acceptodds
Under review as a conference paper at ICLR 2027

RESET: Residual-Energy Subspace Expansion via Tokens for Efficient VLM Inference

Abstract

Vision-Language Models (VLMs) achieve strong visual understanding but incur substantial memory and inference costs due to long visual token sequences. While numerous token compression strategies have been explored, we revisit the problem from a representative subset selection perspective, asking which visual tokens can collectively preserve diverse information with minimal redundancy. Based on this view, we propose RESET, a training-free visual token compression framework based on residual subspace selection. For images, RESET iteratively selects tokens with the highest residual energy, favoring previously unexplained feature directions and reducing spatial redundancy. For videos, RESET constructs a compact historical reference via SVD and conditions current-frame selection on the subspace orthogonal to historical information, thereby suppressing temporal redundancy. RESET is text-independent, plug-and-play, and compatible with FlashAttention, enabling seamless integration into existing image and video inference pipelines without modifying the underlying attention mechanism. Extensive experiments across multiple VLMs and benchmarks show that RESET consistently outperforms strong state-of-the-art token compression methods, validating residual-subspace selection as a simple and effective principle for representative visual token compression.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
Token Compression, Efficient VLM Inference, Multimodal Model

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

This paper proposes RESET, a training-free framework for visual token compression in VLMs. The paper argues that conventional token selection methods primarily estimate the importance of individual tokens without explicitly considering whether candidate tokens provide complementary information beyond the tokens already selected. To address this issue, RESET formulates visual token compression as a residual subspace selection problem and develops a unified framework for image and video understanding. For image compression, RESET iteratively selects tokens with the largest residual energy using a QRCP-based procedure, encouraging the retained tokens to capture complementary feature directions. For video compression, it constructs a low-rank historical reference subspace to remove information already explained by previous visual content before selecting tokens from the residual representation. The framework also incorporates grouped parallel execution to reduce the computational overhead of sequential token selection. Experiments across multiple VLM architectures and image and video understanding benchmarks show that RESET achieves competitive performance. Furthermore, since the method does not require access to attention information, it can be integrated with methods such as FlashAttention.

Critical strengths

1. A novel perspective on token importance and a unified framework for image and video compression. RESET takes a different perspective on visual token importance by evaluating a token based on its residual energy after accounting for the information already captured by previously selected tokens, rather than assessing its importance in isolation. This allows the selection process to favor tokens that contribute complementary feature directions, providing a meaningful perspective on reducing redundancy in visual token compression. The paper further extends this principle to video by introducing a low-rank historical reference subspace to account for information already represented in previous frames. While the selection procedure builds on established techniques such as QRCP, the contribution lies in applying residual based selection as a common principle for image and video compression. This provides a coherent framework for addressing both spatial and temporal redundancy.

2. Evaluation across image and video benchmarks. The paper evaluates RESET on multiple VLM architectures and image and video understanding benchmarks. It achieves the highest average performance retention among the compared methods at three token budgets on LLaVA-1.5-7B, and outperforms FluxMem on four of five video benchmarks. Ablation studies examine the effects of historical reference and group size, while latency measurements evaluate the efficiency of the method. Although the results are not uniformly better across tasks, the experiments provide empirical support for RESET's effectiveness under the evaluated settings.

Critical weaknesses

1. The methodological novelty of the token selection strategy remains insufficiently isolated. RESET builds on classical QRCP-based residual selection, while prior work such as ApET also adopts a reconstruction-based perspective. The paper does provide a matched comparison between ApET and RESET under the same image-compression pipeline and token budgets, showing consistent improvements in average performance retention. However, the experiments do not isolate the effects of the key design choices in RESET, particularly the use of residual-energy-based pivot selection and the iterative residual updates that account for previously selected directions.

2. The computational trade-offs and cross-method comparability are not fully characterized. Although RESET and FluxMem are both evaluated at 1 fps, the broader comparisons in Table 2 involve heterogeneous frame-sampling configurations, making it difficult to attribute performance differences solely to the compression methods. In addition, while the paper reports compression and end-to-end latency under different group sizes and context lengths, it does not provide a component-wise breakdown of the compression cost, particularly the overhead attributable to iterative QRCP selection versus historical subspace construction and updates. Such profiling would help clarify the practical efficiency benefits of RESET, especially under mild compression, where the paper acknowledges that residual-selection overhead becomes more significant.

Questions for authors

1. Contribution of the QRCP-based selection strategy. Table 1 already compares RESET with the original ApET selector under the same image-compression pipeline and token budgets. Could the authors provide additional ablations that isolate the contributions of iterative residual updates and QRCP-based pivot selection, for example by comparing against alternative residual-based selection criteria under the same pipeline? This would help clarify which design choices account for the observed performance gains and what methodological advantage QRCP provides beyond existing reconstruction-based selection approaches.

2. Fairness of video comparisons and the computational cost of selection. Table 2 reports different or unspecified frame-sampling settings for several video baselines, which limits the comparability of cross-method results. Could the authors provide matched-setting comparisons where feasible, or clarify which baseline results were obtained under the same input-frame configurations as RESET? In addition, could the authors specify the operations included in the reported compression latency and provide a breakdown of the costs of QRCP selection, historical subspace construction and updates? This would help quantify the overhead attributable to residual selection and clarify the practical efficiency benefits of RESET under different compression ratios.

3. Preservation of fine-grained spatial information after compression. Although the paper reports aggregate MME and GQA results and video spatial-understanding subtasks, the impact of aggressive token compression on localization and grounding remains unclear. Could the authors provide spatial-reasoning subtask results or evaluate RESET on a localization or grounding benchmark under different token budgets? Additionally, are the original positional indices of retained tokens preserved after pruning, or reassigned according to the compressed sequence, particularly for architectures using M-RoPE? Clarifying these points would help assess the preservation of fine-grained spatial information.

6 comments

We thank the reviewer for highlighting the need to isolate the selection mechanism, improve the comparability of video results, and examine fine-grained spatial information. We provide additional selector ablations, matched video comparisons, component-level profiling, and grounding results below.

W1 & Q1: Contributions of Iterative Selection and Pivot Scoring

QRCP is a classical algorithm. RESET's contribution lies in using iterative residual selection with reference conditioning for image and causal video compression. Unlike a fixed reference used for scoring, RESET updates the remaining residuals after each selected token, accounting for directions already retained. In video, a compact historical reference additionally suppresses redundant information across frames within causal groups. To isolate these design choices, we compare three selectors within the same image-compression pipeline on MME using LLaVA-1.5-7B. Static Residual Top-K ranks tokens once without residual updates. Iterative Normalized Residual updates residuals but normalizes pivot scores by original feature energy. RESET uses iterative updates with absolute residual energy.

Token budgetStatic Top-KNormalized ResidualRESET
1921582.991787.711765.30
1281204.201789.261768.30
64917.861665.391700.70

The large gap over Static Top-K highlights the value of iterative residual updates. Both absolute and normalized residual scoring identify complementary information, with comparable results overall. Neither feature magnitude nor normalized novelty is guaranteed to align with downstream relevance. We therefore adopt the simpler absolute residual criterion, which also performs best at the tightest budget.

W2 & Q2: Matched Video Comparisons and Compression Cost

Due to the limited rebuttal time, we focus our additional matched comparisons on the Real-Time tasks of OVO-Bench. We rerun CausalMem and StreamingTOM with the same Qwen2.5-VL-7B model, 1 fps video sampling, dynamic resizing, prompts, and generation settings as RESET. Their hyperparameters are adjusted to achieve similar overall visual-token retention. Actual retention rates are reported below.

RetentionMethodSTUOJRATRACROCRFPDAvg.
8.42%RESET58.9968.4874.1466.9792.6274.2672.58
8.35%StreamingTOM51.1255.4358.6255.9671.8168.3260.21
8.42%CausalMem57.3062.5072.4167.8991.9570.3070.39
11.45%RESET56.1867.3973.2863.3091.9576.2471.39
11.29%StreamingTOM48.8858.7059.4855.0572.4872.2861.14
11.45%CausalMem55.6261.9673.2867.8991.2867.3369.56
18.85%RESET56.7462.5070.6961.4787.2575.2568.98
18.82%StreamingTOM49.4459.2462.0755.0575.1770.3061.88
18.85%CausalMem56.1860.8768.9766.0691.2869.3168.78

RESET obtains the highest average accuracy in all three matched settings.

We further profile compression on a single NVIDIA A800 using three OVO-Bench clips sampled at 512, 1024, and 2048 frames. We report the costs of historical projection, iterative QRCP selection, historical subspace construction and updates, token merging, and other operations. The component-level results are summarized below.

KeepFramesRetentionProjectionQRCPHistory BuildHistory UpdateMergeOther Ops.CompressionEnd-to-end (E2E)Speedup
6.25%5123.69%36.3439.6750.60166.8410.2678.44382.242052.981.53×
6.25%10242.63%72.0376.15102.72338.4521.31156.93767.703317.401.89×
6.25%20482.09%144.98154.65208.60691.4143.99320.221563.985867.252.43×
12.5%5124.47%35.8073.7661.18204.3510.0476.14461.362152.071.46×
12.5%10243.02%72.57148.73125.25417.4621.47156.74942.313502.561.79×
12.5%20482.29%146.81301.23253.62843.5444.48319.961909.756207.292.29×
25%51210.06%36.68173.4488.96300.8010.5979.13689.712418.971.30×
25%10248.15%73.41346.55181.57612.0722.03159.751395.484080.551.54×
25%20487.20%147.03693.59366.651225.3045.15319.712797.548004.901.78×

All latency values are reported in milliseconds (ms). Other Ops. includes selection preparation, postprocessing, and remaining compression operations. The results show that all computation overhead increase with the token budget and video length. RESET achieves consistent E2E acceleration across all evaluated settings, with greater speedups on longer videos.

Q3: Grounding Performance and Positional Preservation

Spatial Understanding results are already reported in Table 7. To further assess localization, we evaluate RESET and ApET on RefCOCO TestB using Qwen2.5-VL-7B, with matched inputs and generation settings. We report Accuracy@IoU ≥ 0.5.

Token retentionRESETApET
20%41.0222.87
10%16.115.14
5%6.481.02

The uncompressed baseline scores 86.34. RESET consistently outperforms ApET, but both methods exhibit substantial degradation relative to the full input. These results reveal an important limitation of training-free token compression for fine-grained localization under aggressive budgets, which we will acknowledge in the revision.

Regarding positional information, RESET preserves the original spatial and temporal indices of retained tokens, including their M-RoPE coordinates, instead of reassigning positions based on the compressed sequence. Merged features inherit their retained anchor's coordinates. This maintains positional consistency for the remaining representations, although information from removed locations may still be lost.

Thank you for the detailed responses and additional experiments. I appreciate the efforts in addressing my previous concerns. Regarding the RefCOCO evaluation, could the authors additionally report grounding performance at 70% and 50% token retention? Furthermore, could the authors provide performance-versus-pruning-ratio and speedup-versus-pruning-ratio curves by combining these additional results with the existing results (20%, 10%, and 5% retention)? The speedup should account for the overhead introduced by RESET. These results would help better understand the trade-off between localization performance and compression efficiency, and clarify the practical applicability of the proposed method.

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

This paper proposes RESET, it is a training-free visual token compression framework based on residual subspace selection for both image and video understanding The design is motivated by the token importance should be measured by the additional information it adds compared with the already existing content. The results show improvement on image understanding across multiple VLMs and video understanding tasks.

Critical strengths
  • The paper is well-written and easy to follow.
  • The paper provides extensive experiments across multiple VLM backbones, including different image and video understanding benchmarks, covering different compression ratios and both online and offline video settings.
  • The paper provides comprehensive ablations and evaluate the contribution of different components.
Critical weaknesses
  • The token selection strategy is based on classical QR decomposition, while residual-based reconstruction has already been explored in prior work. The main novelty seems to be adapting and combining existing approaches rather than introducing a new compression mechanism, while this itself does not necessarily diminish the impact of this work.
  • For long term temporal memory in the long video sequence setting, it may discard information important for future queries since the historical reference compresses previous information into a bounded space.
  • One assumption that motivates the method design is that tokens with high residual energy contain important information that's not captured by previous tokens, yet high residual energy does not necessarily correspond to the relevance to downstream questions. It would be great if the paper can provide more analysis into whether the proposed method preserves important information and how's that correlate to the feature space residual energy.
Questions for authors
  • What is the end-to-end speedup over existing token compression methods under the same conditions?
  • For long video sequences, will the previously observed but important information eventually be discarded and how to handle this type of situation?
2 comments

We thank the reviewer for acknowledging the clarity of our presentation, comprehensive evaluation, and ablation studies. We clarify the methodological contributions and limitations of RESET below and provide additional end-to-end efficiency comparisons.

W1: Methodological Novelty Beyond Classical QR Selection

We agree that QR decomposition and reconstruction-based token selection are established techniques. Our contribution is not a new QR algorithm, but a training-free framework that selects tokens according to their incremental contribution to the retained feature subspace. For video, this selection is conditioned on a recursively updated historical reference, enabling causal redundancy reduction with bounded memory. Adaptive rank estimation and grouped parallelization further improve its practicality for streaming inputs. We will distinguish these contributions more clearly from the classical QRCP operator in the revised paper.

W2: Information Preservation in Bounded Historical Memory

A bounded historical representation cannot preserve every detail of an arbitrarily long video. RESET recursively updates a compact feature basis so that previously represented information can guide subsequent token selection even after the original tokens are discarded. However, compression and group boundaries can cause information loss, and the historical reference is not a permanent representation of the entire video. We will clarify this limitation and the intended scope of the memory mechanism.

W3: Residual Energy and Downstream Relevance

We agree that residual energy reflects feature novelty rather than direct relevance to downstream questions. To examine this relationship, we analyze reconstruction error and downstream accuracy on StreamingBench. For Counting task, increasing the retention ratio from 6.25% to 25% reduces reconstruction error from 0.7652 to 0.4589 and improves accuracy from 44.68% to 50.00%, suggesting that preserving more feature information benefits counting tasks. The historical projection visualizations show that suppressing residual energy may weaken evidence for small but important objects, while historical conditioning can also improve understanding in certain cases. The reconstruction error analysis and visual examples reveal that lower reconstruction error does not always consistently improve downstream accuracy. These findings indicate that preserving complementary feature information generally benefits visual understanding but cannot guarantee the retention of all question-relevant details. Nevertheless, RESET's training-free and query-agnostic design enables effective redundancy reduction without task-specific supervision or prior knowledge of future queries, making it particularly suitable for general-purpose and streaming visual compression.

Q1: End-to-End Efficiency Compared with Existing Methods

We compare RESET with CausalMem [1] and StreamingTOM [2] for video, and ZOO-Prune [3] and AgilePruner [4] for images. All tests use Qwen2.5-VL-7B on a single NVIDIA A800 GPU. End-to-end latency includes visual encoding, compression, LLM prefill, and generation. For video, we evaluate three clips from OVO-Bench, sampled uniformly at 1 fps to obtain 512, 1024, and 2048 frames. Baseline hyperparameters are adjusted to match RESET's overall visual-token retention as closely as possible. Each table entry reports measured retention (%) and speedup (×) relative to the uncompressed control in that method's backend.

RESET per-stage retention ratioMethod512 frames1024 frames2048 frames
6.25%RESET3.69% / 1.52×2.63% / 1.89×2.09% / 2.43×
CausalMem3.69% / 1.45×2.63% / 1.68×2.09% / 2.06×
StreamingTOM3.13% / 1.22×3.13% / 1.40×3.13% / 1.69×
12.5%RESET4.47% / 1.47×3.02% / 1.80×2.29% / 2.30×
CausalMem4.47% / 1.44×3.02% / 1.68×2.29% / 2.06×
StreamingTOM3.13% / 1.22×3.13% / 1.41×3.13% / 1.70×
25%RESET10.06% / 1.30×8.15% / 1.55×7.20% / 1.94×
CausalMem10.06% / 1.39×8.15% / 1.60×7.20% / 1.89×
StreamingTOM9.38% / 1.19×9.38% / 1.32×6.25% / 1.63×

CausalMem matches RESET's measured retention, while StreamingTOM's discrete budgets allow only approximate matching. RESET achieves a backend-matched speedup of 1.30× to 2.43× and shows increasing gains for longer videos. It achieves larger speedups than CausalMem in most settings and than StreamingTOM in all settings.

For images, we evaluate the same 64 examples from MME with approximately 10% visual-token retention.

MethodE2E latency (ms) ↓
RESET138.41
ZOO-Prune142.61
AgilePruner143.86

RESET achieves the lowest measured end-to-end latency among these image compression methods, demonstrating that its selection overhead remains competitive in practice.

Q2: Loss of Earlier Information and Possible Mitigation

Under a fixed memory budget, previously observed details can eventually be discarded, even when they may be important to an unknown future question. RESET prioritizes complementary information without requiring future queries, but does not guarantee exact long-term recall. For applications requiring retrieval of specific past details, a separate archive and retrieval mechanism could complement the compact historical reference. We regard this as a possible extension of the current method.

[1] Song, Baiyang, et al. "Towards a Dynamic and Fixed-budget Memory Bank for Efficient Streaming Video Understanding." arXiv preprint arXiv:2606.25658 (2026).

[2] Chen, Xueyi, et al. "Streamingtom: Streaming token compression for efficient video understanding." Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. 2026.

[3] Kim, Youngeun, et al. "Zoo-prune: training-free token pruning via zeroth-order gradient estimation in vision-language models." Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. 2026.

[4] Baek, Changwoo, et al. "AgilePruner: An empirical study of attention and diversity for adaptive visual token pruning in large vision-language models." arXiv preprint arXiv:2603.01236 (2026).

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

RESET is a training-free visual token compression method that selects tokens with the highest residual energy relative to retained features. It extends this approach to videos using a historical subspace and shows improved performance retention across image and video benchmarks.

Critical strengths

The method is simple and practically useful. It requires no additional training or text-conditioned importance estimation and can be integrated without changing the attention mechanism. Selecting tokens according to what remains unexplained by earlier selections is a reasonable way to avoid spending a limited budget on redundant features. The comparison with ApET is particularly useful because the surrounding image compression pipeline is kept the same.

The results support its usefulness under aggressive compression. On LLaVA-1.5-7B, RESET retains 97.61% of average relative performance at the 64-token average budget, compared with 95.64% for ApET. Improvements also extend to other image backbones and several video benchmarks. The historical-reference ablation and grouped execution experiments provide further evidence that the approach can be useful in practice, although the gains are not uniform across tasks.

Critical weaknesses

The evaluation does not sufficiently cover information-dense visual inputs. Much of the motivation assumes that visual tokens contain substantial redundancy, but this assumption is less clear for dense document pages containing text, numbers, formulas, and tables. Similar-looking tokens can carry distinct information that cannot safely be discarded. TextVQA provides some evidence for text-related understanding, but does not establish whether the method preserves enough information for dense document VQA, multi-page reasoning, or complete document parsing.

Residual energy and feature reconstruction are useful selection criteria, but neither directly measures task importance. Small changes to a digit, punctuation mark, or object state may have low residual energy while being essential to the answer. Likewise, representing the feature space well does not necessarily preserve object counts, spatial relationships, or temporal order. The Counting regression on StreamingBench, from 52.1 with FluxMem to 45.2 with RESET, suggests that these limitations deserve closer analysis.

Questions for authors
  1. Could the authors evaluate RESET on information-dense tasks such as document VQA and multi-page document VQA, reporting performance across compression ratios? My main concern is whether compression discards distinct information needed to answer questions about dense text, numbers, and tables. If additional results address this concern, I would be willing to raise my rating.

  2. How does residual-energy selection handle small but important visual differences, such as a changed digit, a decimal point, or a subtle object-state change? In videos, could historical projection suppress these changes because most of the current features are already explained by previous frames? Targeted experiments or failure examples would help assess this risk.

  3. Does lower feature reconstruction error reliably correspond to better downstream understanding, particularly for counting, spatial relations, and temporal order? Please examine this relationship and explain the Counting drop relative to FluxMem. This would help clarify what information the selected subspace preserves and what it may lose.

3 comments

We thank the reviewer for the constructive feedback. We have conducted additional experiments on document understanding, historical projection, and the relationship between reconstruction error and downstream performance to clarify the strengths and limitations of RESET.

W1 & Q1: Performance on Information-Dense Documents

We evaluate RESET on DocVQA-Lite, a subset of DocVQA, using LLaVA-1.5-7B. We compare against ApET and the uncompressed model on the same questions under three token budgets. Results are reported as ANLS × 100.

Token budgetVanilla (576)ApETRESET
192 (33.3%)28.8720.4225.17
128 (22.2%)28.8719.5022.38
64 (11.1%)28.8716.3520.77

RESET consistently outperforms ApET across all budgets, although it remains below the uncompressed model. Document visualizations and failure cases demonstrate successful recognition of small text, numerical entries, and name-column associations. We also identify failures involving exact numbers, names, and table headings, indicating that spatial coverage alone does not guarantee information preservation. These results support RESET's effectiveness on dense document VQA, while acknowledging that fine-grained information loss remains a limitation.

W2 & Q2: Sensitivity to Subtle Changes and Historical Projection

RESET selects tokens based on residual energy after removing components explained by historical features. Small but critical changes, such as digits or object states, may receive low residual scores and consequently be suppressed. To examine this effect, we conduct a controlled ablation with and without historical projection on the Counting and Spatial Understanding tasks from StreamingBench, using Qwen2.5-VL-7B. Results are reported as accuracy (%).

Retention ratioCounting w/ HistoryCounting w/o HistorySpatial w/ HistorySpatial w/o History
6.25%44.6844.1571.5473.17
12.5%45.7446.2871.9572.36
25%50.0050.5366.6766.26

The results show modest but mixed effects from historical projection. Historical projection visualizations illustrate both outcomes. In one flag-counting example, the relevant region retains only 34.4% of its original energy after projection, leading to an undercount that disappears without history. Conversely, historical projection helps resolve a left-right spatial relationship in another example. We acknowledge that historical projection can suppress task-critical details in individual cases, while remaining useful for reducing temporal redundancy and improving selected predictions.

W2 & Q3: Reconstruction Error and Downstream Understanding

Lower feature reconstruction error does not necessarily guarantee better downstream understanding. RESET preserves representative feature directions without explicitly optimizing for object counts, spatial relations, or temporal order. We further analyze reconstruction error using the subspace spanned by the actual delivered tokens after merging, without including historical reference vectors. The following results use RESET with history enabled.

Retention ratioCounting error ↓Counting accuracy (%) ↑Spatial error ↓Spatial accuracy (%) ↑
6.25%0.765244.680.774671.54
12.5%0.629845.740.653671.95
25%0.458950.000.488366.67

For Counting, increasing the token budget reduces reconstruction error and improves accuracy from 44.68% to 50.00%. However, Spatial Understanding drops from 71.54% to 66.67% despite a substantial reduction in reconstruction error. This confirms that feature reconstruction quality is not a reliable monotonic predictor of task performance. Error-accuracy analysis and video visualizations further illustrate how larger budgets can recover information needed to distinguish objects across frames and identify spatial relationships. Regarding the reported Counting regression relative to FluxMem, our findings suggest that aggressive compression can weaken distinctions between similar or repeated instances. Increasing the token budget mitigates this limitation by reducing reconstruction error and improving Counting accuracy.

Thanks for your response. My major concerns are solved, and I will retain my original rating.

We are pleased to know our response has clarified your concerns. Thank you once again for leaving a thoughtful comment and for dedicating your time and effort to reviewing our work.

Comments on the paper

2 comments
Area_Chair_LEXbOpenReview ↗

With the authors' rebuttal now posted, I would ask each reviewer to respond in the discussion thread.

This is a borderline paper. Now that the authors' rebuttal has been posted, I ask each reviewer to respond in the discussion thread—especially on whether your original concerns have been addressed or remain unresolved—so that our discussion can meaningfully inform the final recommendation before the discussion period closes.

Summary of Rebuttal Discussion and Reviewer Feedback

To facilitate the final assessment, we provide a brief summary of our rebuttal discussions, highlighting the additional experiments, clarifications, and reviewer feedback.

1. Reviewer zZpD (Weak Accept, Major Concerns Resolved)

The reviewer requested additional evaluations on information-dense tasks and case studies. We provided detailed experimental results and case analyses to address these concerns. The reviewer maintained their Weak Accept rating.

2. Reviewer ZLFg (Weak Reject, Awaiting Response)

The reviewer raised questions regarding our methodological design and end-to-end efficiency compared with existing methods. We provided detailed explanations of the proposed mechanisms and conducted additional latency comparisons against representative image and video compression baselines under matched experimental settings. The reviewer has not yet responded to our rebuttal.

3. Reviewer eexf (Weak Reject, No Further Questions)

We addressed the reviewer's concerns regarding the contributions of iterative selection and pivot scoring, performance comparisons under matched token budgets, and the accuracy-efficiency experiments on the RefCOCO grounding benchmark. We also provided additional experimental results and analyses. The reviewer acknowledged our responses without raising further questions, although the rating remained unchanged.

Conclusion

We appreciate the reviewers' constructive feedback. Throughout the rebuttal, we have provided detailed clarifications and substantial additional experimental evidence to address the concerns raised by all three reviewers. We hope these responses, together with the explicit acknowledgment from Reviewer zZpD, will be taken into consideration in the final assessment.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.