlapping adjacent blocks and selectively revising tokens generated under incomplete context. It introduces VR-T2T to identify unreliable draft tokens using updated multimodal context, retaining reliable tokens while regenerating the others. The method achieves promising speedup with some quality degradation.

Critical strengths
  1. The method achieves promising speed up. Through experiment, Viswave is effective across different diffusionVLMs with small quality degration. Consider it's training-free, this is useful and might be generalized to many models.
  2. A mechanism to check if decoded token is realiable. The proposed VR-T2T mechanism is simple to verify if the decoded token with incomplete token is correct or not in heuristic way. It does not require the model have native capbility of text-to-text refinement.
Critical weaknesses
  1. Lack of comparison to FlowBlock. My main concern is that the overall framework is conceptually very similar to FlowBlock. . While VR-T2T is useful for models without native T2T editing capabilities, it is unclear how muc it improves upon existing methods. Since the proposed mechanism does not inherently depend on visual inputs, I would like to see experiments on text-only diffusion models, ideally with a direct comparison to FlowBlock under the same setting.
  2. Reliability of the token selection mechanism. The method uses confidence scores from an incomplete context to decide which tokens need verification. However, these scores may be biased or poorly calibrated since the incomplete context can be out-of-distribution. Some high-confidence tokens may actually be incorrect under the final context but are never selected for verification. Although the paper reports verification accuracy on revised tokens in Appendix F, I think it would be valuable to measure the recall of unreliable tokens across the entire draft rather than accuracy only over the selected tokens. Specifically, the recall of VR-T2T seems not reported but only precision.
Questions for authors

Same as weakness. I'm willing to raise the score if the concern is addressed.

3 comments
AuthorsOpenReview ↗

Response to Reviewer 1 (Part 1): Additional FlowBlock Comparisons, Backbone Applicability, and Draft Verification Reliability

Thank you for your careful review and constructive feedback. We have added experiments and analyses to address your concerns.

Q1. Differences from FlowBlock and direct comparisons on text-only diffusion models

Response: Our motivation is to accelerate long-form generation in diffusion VLMs.VisWave shares inter-block wavefront parallelism with FlowBlock; the main difference lies in draft revision. FlowBlock revises visible drafts through native joint M2T/T2T updates. VR-T2T preserves the logical draft, temporarily masks candidate input positions, and re-predicts them under the updated context to decide Keep/Remask. This uses the backbone's existing masked-prediction capability without additional training or reliance on reliable visible-token editing. We agree that the verification operator also applies to text-only backbones.

To examine this difference, we added a direct DOCCI long-description comparison on BARD-VL (Table R1-1). FlowBlock showed substantial output degradation, with METEOR 0.67 versus VisWave's 17.24. VisWave achieved 1.62× speedup over native decoding, with some loss in description quality; complete metrics are below. These results support VR-T2T's applicability to long-form generation on the evaluated Mask checkpoint and demonstrate its quality–efficiency trade-off.

Table R1-1. DOCCI long-description generation on BARD-VL-4B.

MethodCIDEr ↑METEOR ↑Latency (s) ↓NFE ↓TokensEmpty outputs / 100
BARD-VL-4B11.2417.998.110253.51202.800
Serial Joint-T2T≈00.603.48895.2193.8442
FlowBlock≈00.671.97048.06100.3241
VisWave7.6717.245.005142.30183.000

Following your suggestion, we added GSM8K and MATH500 comparisons on LLaDA2.0-mini, LLaDA2.1-mini, and SDAR-4B-Chat (Table R1-2). VisWave's wavefront-launch and VR-T2T thresholds were carried over from the DVLM setting, without selecting new thresholds for each backbone. On LLaDA2.0-mini, VisWave achieved 1.139× and 1.210× speedups, respectively; on SDAR-4B-Chat, it reduced NFE by approximately 21%. Both exhibited accuracy losses,These results support the applicability of VR-T2T in the current Mask checkpoint long text generation and present its quality efficiency trade-off. On LLaDA2.1-mini, officially supported by FlowBlock, current VisWave did not outperform FlowBlock in quality or efficiency. These results show backbone-dependent applicability.

Table R1-2. Comparisons on text-only models

Model / methodGSM8KMATH500
Acc. (%) ↑Latency (s) ↓NFE ↓TokensAcc. (%) ↑Latency (s) ↓NFE ↓Tokens
LLaDA2.0-mini91.133.876129.97251.1445.405.517186.35447.84
FlowBlock0.9916.565541.66511.661.0014.961491.95505.28
VisWave87.953.403109.94262.4943.604.561148.11452.28
LLaDA2.1-mini87.872.02167.30265.2547.602.70090.47434.29
FlowBlock89.921.63953.06268.8948.002.17470.70446.25
VisWave73.772.79489.82275.1032.603.915126.59443.57
SDAR-4B-Chat91.748.124115.60162.7352.6020.559278.87417.75
FlowBlock0.6814.250187.20418.251.2011.043136.44274.00
VisWave87.196.77390.94164.7850.8017.339220.19411.79

To further analyze the performance differences on LLaDA2.1-mini, we added a 100-question controlled ablation and same-state audit. Disabling native T2T editing reduced accuracy from 70% to 56% and increased mean NFE from 80.89 to 156.54, showing the importance of native editing to this backbone's quality and efficiency. The same-state audit also found that some draft tokens retained by a native-editing continuation failed VR-T2T's masked verification and were subsequently remasked, thereby affecting subsequent generation quality and efficiency.

AuthorsOpenReview ↗

Response to Reviewer 1 (Part 2): Additional FlowBlock Comparisons, Backbone Applicability, and Draft Verification Reliability

Q2. Missed high-confidence tokens and recall over the entire draft

Response Thank you for suggesting draft-wide recall evaluation. Building on Appendix F, Table F1, we audited 400 questions from the same two datasets, MMStar and AI2D. At predecessor commit, we examined all revealed tokens in the successor draft, obtaining 9,089 draft events. This extends evaluation beyond verification reliability at selected positions to candidate coverage and pipeline recall. At the default selection threshold of 0.4, selection recall was 54.01%, and conditional recall after selection was 99.08%.

We then fixed the original drafts and contexts, varied the selection threshold, and reran joint-mask verification for each corresponding candidate set. At 0.8, 59.56% of draft events entered verification, selection recall reached 95.74%, and conditional recall was 96.33%. On these fixed draft states, combined selection-and-verification recall reached 92.22%. Thus, expanding candidate coverage improves recall of unsupported drafts while maintaining high conditional detection rates. Table R1-2 reports stage-wise recall and confidence intervals.

Table R1-3. Candidate coverage and verification recall on fixed drafts

τrSelected (%)Selection recall (%)Conditional recall (%)Pipeline recall (%)
0.417.7654.01 [51.71, 56.25]99.08 [98.49, 99.62]53.52 [51.24, 55.78]
0.641.9786.52 [84.68, 88.17]97.19 [96.35, 97.97]84.09 [82.25, 85.85]
0.859.5695.74 [94.71, 96.68]96.33 [95.43, 97.18]92.22 [90.89, 93.50]

Thanks for the detailed response. My concern are addressed and I will raise my score to 3.

Comments on the paper

1 comment
Area_Chair_jdrJOpenReview ↗

Rebuttal ddl

Dear reviewers, the ICLR deadline is only 4 days away. The authors have already rebuttaled the issues raised. If you still have questions, please participate in the rebuttal.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.