acceptodds
Under review as a conference paper at ICLR 2027

Still There, No Longer Seen: Exposing Compression-Induced Risk In Large Vision-Language Models

Abstract

Visual token compression reduces the inference cost of Large Vision-Language Models (LVLMs). However, aggregate robustness measures do not reveal whether a particular adversarial failure is induced by compression or inherited from the underlying model. We define a compression-specific failure (CSF) as an adversarial input that remains correct under full-token inference but fails after compression, casting compression-induced risk as a paired failure attribution problem. Within a controlled diagnostic cohort, counterfactuals show that retained-set allocation causally changes compressed correctness and reveal a negative association between recovery and representation drift in displaced evidence. Motivated by these findings, we propose CIRA, a Compression-Induced Risk Attack for Large Vision-Language Models. Under a vision-encoder white-box setting, CIRA optimizes image perturbations through encoder-side objectives that manipulate token priorities across candidate compression budgets while preserving displaced evidence. CIRA uses no downstream questions or labels and requires no access to the language model, deployed compressor, or exact compression budget. Across 12 dataset–compressor settings evaluated at four budgets, CIRA achieves a mean CSFR of 20.35% while limiting full-token attack success to 6.92%, with similar behavior on additional LVLM families. A cross-view selection-stabilization defense substantially suppresses CIRA, although Adaptive CIRA partially restores its effectiveness. These results show that compression-specific failures persist under restricted access and support paired evaluation of full-token and compressed inference for attributing risk to visual-token compression.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
large vision-language models, visual token compression, adversarial robustness

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_5pKPOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

This paper investigates adversarial vulnerabilities introduced by visual token compression in large vision-language models. It defines compression-specific failures (CSFs) as cases where full-token inference remains correct while compressed inference fails. Through controlled counterfactual analyses, the authors examine how token selection and representation drift contribute to these failures. Based on these findings, they propose CIRA, an encoder-only attack that combines Global Selection Hijacking (GSH) and Hidden-Evidence Preservation (HEP) to manipulate token priorities while preserving critical visual information. Experiments across multiple datasets, compressors, and models demonstrate its effectiveness under unknown compression configurations. The paper also introduces a cross-view selection-stabilization defense and evaluates its robustness against adaptive attacks.

Critical strengths
  1. The clean-conditioned, paired CSF evaluation and controlled retained-set interventions provide a clear framework for analyzing failures specifically associated with visual token compression.
  2. The combination of global priority reallocation and selective evidence preservation is well motivated, and the experiments demonstrate promising attack selectivity across multiple compressors, budgets, and LVLM architectures.
Critical weaknesses
  1. The novelty is somewhat limited, as compression-specific failures and token-ranking manipulation were previously investigated in [1, 2], while encoder-only attacks and probabilistic modeling of unknown compression budgets were explored in [3, 4]; the main novelty therefore lies in the particular combination of GSH and HEP.
  2. The mechanistic counterfactual analysis is restricted to LLaVA–VisionZip and relies on a post-hoc answer-aware oracle, leaving unclear whether the observed relationship between retained-set allocation, representation drift, and correctness generalizes to other compression mechanisms.
  3. The attack relies on encoder-side proxy rankings that may differ substantially from actual token-selection decisions, particularly for merging-based compressors and LLM-attention-based methods such as FastV, but this proxy mismatch is not systematically quantified.

[1] Sun et al., Does Acceleration Cause Hidden Instability in Vision Language Models? Uncovering Instance-Level Divergence Through a Large-Scale Empirical Study.

[2] Zhang et al., Less Is More — Until It Breaks: Security Pitfalls of Vision Token Compression in Large Vision-Language Models.

[3] Zhang et al., On the Adversarial Robustness of Large Vision-Language Models under Visual Token Compression.

[4] Mei et al., VEAttack: Downstream-agnostic Vision Encoder Attack against Large Vision Language Models.

Questions for authors
  1. Can the retained-set counterfactual analysis and TCS defense be extended to compressors beyond VisionZip?
  2. Does HEP consistently outperform selective perturbation [1] and simpler feature-preservation baselines under matched optimization settings?
  3. Can the authors quantify the agreement between encoder-side proxy rankings and actual retained-token selections across different compressors, and relate this agreement to CSFR?

[1] Zhang et al., Less Is More — Until It Breaks: Security Pitfalls of Vision Token Compression in Large Vision-Language Models.

4 comments
AuthorsOpenReview ↗

[1/2] Response to Reviewer 5pKP

We thank the reviewer for the positive assessment and constructive questions.

Question 1 & Weakness 2 The mechanistic counterfactual analysis is restricted to LLaVA–VisionZip and relies on a post-hoc answer-aware oracle, leaving unclear whether the observed relationship between retained-set allocation, representation drift, and correctness generalizes to other compression mechanisms. Can the retained-set counterfactual analysis and TCS defense be extended to compressors beyond VisionZip?

Thank you for this question! We extended both to VisPruner, PruMerge and FastV. The answer-aware oracle selects post-hoc diagnostic exchanges at fixed failure states; CIRA's encoder-only optimization uses no reference answers.

  • Cross-compressor diagnostics.

    • We evaluate LLaVA-1.5-7B on three benchmarks and K = 32, 64, 128 and 192. On 2,269 first-CSF observations, we fix adversarial features and budgets, exchange about 7.5% of direct/representative slots, and rebuild native compression.
    • Factorial controls compare incoming tokens with high reference-answer support against random incoming tokens, under both random and low-support outgoing choices.
    • All four compressors show positive mean restoration effects and negative pooled within-setting drift-recovery slopes.
    CompressorRestoration effect (pp)Drift slope (pp/0.1)Paired CSFR Native / TCS (%)
    VisionZip+14.33-3.3518.45 / 2.70
    VisPruner+16.42-7.1314.30 / 4.76
    PruMerge+20.31-7.8532.15 / 9.89
    FastV+27.58-10.2013.88 / 6.65
    • Restoration effects and CSFR average 12 dataset-budget settings per compressor. Drift slopes pool within-setting variation and report percentage points per 0.1 increase in drift.
  • TCS defense.

    • TCS uses fixed standard-CIRA adversarial images and 1,000 image–question pairs per benchmark. Paired CSFR uses the common cohort clean-correct under full-token, native-compressed and TCS-compressed inference.
    • TCS lowers paired CSFR in all 48 settings, reducing per-compressor mean CSFR by 52.1%–85.4%.

Question 2 & Weakness 1 Compression-specific failures and token-ranking manipulation were previously investigated, while encoder-only attacks and probabilistic modeling of unknown compression budgets were also explored; the main novelty therefore lies in the particular combination of GSH and HEP. Does HEP consistently outperform selective perturbation [1] and simpler feature-preservation baselines under matched optimization settings?

Thank you for this question! Building on prior ranking manipulation and encoder attacks, CIRA couples global priority inversion with targeted preservation of displaced tokens.

  • Contribution of HEP.

    • HEP weights clean-high-priority tokens by whether their adversarial ranks cross candidate retention boundaries, marginalizing over candidate budgets, and constrains their representation drift.
    • In the original POPE-VisionZip component ablation, GSH alone yields 18.56% average CSFR and 15.72% Full ASR. Adding HEP yields 17.77% average CSFR and 4.96% Full ASR, reducing full-token failures by 68.4% while retaining 95.8% of the CSFR.
  • Matched preservation controls.

    • Global Pres. (Table 3) applies uniform preservation, Linear Pres. uses static clean-rank weights, and the CAA-style control perturbs only the lowest-priority 40% of patches according to the same clean encoder proxy.
    • All four methods use the same GSH objective, encoder proxy, perturbation budget of 4/255, step size of 1/255, 100 optimization steps and clean initialization.
    • The matched reruns use the same 1,000 POPE-VisionZip pairs. Average CSFR equally weights K = 32, 64, 128 and 192.
    Preservation strategyAvg. CSFR (%) ↑Full ASR (%) ↓
    HEP17.915.09
    Global Pres.15.764.38
    Linear Pres.17.696.15
    CAA-style selective perturbation19.715.92
    • HEP improves average CSFR over Global Pres. by 2.15 percentage points and achieves higher average CSFR with lower Full ASR than Linear Pres.
    • Compared with the spatial mask, HEP achieves lower Full ASR, 5.09% versus 5.92%, with average CSFR of 17.91% versus 19.71%.
    • HEP constrains displaced-token representations while allowing image-wide perturbations; selective perturbation constrains spatial support.
AuthorsOpenReview ↗

[2/2] Response to Reviewer 5pKP

Question 3 & Weakness 3 The attack relies on encoder-side proxy rankings that may differ substantially from actual token-selection decisions, particularly for merging-based compressors and LLM-attention-based methods such as FastV, but this proxy mismatch is not systematically quantified. Can the authors quantify the agreement between encoder-side proxy rankings and actual retained-token selections across different compressors, and relate this agreement to CSFR?

Thank you for this question! We traced native patch IDs and predictions on three benchmarks, 1,000 pairs each and four budgets.

  • Proxy–native agreement measurement.

    • Proxy Top-D matches each compressor's native direct/representative capacity, including PruMerge's representative patches and FastV's actual decoder-side selections.
    • The uniform reference fixes the native set and uniformly selects D patches from all 576 patches. We compute its exact expected Jaccard and average using the same observation weights as the measured adversarial retained-set Jaccard.
    • For each selection rule, the displaced set contains patches retained on the clean input that lose retained/representative membership after attack. We also compute Jaccard between the proxy and native displaced sets.
  • CSFR analysis and results.

    • Among observations clean-correct under full-token and native-compressed inference, we split within each compressor–dataset–budget setting by displaced-set Jaccard: lower is at or below the median; higher is above it. Tied values stay together.
    • Each group contains at least 174 observations per setting. Group CSFR averages 12 dataset-budget settings equally; retained-set Jaccard averages all evaluated observations.
    CompressorRetained Jaccard: proxy / uniformCSFR: lower / higher displaced agreement (%)
    VisionZip0.457 / 0.08417.44 / 21.23
    VisPruner0.329 / 0.10314.65 / 16.01
    PruMerge0.543 / 0.10230.77 / 37.49
    FastV0.328 / 0.10315.76 / 16.80
    • Per-compressor mean retained agreement is 3.17–5.41 times the uniform reference; all 48 setting means exceed their matched reference.
    • The higher displaced-agreement groups have higher mean CSFR for all four compressors, by 3.79, 1.36, 6.72 and 1.04 percentage points, respectively.

Thank you for your response. It addressed my questions, and I will maintain my score.

Thank you for your time and constructive feedback, which have helped us improve the quality of our paper! We are pleased that our response addressed your questions.

Reviewer_Ltz5OpenReview ↗
Rating
3(+1)
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper studies adversarial failures that occur under visual-token compression while full-token inference remains correct. It introduces a clean-conditioned, paired evaluation metric and proposes CIRA, an encoder-only attack combining global token-priority reversal with representation preservation for displaced tokens. Experiments cover three benchmarks, four compressors, and additional LVLM families. The paper also presents retained-set counterfactual diagnostics and a translation-consensus defense with adaptive evaluation. On LLaVA, CIRA achieves a reported mean compression-specific failure rate of 20.35% with a full-token attack success rate of 6.92%.

Critical strengths
  1. The paired evaluation addresses a useful attribution problem. Requiring both inference paths to be correct before attack, and full-token inference to remain correct afterward, distinguishes compression-specific failures from failures shared with the underlying model. This provides a more informative assessment than compressed accuracy alone.

  2. The experimental coverage and component analyses are useful. The evaluation spans several compression mechanisms and model families. Table 3 supports the distinct roles of priority manipulation and representation preservation, while the adaptive defense evaluation acknowledges that standard-attack suppression does not establish robust protection.

Critical weaknesses
  1. The comparisons do not sufficiently establish the incremental contribution over closely related attacks. Section 2 already describes CAA as manipulating selection rankings and CAGE as addressing unknown compression settings. However, VEAttack and CAGE are evaluated at operating points with much higher full-token failure rates, whereas CSFR explicitly rewards full-token preservation. The results therefore do not isolate how much improvement comes from CIRA’s specific design versus adding preservation to an existing attack. Table 3 provides a useful global-preservation control, but only at the reported preservation setting. Comparisons against preservation-augmented baselines and tradeoff curves at comparable Full ASR would make the contribution substantially more convincing. The CAA configuration also needs fuller specification.

  2. The validation protocol leaves uncertainty about generalization and the reliability of smaller gains. Appendix D.4 selects the scoring-layer configuration using downstream selectivity on the same three benchmark sources and three compressors used in the final evaluation, without explicitly establishing separation between design and test samples. This does not invalidate encoder-only optimization, but weakens the evidence for generalization to previously unseen compression configurations. Moreover, the paper does not report per-setting clean-eligible sample counts, diagnostic cohort sizes, or uncertainty estimates. These omissions matter because CSFR uses different eligible subsets across settings, and the targeted-preservation advantage over global preservation is relatively modest.

Questions for authors
  1. Does CIRA retain its advantage at comparable Full ASR when CAGE is augmented with the same representation-preservation term, or when global preservation is independently tuned? Please also specify the exact CAA objective and its compressor/budget assumptions.

  2. Are the 100-example design subsets in Appendix D.4 disjoint from the final evaluation samples, including at the image level? Does the selected configuration remain effective on a compressor excluded from configuration selection?

  3. What are the clean-eligible sample counts and diagnostic cohort sizes for each setting? Can image-level bootstrap confidence intervals establish the reliability of the targeted-versus-global preservation gains and the reported drift–recovery association?

4 comments
AuthorsOpenReview ↗

[1/2] Response to Reviewer Ltz5

We thank the reviewer for the constructive comments.

Question 1 & Weakness 1 The comparisons do not sufficiently establish how much improvement comes from CIRA’s specific design versus adding preservation to an existing attack, and Table 3 provides a useful global-preservation control only at the reported preservation setting. Does CIRA retain its advantage at comparable Full ASR when CAGE is augmented with the same representation-preservation term, or when global preservation is independently tuned? Please also specify the exact CAA objective and its compressor/budget assumptions.

Thank you for this question!

  • Experimental setup.

    • We added the same HEP or global-preservation term to CAGE, and replaced HEP in CIRA with global preservation.
    • Each variant independently swept λ∈{0.2,0.4,0.6,0.8,1.0} using the same 1,000 POPE pairs, LLaVA–VisionZip, ε=4/255, and 100 steps.
    • Cells show Avg. CSFR / Full ASR (%); CSFR equally averages K=32/64/128/192.
    Methodλ=0.2λ=0.4λ=0.6λ=0.8λ=1.0
    CIRA (GSH + HEP)18.92/8.7617.49/7.2218.08/6.1517.91/5.0916.81/4.62
    CAGE + HEP9.17/37.5113.96/27.5715.55/19.8815.70/17.6314.79/14.32
    CAGE + Global Pres.7.22/40.957.48/38.587.63/36.097.26/36.928.83/31.36
    GSH + Global Pres.18.81/10.5316.70/8.0515.45/6.3915.76/4.3814.94/4.50
  • Contribution of GSH.

    • Across the tested grid, CIRA achieves higher CSFR and lower Full ASR than both preservation-augmented CAGE variants.
    • CAGE + HEP's minimum Full ASR is 14.32%, versus CIRA's maximum of 8.76%.
    • GSH + Global Pres. also improves both metrics over CAGE + Global Pres., supporting global priority inversion.
  • Contribution of HEP.

    • Within CAGE, HEP improves both metrics over global preservation at every tested weight.
    • At comparable Full ASR, CIRA (λ=1.0) achieves 16.81% CSFR / 4.62% Full ASR, versus 15.76% / 4.38% for GSH + Global Pres. (λ=0.8): a 1.05 pp CSFR point gain.
  • CAA† configuration.

    • Objective: question-conditioned scores and relative patch groups from uncompressed LLM layer 2; hierarchical BPR ranking (weights 20/20/10, threshold 0.1), key-direction matching (5/5), and semantic erasure (10/10, margin 7).
    • Noise/least patch fractions: 0.2/0.4 (POPE), 0.1/0.2 (TextVQA), 0.1/0.3 (MME).
    • SGD: learning rate 1.0, random initialization, ε=4/255, 100 steps.
    • Perturbations are reused across compressors and budgets. We will add these details to Appendix B.4.

Question 2 & Weakness 2 Appendix D.4 selects the scoring-layer configuration using downstream selectivity on the same three benchmark sources and three compressors used in the final evaluation, without explicitly establishing separation between design and test samples. Are the 100-example design subsets in Appendix D.4 disjoint from the final evaluation samples, including at the image level? Does the selected configuration remain effective on a compressor excluded from configuration selection?

Thank you for this question!

  • The 100-example scoring-layer design subsets in Appendix D.4 were image-disjoint from their corresponding 1,000-pair main evaluation sets.

  • Scoring layers were selected using VisionZip, PruMerge, and VisPruner. Table 2 also includes FastV, which did not participate in layer selection. With the same selected layers, CIRA achieves Avg. CSFR of 16.55%/19.96%/12.23% on FastV for POPE/TextVQA/MME, supporting transfer beyond the selection compressors.

AuthorsOpenReview ↗

[2/2] Response to Reviewer Ltz5

Question 3 & Weakness 2 The paper does not report per-setting clean-eligible sample counts, diagnostic cohort sizes, or uncertainty estimates. These omissions matter because CSFR uses different eligible subsets across settings, and the targeted-preservation advantage over global preservation is relatively modest. What are the clean-eligible sample counts and diagnostic cohort sizes for each setting? Can image-level bootstrap confidence intervals establish the reliability of the targeted-versus-global preservation gains and the reported drift–recovery association?

Thank you for this question!

  • Original Table 2 CIRA runs use 1,000 pairs per benchmark. Clean eligibility requires correct clean full-token and compressed predictions. Counts follow K=32/64/128/192.

    BenchmarkVisionZipVisPrunerPruMergeFastV
    POPE675/752/804/815698/765/805/818638/661/686/713625/697/759/781
    TextVQA470/518/549/564481/528/554/562456/476/483/492346/417/465/489
    MME636/687/721/743654/690/730/742640/657/663/683613/657/694/723
  • Each diagnostic dataset–budget setting has 45 images. Question counts:

    KPOPETextVQAMME
    32/64/128/19253/52/53/4945/45/45/4547/46/48/47

    In total: 575 question–budget observations, 252 distinct questions, and 231 distinct images.

  • The new POPE–VisionZip sweep uses 1,000 pairs from 453 images, with common clean-eligible counts of 677/752/802/812. At λ=0.8, HEP/global preservation achieve 17.91%/15.76% Avg. CSFR and 5.09%/4.38% Full ASR. The paired gain and original diagnostic slope are:

    QuantityEstimate (pp)95% bootstrap interval (pp)
    HEP − Global Pres., Avg. CSFR difference+2.15[0.64, 3.65]
    Drift–recovery slope per 0.1 increase in drift−8.07[−13.90, −2.53]
    • Preservation gain. The interval uses 10,000 paired image-cluster percentile bootstrap resamples, grouping each image's questions and budgets and recomputing CSFR denominators.
    • Drift–recovery association. The interval uses 10,000 Bayesian image-cluster bootstrap draws. The slope is estimated from within-dataset–budget variation in the original fixed first-CSF cohort.

    Both 95% bootstrap intervals exclude zero, supporting the CSFR gain of HEP over global preservation at λ=0.8 and the association between greater representation drift and smaller evidence-restoration benefits in the original diagnostic cohort.

The rebuttal substantially addresses my main concerns. In the POPE–VisionZip setting, CIRA achieves higher CSFR and lower Full ASR than the preservation-augmented CAGE baselines across the tested weight grid. The clarification that the design and evaluation sets are image-disjoint, and that FastV was excluded from layer selection, also strengthens the evaluation. The incremental benefit of HEP over global preservation at comparable Full ASR remains modest, and its uncertainty is not reported for that comparison. On balance, I am raising my rating to weak acceptance.

Thank you for your prompt reply! At comparable Full ASR, HEP achieves Avg. CSFR gains of 2.63 pp (95% CI: [0.98, 4.32]) at λ=0.6 and 1.87 pp (95% CI: [0.40, 3.37]) at λ=1.0, outperforming global preservation.

We sincerely appreciate your thoughtful feedback and your decision to raise the rating to Weak Accept!

Reviewer_PC25OpenReview ↗
Rating
3
Confidence
3
Informativeness
3
Presentation
3
Soundness
3
Read review
Paper summary

This paper investigates the adversarial risks introduced by visual token compression in Large Vision-Language Models (LVLMs). The authors propose CIRA, a compression-specific adversarial attack that combines Global Selection Hijacking (GSH) and Hidden-Evidence Preservation (HEP). CIRA aims to induce failures in compressed models while preserving the correctness of full-token inference. Extensive experiments across multiple LVLMs, compression methods, and benchmarks demonstrate the effectiveness of the proposed approach.

Critical strengths

1.Clear motivation and meaningful problem formulation: This paper investigates an important yet relatively underexplored problem concerning the adversarial risks introduced by visual token compression.

2.Effective and well-motivated method design: CIRA integrates two complementary components, GSH and HEP, to manipulate visual token priorities while preserving important visual representations. This method is well aligned with the motivation, and the ablation studies provide convincing evidence supporting the contributions of the paper.

3.Comprehensive evaluation: The authors conduct extensive experiments across various LVLMs, compression methods, compression budgets, and benchmarks. The mechanistic analyses, defense experiments strengthen the contributions.

Critical weaknesses

See questions

Questions for authors
  1. Could the authors provide further insights into why CIRA achieves particularly strong performance on PruMerge compared with other compression methods?

  2. Could the authors briefly discuss how the choice of the candidate compression budget range affects the performance of CIRA in practical scenarios?

  3. Could the authors further discuss the computational overhead introduced by TCS and its potential impact on inference efficiency?

3 comments
AuthorsOpenReview ↗

[1/1] Response to Reviewer PC25

We thank the reviewer for the positive assessment and constructive questions.

Question 1 Could the authors provide further insights into why CIRA achieves particularly strong performance on PruMerge compared with other compression methods?

Thank you for this question! A plausible explanation is that PruMerge depends on attention at both selection and aggregation.

  • PruMerge selects attention-ranked representatives and uses attention weights to aggregate neighboring and residual features.

    • GSH can therefore affect both representative identities and the contribution of clean high-priority evidence, while HEP constrains representation drift.
    • VisionZip’s contextual anchors and VisPruner’s diversity-based selection provide additional retention routes.
  • We measured how closely the proxy priorities correspond to native selections on three benchmarks, with 1,000 pairs each and four budgets.

    • Proxy Top-D matches each compressor’s native direct/representative capacity.
    • Jaccard averages all evaluated observations across the three benchmarks and four budgets.
    MetricPruMergeVisionZipVisPrunerFastV
    Adversarial retained-set Jaccard0.5430.4570.3290.328
  • PruMerge has the highest observed adversarial retained-set agreement (Jaccard 0.543). This supports the connection between CIRA’s encoder priorities and PruMerge’s native representative choices. Together with PruMerge’s reliance on attention in both selection and aggregation, it offers a plausible explanation for the stronger effect observed on this compressor.

Question 2 Could the authors briefly discuss how the choice of the candidate compression budget range affects the performance of CIRA in practical scenarios?

Thank you for this question! The interval controls HEP’s budget-marginal weights, while GSH optimizes the full priority ranking.

  • With evaluation budgets fixed at K∈{32, 64, 128, 192}, the appendix sensitivity study (Table 12) reports the following results on LLaVA–POPE–VisionZip.

    Candidate intervalAvg. CSFR (%) ↑Full ASR (%) ↓
    [32,192], default17.774.96
    [16,192]17.704.62
    [64,192]17.285.56
    [32,64]18.575.68
    [32,128]17.214.97
    [32,384]16.805.33
  • Across these six intervals, average CSFR varies by only 1.77 percentage points, while Full ASR stays within 4.62%–5.68%. This stability supports using a coarse range of plausible deployment budgets.

Question 3 Could the authors further discuss the computational overhead introduced by TCS and its potential impact on inference efficiency?

Thank you for this question!

  • We measured LLaVA-v1.5-7B on RTX 4090 in FP16, using the same 32 POPE inputs, batch size 1, and exactly 32 generated tokens. After warm-up, we measured 64 requests per budget and condition.

    • TCS increases end-to-end median latency by 0.48%–3.95%.

      KVisionZip p50 (ms)+TCS p50 (ms)Overhead
      32773.97785.831.53%
      64783.24786.980.48%
      128793.46824.793.95%
      192786.74804.932.31%
    • The additional visual-stage latency is 5.26–5.43 ms, and peak allocated GPU memory increases by 0.84 GiB (6.24%). TCS batches four views through the vision encoder and performs one LLM generation. End-to-end timing starts from preprocessed GPU-resident images and includes tokenization, visual processing, generation, and decoding.

  • These measurements show that TCS provides cross-view selection stabilization with less than 4% additional end-to-end median latency under the tested setting. This supports a practical balance between selection stability and inference efficiency.

Reviewer_PC25OpenReview ↗

Official comments by Reviewer PC25

Thanks for your feedback. The rebuttal addresses my concerns. I will maintain my score. Good luck!

Thank you for your time and constructive feedback, which have helped us improve the quality of our paper! We are pleased that our response addressed your questions.

Comments on the paper

1 comment
AuthorsOpenReview ↗

Author Final Remarks

Dear Area Chair and Reviewers,

Thank you for your time and constructive feedback. Our paper received initial scores of 3/3/2. In our rebuttal, we added experiments and clarified the following points:

  • PC25: Performance on PruMerge, budget sensitivity, and TCS overhead.
  • Ltz5: Baseline comparisons, evaluation design, and statistical uncertainty.
  • 5pKP: Cross-compressor validation, HEP’s contribution, and proxy–selection agreement.

Following the rebuttal, Ltz5 raised their score from 2 to 3, noting that their main concerns had been substantially addressed. PC25 and 5pKP confirmed that their questions had been answered and maintained their scores of 3. The final scores are 3/3/3, with all three reviewers recommending Weak Accept.

We appreciate the reviewers’ feedback throughout the discussion and thank the Area Chair for considering our paper.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.