acceptodds
Under review as a conference paper at ICLR 2027

Faster and Better Reference-Based Generation: Early Token Reduction via Downsampled Self-Probing

Abstract

Recent generic reference-based image generation models natively support multiple conditions by concatenating all reference image and text tokens into a joint sequence processed by a single Diffusion Transformer (DiT). However, this versatile paradigm comes at an efficiency cost: accommodating multiple references leads to a token explosion, incurring massive computational complexity. Most acceleration methods operate inside the DiT, showing limited capability in reducing spatial redundancy and failing to alleviate peak VRAM. While some pre-model token reduction alternatives exist, they inevitably introduce external models. This raises a question: Is it possible to achieve acceleration and memory savings while preserving compatibility with in-model techniques and requiring no additional parameters? In this work, we realize this goal by proposing an early token reduction framework via downsampled self-probing. By executing a few denoising steps on spatially downsampled reference images, this cheap probing stage utilizes the DiT model itself to effectively and efficiently extract reference token importance. Using the probed importance, we apply a hierarchical reduction before the main denoising stage: we completely discard low-importance tokens, substitute mid-importance tokens with their downsampled counterparts to cheaply preserve global context, and keep high-importance tokens. Experiments across three base models and two benchmarks demonstrate that our early reduction accelerates inference significantly and mitigates peak GPU memory. Furthermore, we can improve generation quality by safely filtering out noisy, uninformative reference tokens, enabling the model to concentrate its limited attention capacity on genuinely useful tokens. We will release the code upon acceptance.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
DiT acceleration, image generation, image editing, token pruning

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_At5AOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper addresses the token explosion in generic multi-reference DiT models, where reference tokens dominate the joint sequence. The authors run a cheap probing pass of the generator itself on 2× downsampled inputs with a sparse timestep schedule, and score each reference token by the attention it receives from generative and prompt tokens. Before the main denoising, high-score 2×2 blocks are kept at full resolution, mid-score blocks are replaced by a single low-resolution token, and the rest are discarded. The method is training-free and needs no external model. Experiments on three base models and two benchmarks show better quality–speed trade-offs than ToMeSD and VisPruner, lower peak activation memory, and compatibility with in-model acceleration methods.

Critical strengths
  1. Well-motivated problem and a clean design. Reducing reference tokens before the main denoising, rather than inside the DiT, is an underexplored but practically important direction for multi-reference generation. It is also what makes the peak-VRAM reduction and the composability with in-model methods (Table 2) possible. Using the generator itself as a low-resolution probe avoids an external saliency model. The keep, substitute, discard tiers are simple but clearly motivated, and Table 3 shows the intermediate tier matters.

  2. Solid efficiency gains without quality loss. On FLUX.2-Klein, the fast setting gives roughly 2.3–2.5× end-to-end speedup and about halves activation VRAM while matching baseline quality. The slow setting improves average quality on both benchmarks (4.70→4.88 on MultiBanana, 6.18→6.27 on MACRO). The trend holds across three base models of different architectures, including the pixel-space HiDream-O1. This is what most supports my positive assessment.

Critical weaknesses

The paper's central claim is that early reduction is preferable to in-model reduction, yet the most relevant in-model method, ToPi (Lin et al., 2026), which prunes reference tokens inside the DiT for in-context generation, is only discussed in related work. The compared methods, ToMeSD and VisPruner, both reduce tokens before the model, and TransPrune in Table 2 is also a VLM method. Without comparing to ToPi at matched speedup, including peak VRAM, this central claim remains untested.

Questions for authors

Please address the concern raised under Critical Weaknesses.

3 comments

We sincerely thank the reviewer for the careful reading and the valuable suggestion. We have revised the manuscript accordingly (the blue text in the PDF) and respond below.

We implement and evaluate the in-model token reduction method ToPi [1] on FLUX.2-Klein-4B under the same MultiBanana evaluation protocol:


Table: Comparison with ToPi.

MethodOverall ↑Latency (s) ↓Speedup ↑VRAM ↓
Baseline4.705.741.00×15.3
Fast
ToMeSD4.482.622.19×11.2
VisPruner4.532.552.25×8.7
ToPi4.583.321.73×15.2
Ours4.722.502.30×7.3
Slow
ToMeSD4.573.461.66×10.7
VisPruner4.753.521.63×10.2
ToPi4.684.021.43×15.2
Ours4.883.431.67×9.8

Ours achieves better quality, faster inference, and lower peak activation VRAM than ToPi under both Fast and Slow settings, showing the advantage of our early reduction over the in-model reduction. The full per-dimension quality metrics are provided in Appendix C.2 (Table 11).

Compatibility with ToPi: Early token reduction and in-model token reduction do not compete with each other, as they act at different stages. Our method reduces the reference sequence before the main denoising stage, while ToPi can further progressively prune reference tokens inside the DiT. This allows the combination to remove additional reference-token redundancy after our early reduction.


Table: Compatibility between Ours and ToPi.

MethodOverall ↑Latency (s) ↓Speedup ↑VRAM ↓
Baseline4.705.741.00×15.3
Ours (Fast)4.722.502.30×7.3
Ours (Fast) + ToPi4.702.232.57×7.1
Ours (Slow)4.883.431.67×9.8
Ours (Slow) + ToPi4.863.071.87×9.5

Combining Ours with ToPi further improves speedup while maintaining comparable generation quality. The full per-dimension quality metrics are provided in Appendix C.2 (Table 12).

Implementation details: We reproduce ToPi for FLUX.2-Klein-4B based on its published implementation details [1], as the official code has not been released. The details are provided in Appendix C.2. Speedup alignment: Achieving a strictly exact matched-speedup is difficult because ToPi controls sparsity via a "cumulative token influence mass" threshold rather than explicit retention ratios. Coupled with its in-model progressive pruning nature, its end-to-end latency cannot be precisely controlled. Nevertheless, as shown in the table, even though ToPi operates at a lower compression rate (e.g., a lower speedup of 1.73× vs. our 2.30× in the Fast setting), it underperforms our method in quality.

[1] Lin et al., Token Pruning for In-Context Generation in Diffusion Transformers, arXiv:2602.01609, 2026.

Thank you for the detailed response and the additional experiments. The new comparison with ToPi, together with the compatibility results, addresses my concern. After considering the rebuttal, I maintain my original rating.

We sincerely thank the reviewer for the follow-up and are glad that our additional experiments have addressed your concern. We appreciate your time and consideration.

Reviewer_iG4iOpenReview ↗
Rating
3
Confidence
3
Informativeness
3
Presentation
3
Soundness
3
Read review
Paper summary

The authors identify efficiency bottlenecks, particularly inference latency, and potential generation-quality degradation in models that jointly process multiple reference images through self-attention. Inspired by token pruning, they propose a training-free method that uses low-resolution self-probing to estimate reference-token importance and selectively preserve, replace, or discard these tokens before the main denoising stage. Experiments across three base models and two benchmarks show faster inference, lower peak activation memory, and improved overall generation quality.

Critical strengths
  1. Clear and complete presentation. The paper is clearly written and provides a complete explanation of the approach. The figures are intuitive and help readers understand the method.

  2. Well-motivated problem and method. The problem identified by the authors is meaningful, and the proposed method follows logically from it. Using the model itself to identify and reduce less useful reference tokens is a sensible approach to improving efficiency.

  3. Thorough experiments. The experimental evaluation is extensive, covering multiple base models and benchmarks, comparisons with existing methods, ablations, and additional analyses. These experiments provide substantial support for the effectiveness of the proposed approach.

Critical weaknesses
  1. Evidence for attention interference. The paper argues that irrelevant reference tokens interfere with generation. Please support this explanation with more direct experiments or references to prior findings.

  2. Evaluation of adaptive retention budgets. Useful regions can occupy different proportions of different reference images. Please evaluate the adaptive retention strategy in Appendix B.2.2 against the default fixed-ratio strategy. Include cases with large and small reference subjects.

  3. Hyperparameter selection and sensitivity. Please explain the choice of the 15%/30% ratios for Fast and the 30%/60% ratios for Slow. The two settings change both ratios together. Please vary each ratio separately and report generation quality and inference cost.

  4. Qualitative comparison of two-tier and three-tier reduction. Table 3 already provides a quantitative comparison. Please also show side-by-side generation results and token retention/substitution maps at the same reduction ratio. This would clarify what context the substitute tier preserves and which details it improves.

  5. Figure readability. Some text is difficult to read at normal viewing size, especially the prompts and annotations in Figure 6 and the axes and legends in Figure 3. Please increase the font sizes and simplify the layouts or split the panels.

Questions for authors

see Weaknesses.

2 comments
AuthorsOpenReview ↗

This is Part 1/2 of our reply

We sincerely thank the reviewer for the detailed review. We have revised the manuscript accordingly (the green text in the PDF) and respond below.

Due to word count limitations, we will answer all your questions in two parts. Below is the first part.

W1. Evidence for attention interference

  • Aligning What Vision-Language Models See and Perceive with Adaptive Information Flow (CVPR 2026) reports that text tokens distribute too much attention to irrelevant visual tokens, leading to incorrect answers.
  • The two examples in the first row of Figure 6 and the example in Figure 7 also illustrate this issue: the baseline introduces unwanted content from distracting reference regions, whereas our reduction removes these artifacts in the generated results.
  • Quantitatively, Table 4 shows that the Effective Attention Ratio (EAR) increases from 43.6% for the baseline to 52.7% with our method.

W2. Evaluation of adaptive retention budgets

(a). Quantitative comparison

We compare the two decisions quantitatively by selecting thresholds that yield similar inference times:


Table: Comparison of two decision policies on MultiBanana (FLUX.2-Klein-4B).

SettingDecision policyOverall ↑Latency (s) ↓Speedup ↑VRAM ↓
FastDefault (Ratio-based decision)4.722.502.30×7.3
FastAdaptive ratio (Threshold-based decision)4.732.692.13×8.0
SlowDefault (Ratio-based decision)4.883.431.67×9.8
SlowAdaptive ratio (Threshold-based decision)4.843.101.85×10.3

Both decisions achieve similar trade-offs at the Fast and Slow settings. Full per-dimension results are provided in Appendix B.2.3 (Table 7).

(b). Explanation and visual examples for large/small useful regions

  • Reference-level (intra-sample) adaptivity (Figure 10(a)): Default (Ratio-based decision) already supports adaptive allocation across reference images within each sample. Specifically, p1p_1 and p2p_2 define a global token budget over all reference images, rather than independent fixed ratios for each reference. We rank reference-token importance globally, allowing references with larger useful regions to receive more tokens, subject to the minimum per-reference Full-Resolution retention ratio pmin⁡=23p1p_{\min}=\frac{2}{3}p_1. Figure 10(a) visualizes how the retained tokens are allocated across references within one sample.

  • Sample-level adaptivity (Figure 10(b)): Adaptive ratio (Threshold-based decision) in Appendix B.2.2 introduces another level of flexibility, allowing the total retained token budget to vary across samples according to their reference importance distributions. In contrast, Default (Ratio-based decision) uses a predetermined global budget per sample, subject to the per-reference minimum retention guarantee. This distinction becomes particularly relevant when the useful regions across the references of a sample are all large or small. Figure 10(b) illustrates both cases:

    • In the large-region example, the useful regions exceed the budget of Default (Ratio-based decision); Adaptive ratio (Threshold-based decision) retains more tokens (55.9% versus 22.9%), improving generation quality.
    • In the small-region example, Adaptive ratio (Threshold-based decision) reduces retention to 18.3%, compared with 24.6% for Default (Ratio-based decision), while maintaining the same generation quality.
AuthorsOpenReview ↗

This is Part 2/2 of our reply

W3. Hyperparameter selection and variations

(a). Selection

Our method is designed as a flexible framework that pushes the quality-speed Pareto frontier across a wide operating range (Figure 5), rather than proposing two specific variants. The Fast and Slow configurations are simply two representative operating points from this curve to illustrate fine-grained metrics in Table 1. Therefore, instead of tuning the ratios (p1p_1 and p2p_2) to find a singular optimal configuration, we adopted a simple proportional rule (p2=2p1p_2 = 2p_1): setting p1=15%p_1=15\% to obtain the Fast (15%/30%), and doubling both to obtain the Slow (30%/60%).

(b). Configuration (retention ratios) variations

We evaluated neighboring configurations by varying the Full-Resolution Tier ratio p1p_1 and Low-Resolution Tier ratio p2p_2 separately by ±5:


Table: Comparison of different Full-Resolution (p1p_1) and Low-Resolution (p2p_2) Tier ratios on MultiBanana (FLUX.2-Klein-4B).

SettingVariationp1p_1p2p_2Overall ↑Latency (s) ↓Speedup ↑VRAM ↓
Base model—100%0%4.705.741.00×15.3
Fastp1−5p_1-510%30%4.692.312.48×6.7
Fastp2−5p_2-515%25%4.682.412.38×7.1
FastDefault15%30%4.722.502.30×7.3
Fastp2+5p_2+515%35%4.762.502.30×7.3
Fastp1+5p_1+520%30%4.762.662.16×7.9
Slowp1−5p_1-525%60%4.833.151.82×9.1
Slowp2−5p_2-530%55%4.853.301.74×9.6
SlowDefault30%60%4.883.431.67×9.8
Slowp2+5p_2+530%65%4.843.401.69×9.9
Slowp1+5p_1+535%60%4.833.581.60×10.4

As shown in the table, varying these two ratios demonstrates expected and smooth shifts along our quality-speed spectrum. Across all tested configurations, the Slow variations consistently improve Overall quality (by 0.13–0.18 over the unpruned Base model), while the Fast variations stably match Base model (−0.02 to +0.06). This shows that our framework offers robust, continuous trade-off choices for different inference budgets, rather than requiring delicate tuning.

W4. Visual comparison of two-tier and three-tier reduction

We have added side-by-side generated results and token retention/substitution maps comparing No Low-Resolution Tier (two-tier) with Ours (three-tier) at a matched token budget, as presented in Appendix A.4 (Figure 9). In the illustrated sample, the three-tier strategy preserves broader mid-importance context through its substituted Low-Resolution Tier tokens, whereas the two-tier variant loses part of this context and shows lower reference consistency.

W5. Figure readability

Thanks for your feedback. With the additional page allowance during the rebuttal phase, we have adjusted / enlarged Figures 1, 3, 6, and 7.

Reviewer_sv2vOpenReview ↗
Rating
3(+1)
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper addresses the computational and memory costs of processing many reference tokens in multi-reference image generation. It proposes a training-free approach that uses the generator itself to estimate token importance through a short denoising probe at reduced resolution. The main denoising run then retains full-resolution tokens in high-importance regions, substitutes low-resolution tokens for moderately important regions, and discards low-importance regions. Experiments on three base models and two benchmarks report lower inference latency and peak activation memory, with higher average quality scores under selected configurations.

Critical strengths
  1. The method requires no additional training and uses the generator itself to estimate token importance. Its three-tier strategy preserves detail in important regions while retaining contextual information at lower resolution, without introducing a separate importance-prediction model.
  2. The evaluation covers three base models and two benchmarks. Table 3 includes ablations of the reduction strategy and shows that the complete method outperforms random selection. Table 2 also examines compatibility with several acceleration methods that operate within the model.
Critical weaknesses
  1. The reliability and scope of the reported quality gains need further clarification. All generation experiments use seed 42, quality is assessed by Qwen3-VL, and confidence intervals for the quality differences are not reported. In Table 1, the fast setting increases the average scores from 4.70 to 4.72 and from 6.18 to 6.20, while some individual metrics are slightly below the baseline. The slow setting achieves larger gains, but uncertainty estimates are also needed for these results. Figure 5 shows that quality can fall below the baseline under more aggressive compression. Claims of improved quality should therefore be limited to the supported settings and accompanied by an analysis of their reliability.

  2. The token budgets and evaluation sets in the ablation and analysis experiments need clearer reporting. Table 3 states that the reduction ratio is matched across variants. Before timestep-dependent adjustments, the fast configuration in Appendix B.2 implies retaining approximately 22.5% of the reference tokens, whereas uniform 2x spatial downsampling would ordinarily retain 25%. The actual budgets and any adjustments used to match them are not clear from the description. Table 4 is described as an evaluation on the object-composition subset, yet the quality scores in its Baseline and All layers rows match the full-benchmark Baseline and Slow results in Table 1, respectively. It is unclear whether EAR and quality are evaluated on the same samples. Clarifying these settings is necessary to assess how strongly the experiments support the individual design choices and the proposed explanation for the quality gains.

Questions for authors
  1. Could the authors report paired quality differences and confidence intervals for Table 1 and Figure 5? On a representative subset, does the fast setting's approximately 0.02 average improvement persist across additional generation seeds?
  2. What are the actual full-resolution and low-resolution token budgets for each variant in Table 3, and are timestep-dependent ratios enabled? How are the budgets of the "Only Low-Res Tier" and "No Low-Res Tier" variants matched to that of the complete method?
  3. Are the EAR and quality scores in Table 4 computed on the same object-composition subset? Please identify the evaluation set for each metric, clarify the relationship to the corresponding scores in Table 1, and specify which variants use a matched token-reduction budget.
2 comments
AuthorsOpenReview ↗

This is Part 1/2 of our reply

We sincerely thank the reviewer for the review. We have revised the manuscript accordingly (the orange part in pdf) and make the following response.

Due to word count limitations, we will answer all your questions in two parts. Below is the first part.

W1. Multi-seed runs and Claim clarification

(a). Clarification

We would like to clarify a potential misunderstanding regarding our evaluation design. Our primary goal is not to propose two specific variants ("Fast" and "Slow"), but to introduce a flexible early reduction framework that pushes the quality–speed Pareto frontier, consistently outperforming competing methods across practical operating ranges and even improving the quality of the base model at moderate compression levels (as evidenced by Figure 5).

The "Fast" and "Slow" configurations in Table 1 are simply two representative operating points from this curve to illustrate fine-grained metrics.

  • The Fast configuration is a point that safely matches the unpruned Baseline at a high compression rate (>2x speedup) rather than consistently outperforming it. While slight fluctuations in certain individual dimensions are expected, per-dimension metrics are largely preserved, yielding an overall generation quality that is closely matched with the Baseline.
  • The Slow configuration is a point that shows consistent per-dimension quality gains at moderate speedups (>1.5×).

To validate that the performance is statistically stable and does not depend on a particular seed or VLM judge, we have added three-seed generation runs (see (b) below) and evaluation using Gemini (see (c) below).

Thanks for the feedback. We have clarified the relevant claims in the paper.

(b). Multi-seed runs

We conduct three generation runs in total for the Baseline (Base model), ours, and the most competitive competing method VisPruner. The mean and standard deviation (Mean ± Std), together with 95% confidence intervals (CIs), are reported below.


Table: Three-seed runs (Qwen3-VL judge).

MethodInst. Align ↑Ref. Cons. ↑B-S Match. ↑Phys. Real. ↑Visu. Qual. ↑Overall ↑Δ vs. Baseline (95% CI) ↑Speedup ↑VRAM ↓
Baseline4.61 ±\pm 0.0324.85 ±\pm 0.0463.95 ±\pm 0.0264.39 ±\pm 0.0465.76 ±\pm 0.0404.71 ±\pm 0.026—1.00 ±\pm 0.00815.25 ±\pm 0.00
VisPruner (Fast)4.56 ±\pm 0.0574.57 ±\pm 0.0533.72 ±\pm 0.0204.28 ±\pm 0.0405.68 ±\pm 0.0874.56 ±\pm 0.049−0.147 [−0.248, −0.045]2.20 ±\pm 0.0448.74 ±\pm 0.00
Ours (Fast)4.66 ±\pm 0.0254.83 ±\pm 0.0213.94 ±\pm 0.0424.36 ±\pm 0.0215.77 ±\pm 0.0354.71 ±\pm 0.021+0.003 [−0.052, +0.058]2.31 ±\pm 0.0147.25 ±\pm 0.09
VisPruner (Slow)4.67 ±\pm 0.0104.85 ±\pm 0.0153.88 ±\pm 0.0214.40 ±\pm 0.0325.81 ±\pm 0.0754.72 ±\pm 0.030+0.010 [−0.055, +0.075]1.61 ±\pm 0.02010.18 ±\pm 0.00
Ours (Slow)4.84 ±\pm 0.0205.03 ±\pm 0.0234.06 ±\pm 0.0524.55 ±\pm 0.0365.92 ±\pm 0.0554.88 ±\pm 0.030+0.170 [+0.105, +0.235]1.69 ±\pm 0.0149.72 ±\pm 0.14

The Slow setting prioritizes quality, achieving 4.88 ± 0.030, consistently exceeding the Baseline's 4.71 ± 0.026 across the three seeds. For the Fast setting, per-dimension metrics are largely preserved, yielding an overall generation quality that is closely matched with the Baseline. We have updated Table 1 and Figure 5 accordingly. The per-seed metrics are provided in Appendix C.1.1 (Table 8).

(c). Evaluation using another VLM on multi-seed runs

We additionally evaluate generation quality using another VLM, Gemini 3.1 Flash-Lite, and obtain results consistent with the findings in (a) and (b).


Table: Three-seed runs (Gemini 3.1 Flash-Lite judge).

MethodInst. Align ↑Ref. Cons. ↑B-S Match. ↑Phys. Real. ↑Visu. Qual. ↑Overall ↑
Baseline3.41 ±\pm 0.0853.32 ±\pm 0.0813.19 ±\pm 0.0253.27 ±\pm 0.0383.88 ±\pm 0.0233.41 ±\pm 0.051
VisPruner (Fast)3.45 ±\pm 0.0463.13 ±\pm 0.0323.18 ±\pm 0.0253.31 ±\pm 0.0363.91 ±\pm 0.0553.40 ±\pm 0.035
Ours (Fast)3.55 ±\pm 0.0493.32 ±\pm 0.0063.17 ±\pm 0.0623.27 ±\pm 0.0383.89 ±\pm 0.0323.44 ±\pm 0.035
VisPruner (Slow)3.48 ±\pm 0.0513.34 ±\pm 0.0453.17 ±\pm 0.0233.27 ±\pm 0.0173.85 ±\pm 0.0233.43 ±\pm 0.023
Ours (Slow)3.55 ±\pm 0.0533.43 ±\pm 0.0153.20 ±\pm 0.0403.33 ±\pm 0.0403.94 ±\pm 0.0253.49 ±\pm 0.031

Per-seed results are included in Appendix C.1.2 (Table 10).

AuthorsOpenReview ↗

This is Part 2/2 of our reply

W2. Implementation details in Tables 3 and 4

W2.1. Token-budget matching in Table 3

We apologize for the imprecise description of token-budget matching. The three reduction variants other than Only Low-Res Tier use matched reference-token budgets. At the final timestep, the three matched variants retain approximately 22.5% of the original reference tokens. Only Low-Res Tier directly performs 2× spatial downsampling and thus retains 25% of the original reference tokens. Despite retaining more tokens than ours, Only Low-Res Tier achieves lower quality, supporting the benefit of our tiered reduction.


Table: Token retention budgets at the final timestep.

VariantFull-Resolution tokensLow-Resolution tokensTotal
Only Low-Res Tier0%25%25%
No Low-Resolution Tier22.5%0%22.5%
Random Reduction15%7.5%22.5%
Ours15%7.5%22.5%

Timestep-dependent ratio adjustments are enabled and the retained budgets at the earliest and final timesteps are shown below.


Table: Token retention budgets across timesteps.

VariantEarliest timestepFinal timestep
Only Low-Res Tier25%25%
No Low-Resolution Tier21%22.5%
Random Reduction21%22.5%
Ours21%22.5%

W2.2. Evaluation sets and budgets in Table 4

The quality scores in Table 4 are computed over the same samples as Table 1 and Figure 5, whereas EAR is calculated on the object composition subset, as ground-truth reference importance is only defined for this task. Except for the unpruned Baseline, the other three variants (Top-1 layer, Top 20% of layers, and All layers) use matched token-reduction budgets.

We have clarified all these distinctions in the revised paper (Appendix B.2.1 and B.5).

Comments on the paper

2 comments
AuthorsOpenReview ↗

General Response & Summary of Rebuttal Updates

Dear Area Chair and Reviewers,

We thank all reviewers for their detailed feedback! We appreciate their recognition of our strengths:

  • Well-motivated and clean design.
  • Thorough experiments.
  • Clear and complete presentation.

We have responded to all concerns and have uploaded the revised manuscript with all changes highlighted in corresponding colors. To facilitate your review, we summarize the major updates below:

  • Clarification & multi-seed runs (Addressing sv2v): We clarified that our method is a flexible framework that pushes the quality-speed Pareto frontier (Figure 5), rather than proposing two specific variants. The "Fast" and "Slow" configurations simply illustrate two representative operating points: "Fast" closely matches the quality of Base Model at a high speedup (>2×), while "Slow" yields consistent quality gains at moderate speedups (1.40×–1.83×). To validate the statistical stability of our performance across generation seeds and VLM judges, we added 3-seed generation runs (Table 1, Appendix C.1.1) and an independent Gemini VLM evaluation (Appendix C.1.2).

  • Comparison & compatibility with the in-model method ToPi (Addressing At5A): New experiments show that our method consistently outperforms ToPi in quality and efficiency, and remains compatible with it (Appendix C.2).

  • Additional ablations / analysis (Addressing iG4i):

    • Compared "Adaptive (threshold-based)" and "Default (ratio-based)" decision policies quantitatively (Appendix B.2.3).
    • Swept $\pm$5 p1p_1 and p2p_2 configuration variations (Appendix C.3) to demonstrate smooth shifts along our quality-speed spectrum.
    • Added visual comparisons between the 2-tier and 3-tier reduction strategies (Appendix A.4, Fig. 9).
  • Implementation detail clarifications (Addressing sv2v): We explicitly clarified the fair token budgets and evaluation subsets for Tables 3 and 4 (Appendices B.2.1 and B.5).

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. With the high number of submissions this year, your input matters more than ever.

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.