Paper deep dive

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

Qiankun Li, Yuechen Zhang, Bowen Chen, Shilinlu Yan, Zhenhong Zhou, Kun Wang, Li Sun

Year: 2026Venue: arXiv preprintArea: cs.CVType: PreprintEmbeddings: 105

Intelligence

Status: not_run | Model: - | Prompt: - | Confidence: 0%

Entities (0)

No extracted entities yet.

Relation Signals (0)

No relation signals yet.

Cypher Suggestions (0)

No Cypher suggestions yet.

Abstract

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.

PDF

Open source PDF →Open local PDF →

Trouble viewing inline? Open PDF directly →

Full Text

101,559 characters extracted from source content.

Expand or collapse full text

Still There, No Longer Seen: Exposing Compression-Induced Risk in Large Vision-Language Models Qiankun Li1 Yuechen Zhang211footnotemark: 1 Bowen Chen2 Shilinlu Yan2 Zhenhong Zhou1 Kun Wang1 Li Sun222footnotemark: 2 1Nanyang Technological University 2Beijing University of Posts and Telecommunications cs-qiankun.li,wang.kun@ntu.edu.sg orange.zhangyc05@gmail.com zhenhong001@e.ntu.edu.sg cbcbw,lulu_land,lsun@bupt.edu.cn †thanks: Equal contribution.†thanks: Kun Wang and Li Sun are the corresponding authors. 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%20.35\% while limiting full-token attack success to 6.92%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. Code is provided in the Github. 1 Introduction Figure 1: Overview of the compression-specific failure setting, the CIRA attack framework and the resulting behavior under visual-token compression. Large Vision-Language Models (LVLMs) support broad visual understanding by mapping images to long visual-token sequences (Alayrac et al., 2022; Li et al., 2023a; Liu et al., 2024b). Visual-token compression reduces this inference cost through token selection, aggregation, and compact visual representations (Shang et al., 2025; Yang et al., 2025b; Li et al., 2025; Bulat et al., 2026). Decoder-side methods further reduce visual-token computation within the language model (Chen et al., 2024; Zhang et al., 2024; Xing et al., 2024). Meanwhile, adversarial attacks show that failures can be induced in the underlying LVLM through image-space or joint-modal perturbations (Zhang et al., 2022; Schlarmann and Hein, 2023; Yin et al., 2023). This raises the question: what risks are induced by visual-token compression itself, beyond vulnerabilities already present in the underlying LVLM? When a full-token LVLM is used for robustness assessment while a compressed variant is deployed, failures confined to the compressed path are not captured by that assessment. We therefore evaluate each adversarial input with and without token compression. A compression-specific failure (CSF) occurs when full-token inference remains correct but compressed inference fails. This paired criterion distinguishes failures introduced by compression from those already present in the underlying LVLM. Recent studies increasingly examine robustness under visual-token compression. Evidence shows that compression can reshape robustness in either direction, depending on how visual evidence is selected and retained (Wang et al., 2026a; Gu et al., 2026). Compression-aware attacks further demonstrate that adversarial behavior can depend on the compression path itself (Zhang et al., 2026c; Zhang et al., 2026a). These developments shift the question from whether compression affects robustness to which adversarial failures can be attributed specifically to compression and whether they can be selectively induced under deployment uncertainty. Specifically, can a single perturbation optimized through the target vision encoder preserve full-token correctness while inducing compressed-path failures when the deployed compressor and budget are unknown during optimization? To understand what governs this selectivity, we use controlled counterfactuals to isolate two factors. Retained-set interventions causally change compressed correctness by altering which visual evidence remains accessible after compression. Meanwhile, recovery is negatively associated with representation drift in displaced evidence. These results directly inform the attack design: reallocate token priority while preserving displaced evidence. In this direction, we introduce CIRA (Compression-Induced Risk Attack), an attack framework for inducing compression-specific failures. As illustrated in Figure 1, CIRA combines Global Selection Hijacking (GSH), which globally reallocates encoder-side proxy priorities, with Hidden-Evidence Preservation (HEP), which limits representation drift in displaced clean high-priority tokens. With white-box access restricted to the vision encoder, CIRA uses no downstream questions or labels and requires no access to the language model, deployed compressor, or exact compression budget. Across four compressors, three datasets, four budgets, and multiple LVLMs, CIRA induces compression-specific failures while keeping full-token degradation limited. We also introduce Translation-Consensus Selection (TCS), a cross-view selection-stabilization defense, and evaluate it against both standard and adaptive CIRA. Our contributions are threefold: ❶ Compression Risk Attribution. We formulate compression-induced risk through a clean-conditioned paired CSF criterion and show that retained-set allocation causally affects compressed correctness while recovery decreases with representation drift. ❷ Compression-Induced Risk Attack. We introduce CIRA, a target-encoder-only attack that combines priority reallocation with evidence preservation and induces paired CSFs across compressor–budget settings without configuration-specific optimization. ❸ Comprehensive Evaluation and Analysis. We evaluate CIRA across four compressors, three datasets, four budgets, and multiple LVLM families, with mechanistic evidence of global priority reallocation and limited representation drift in displaced evidence. We also test TCS as a selection-stabilization defense against both standard and adaptive CIRA. 2 Related Work Visual token compression. Token-selection methods exploit visual cues, diversity, and salience–coverage objectives (Zhang et al., 2025b; Alvar et al., 2025; Xu et al., 2026), while learned pruning addresses limitations of attention-based importance estimates (Takezoe et al., 2026). Hybrid methods combine pruning with clustering and merging (Endo et al., 2025; Dhouib et al., 2025; Yang et al., 2025a). Query-conditioned methods use textual instructions for token scoring or aggregation (Yu et al., 2026; Gao et al., 2026; Sun et al., 2026). Adaptive and progressive methods vary token reduction across inputs and layers (Ye et al., 2025; Chen et al., 2026; Li et al., 2026a), while video and multi-turn approaches address temporal redundancy and evolving context (Wang et al., 2025a; Shen et al., 2024; Li et al., 2026b; Wang et al., 2026c). Adversarial robustness of vision-language models. Transfer-based attacks exploit set-level guidance, prompt-robust objectives, and iterative multimodal alignment (Lu et al., 2023; Luo et al., 2024; Liu et al., 2024a; Xie et al., 2025). Encoder-based and self-supervised attacks support cross-task and cross-model transfer (Zhang et al., 2025a; Hu et al., 2025; Zhang et al., 2026b). VEAttack disrupts visual representations (Mei et al., 2026b), while PA-Attack uses prototype-guided gray-box attacks (Mei et al., 2026a). Defenses include adversarial encoder fine-tuning (Schlarmann et al., 2024) and adversarial pre-training and instruction tuning (Wang et al., 2025c). Test-time defenses use prompt adaptation and augmented-view consistency (Sheng et al., 2025; Liu et al., 2025). Robustness under visual token compression. Visual-token compression introduces distinct robustness risks and defense opportunities. Safety-Aware Pruning (SAP) (Wang et al., 2026a) mitigates pruning-induced vulnerabilities, while robustness-oriented pruning (Gu et al., 2026) removes visually misaligned tokens. On the attack side, CAGE (Zhang et al., 2026c) targets tokens expected to survive unknown compression settings. CAA (Zhang et al., 2026a) directly manipulates token-selection rankings, with white-box attacks tailored to known compression configurations and transfer attacks using surrogate models. Our focus is on selectively inducing compression-specific failures under encoder-only access, without knowledge of the deployed compressor or budget. 3 Diagnosing Compression-Specific Failures In this section, we investigate why adversarial inputs can fail only after compression and use controlled diagnostics to inform the design of CIRA. 3.1 Problem Setup and Compression-Specific Failure Let C denote a visual-token compressor and K its compression budget. For a dataset =(xi,qi,yi)i=1MD=\(x_i,q_i,y_i)\_i=1^M, let fθ(xi,qi)f_θ(x_i,q_i) and fθ,K(xi,qi)f_θ^C,K(x_i,q_i) denote full-token and compressed inference, respectively. Given a task-specific binary evaluator χ⁡(y^,y)∈0,1χ( y,y)∈\0,1\, we define ci(x)=χ⁡(fθ(x,qi),yi),ci,,K(x)=χ⁡(fθ,K(x,qi),yi).c_i(x)=χ\! (f_θ(x,q_i),y_i ), c_i,C,K(x)=χ\! (f_θ^C,K(x,q_i),y_i ). For a fixed C, we write ci,K(x)≡ci,,K(x)c_i,K(x)≡ c_i,C,K(x) and omit the compressor subscript below. For each compression budget K, we restrict evaluation to samples for which both full-token and compressed inference produce correct answers before attack: K=i|ci(xi)=1,ci,K(xi)=1.S_K= \i\; |\;c_i(x_i)=1,\;c_i,K(x_i)=1 \. (1) A compression-specific failure (CSF) occurs when an adversarial input remains correct under full-token inference but fails after compression: ℱKCSF=i∈K|ci(xiadv)=1,ci,K(xiadv)=0.F^CSF_K= \i _K\; |\;c_i(x_i^adv)=1,\;c_i,K(x_i^adv)=0 \. (2) Aggregate accuracy can obscure instance-level prediction changes introduced by VLM acceleration (Sun et al., 2025). Our paired definition focuses on adversarial inputs that remain correct under full-token inference but fail after compression. 3.2 Retained-Set Allocation and Compressed Correctness Diagnostic setting. We test whether retained-set allocation can change compressed correctness while the adversarial encoder state and compression capacity remain fixed. LLaVA–VisionZip exposes direct-token membership, providing a controlled interface for retained-set intervention. From each downstream-agnostic VEAttack trajectory, we retain the earliest checkpoint satisfying the CSF criterion, termed a first-observed CSF. Counterfactual intervention. For each CSF, we exchange equal numbers of retained and omitted direct tokens, reconstruct compression, and rerun inference at the same encoder state and capacity. A post-hoc answer-aware oracle selects the exchanged token identities (Appendix C). Recovery criteria. Let ρ denote the fraction of exchanged direct-token slots. We compare guided reallocation with a size-matched random exchange. Cumulative recovery credits correction at any tested ρ′≤ρ ≤ρ, whereas exact recovery uses only ρ. ✰ Observation 1: Controlled retained-set reallocation changes compressed correctness. Figure 2: Cumulative CSF recovery under guided and matched-random retained-set exchanges across direct-token exchange fractions. Figure 2 shows that guided recovery exceeds matched-random recovery for every tested budget–fraction pair with ρ>0ρ>0, with an average advantage of 22.522.5–24.124.1 p across the 12 dataset–budget cells. With encoder state, capacity, and exchange size fixed, the contrast shows that robustness depends on token identity, not only on retention count. 3.3 Retained-Set Allocation Components and Representation Drift Having established retained-set allocation sensitivity, we next separate restoration from removal and ask when restored evidence remains useful. Observations 2–3 use ρ∗=7.5%ρ^*=7.5\% as a shared intermediate diagnostic point, where guided and random recovery are already clearly separated. Representation drift is the restored tokens’ mean half-cosine distance, [1−cos⁡(c,a)]/2[1- (h^c,h^a)]/2, between their clean and adversarial representations; Table 1(b) reports this distance multiplied by 100100. ✰ Observation 2: Recovery depends on what is restored and removed. Table 1(a) decomposes guided reallocation into restoring omitted tokens with high retrospective answer support and removing retained tokens with low support; larger scores indicate stronger support for the reference answer. Exact recovery rises from 17.4%17.4\% under matched random exchange to 42.0%42.0\% under guided reallocation. Both restoration and removal contribute to recovery, with their coordinated use providing the largest gain. Table 1: Mechanistic diagnostics at ρ∗=7.5%ρ^*=7.5\%: exact recovery under retained-set interventions (a) and evidence-restoration effects across representation-drift tertiles (b), with construction and aggregation details in Appendix C. (a) Retained-Set Counterfactuals (b) Association with Representation Drift Retained-Set Change Exact Recovery (%) Avg. K=192K=192 K=128K=128 K=64K=64 K=32K=32 Random 17.9 17.2 18.4 16.1 17.4 ++ Evidence restoration 31.1 33.5 34.0 34.9 33.4 ++ Low-support removal 34.5 31.9 25.4 20.1 28.0 Guided reallocation 46.4 44.4 38.5 38.6 42.0 Drift Tertile Mean Drift (%) Evidence Effect (p) Avg. K=192K=192 K=128K=128 K=64K=64 K=32K=32 Lower 31.4 +16.6+16.6 +18.9+18.9 +18.2+18.2 +28.5+28.5 +20.5+20.5 Middle 39.5 +15.3+15.3 +12.7+12.7 +7.0+7.0 +20.8+20.8 +14.0+14.0 Upper 45.2 +5.4+5.4 +11.5+11.5 +18.0+18.0 +6.1+6.1 +10.3+10.3 Lower–Upper Gap – +11.2+11.2 +7.4+7.4 +0.2+0.2 +22.3+22.3 +10.3+10.3 ✰ Observation 3: Restoration benefits are smaller under greater representation drift. Table 1(b) shows that evidence restoration is most effective for tokens with limited representation drift: the average effect falls from +20.5+20.5 p in the lower-drift tertile to +10.3+10.3 p in the upper. A within-cell continuous analysis shows the same negative association. Design implication. Together, these diagnostics motivate CIRA’s objective: make evidence inaccessible after compression while preserving its utility under full-token inference. 4 Methodology 4.1 Attack Formulation We consider an untargeted, image-specific attack with white-box access only to the vision encoder. The downstream question, reference answer, language model, compressor, and exact deployed compression budget are unavailable during optimization; the attacker knows only the admissible range =K∈ℕ:Kmin≤K≤KmaxK=\K :K_ ≤ K≤ K_ \. For an evaluation tuple (x,q,y)(x,q,y), the ideal compression-specific outcome is to preserve full-token correctness while causing compressed inference to fail: max‖δ‖∞≤ϵ⁡minK∈⁡ℒ⁡(fθ,K(x+δ,q),y)s.t.c⁡(x+δ)=1, _\|δ\|_∞≤ε\; _K L (f_θ^C,K(x+δ,q),y ) .t. c(x+δ)=1, (3) where c⁡(⋅)c(·) denotes full-token correctness and ℒL increases with compressed prediction error. Equation 3 defines the desired compression-specific behavior but is not optimized directly. CIRA instead replaces it with compressor-independent encoder-side objectives and evaluates cross-configuration transfer empirically. 4.2 Global Selection Hijacking Let si(x)s_i(x) denote the encoder-side proxy priority score of visual token i, where larger values indicate higher retention priority. We collect these scores in the priority-score vector ⁡(x)=[s1(x),…,sN(x)]∈ℝNs(x)=[s_1(x),…,s_N(x)] ^N, whose descending order defines the priority ranking over tokens. We write c=⁡(x)s^c=s(x) and a=⁡(x+δ)s^a=s(x+δ) for the clean and adversarial priority-score vectors, respectively. We first construct the clean rank, its inverse-priority encoding, and a stabilized standardization operator: c=rank↓⁡(c),ℛ(c)i=ric−1N−1,ϵs()=−u¯maxN−1/2∥−u¯∥2,ϵs.r^c=rank_ (s^c), (s^c)_i= r_i^c-1N-1, _ _s(u)= u- u1 \! \N^-1/2\|u- u1\|_2, _s \. (4) Here, cr^c ranks tokens from high to low clean priority, while ℛR maps high-priority tokens near zero and low-priority tokens near one. ϵsZ_ _s standardizes its input vector and prevents a degenerate denominator when its variance vanishes. To drive a global priority inversion, CIRA maximizes the standardized alignment between adversarial priorities and the inverse clean priority ranking: ℒGSH(δ)=Alignϵs⁡(a,ℛ⁡(c))≡1N⟨ϵs(a),ϵs(ℛ⁡(c))⟩.L_GSH(δ)=Align_ _s\! (s^a,R(s^c) )≡ 1N _ _s(s^a),Z_ _s\! (R(s^c) ) . (5) This differentiable, scale-invariant objective encourages clean high-priority tokens to move downward while promoting clean low-priority tokens. 4.3 Hidden-Evidence Preservation Priority reallocation may also perturb the representations of displaced evidence, undermining full-token correctness. HEP therefore focuses preservation on clean high-priority tokens that cross candidate retention boundaries. CIRA computes the adversarial descending ranks a=rank↓⁡(a)r^a=rank_ (s^a) and defines ℋK(δ)=i:ric≤K<riaH_K(δ)=\i:r_i^c≤ K<r_i^a\, the clean proxy Top-K tokens whose adversarial ranks move beyond the retention boundary at compression budget K. Under an unknown compression budget, their displacement weights are wi=PrK∼Unif⁡()[i∈ℋK(δ)]=[min⁡(Kmax,ria−1)−max⁡(Kmin,ric)+1]+Kmax−Kmin+1.w_i= _K (K)\! [i _K(δ) ]= [ (K_ ,r_i^a-1)- (K_ ,r_i^c)+1 ]_+K_ -K_ +1. (6) Here, [z]+=max⁡(z,0)[z]_+= (z,0), and wiw_i is the fraction of admissible budgets under which token i becomes hidden. To limit representation drift in these tokens, let ich_i^c and iah_i^a denote their clean and adversarial visual-token features. We normalize their directions and measure the resulting representation drift by ~ic=ic‖ic‖2,~ia=ia‖ia‖2,di=12(1−(~ic)⊤~ia), h_i^\,c= h_i^c\|h_i^c\|_2, h_i^\,a= h_i^a\|h_i^a\|_2, d_i= 12 (1-( h_i^\,c) h_i^\,a ), (7) where the factor 1/21/2 normalizes cosine distance to [0,1][0,1]. Hidden-Evidence Preservation aggregates these token-level distances using the budget-marginal weights: ℒHEP(δ)=−∑i=1Nw~idi,w~i=sg⁡(wi)max⁡∑j=1Nsg⁡(wj),ϵh.L_HEP(δ)=- _i=1^N w_id_i, w_i= sg(w_i) \! \ _j=1^Nsg(w_j), _h \. (8) The stop-gradient freezes rank-derived weights within an update, while ϵh _h stabilizes normalization; weights are recomputed at the next update so evidence displaced across more of K receives greater protection. 4.4 Joint Optimization CIRA combines the two objectives as maxδℒCIRA(δ)=ℒGSH(δ)+λ⋅ℒHEP(δ),s.t.‖δ‖∞≤ϵ. _δ _CIRA(δ)=L_GSH(δ)+λ·L_HEP(δ), .t. \|δ\|_∞≤ε. (9) Here, λ controls the preservation strength. We maximize equation 9 using projected sign-gradient ascent over the valid-image domain. The complete optimization procedure is given in Appendix A. 5 Experiments 5.1 Experimental Setup Models and Benchmarks. We evaluate LLaVA-v1.5-7B (Liu et al., 2024b) on 1,000 randomly sampled image–question pairs from each of POPE (Li et al., 2023b), TextVQA (Singh et al., 2019), and MME (Fu et al., 2025), covering object hallucination, scene-text understanding, and general visual perception and reasoning, respectively. We additionally evaluate Qwen3-VL-8B-Instruct (Bai et al., 2025) and InternVL3.5-8B (Wang et al., 2025b). Compression Settings. We evaluate VisionZip (Yang et al., 2025b), VisPruner (Zhang et al., 2025b), PruMerge (Shang et al., 2025), and FastV (Chen et al., 2024) at eval=32,64,128,192K_eval=\32,64,128,192\. For each input, one adversarial image is reused across all compressor–budget settings. Baselines. We compare CIRA with the downstream-agnostic VEAttack (Mei et al., 2026b) and CAGE (Zhang et al., 2026c); CAA (Zhang et al., 2026a) is reported as a stronger-access reference. Evaluation Metrics. On the clean-eligible set KS_K defined in Eq. 1, we report the Compression-Specific Failure Rate (CSFR) and full-token attack success rate: CSFRK=1|K|∑i∈Kci(xiadv)[1−ci,K(xiadv)],ASR=1|full|∑i∈full[1−ci(xiadv)].CSFR_K= 1|S_K| _i _Kc_i(x_i^adv) [1-c_i,K(x_i^adv) ], = 1|S_full| _i _full [1-c_i(x_i^adv) ]. (10) Here, full=i:ci(xi)=1S_full=\i:c_i(x_i)=1\, and Avg. CSFR is the arithmetic mean of CSFRKCSFR_K over K∈evalK _eval for a fixed dataset–compressor setting. Broader summaries weight each reported dataset–compressor setting equally. CSFR counts post-attack failures confined to the compressed path, normalized over inputs answered correctly by both clean inference paths; ASR is normalized over inputs answered correctly by clean full-token inference. Clean and post-attack accuracy results are reported in Appendix G. Implementation Settings. We run all experiments on a single NVIDIA GeForce RTX 4090 GPU. All downstream-agnostic attacks use ϵ=4/255ε=4/255 and 100 optimization steps; CIRA uses projected sign-gradient ascent with λ=0.8λ=0.8. Further protocol details and sensitivity analyses appear in Appendix B and Appendix D. 5.2 Main Results In this section, we evaluate CIRA’s compression selectivity and cross-configuration transfer, and compare it with stronger-access attacks. Table 2: CSFR and Full ASR on LLaVA-v1.5-7B. For downstream-agnostic attacks, red and blue cells mark highest CSFR and lowest Full ASR, and underlining marks second best; CAA† uses the downstream question and language model during optimization and is excluded from the markings. Metric POPE TextVQA MME VEAttack CAGE CAA† CIRA VEAttack CAGE CAA† CIRA VEAttack CAGE CAA† CIRA Full-token Inference (Compressor-independent) Full ASR ↓ 38.65 39.48 2.72 4.96 60.95 71.35 3.82 10.61 38.26 39.77 3.41 5.18 VisionZip CSFR@192 ↑ 4.05 3.19 1.72 7.61 4.87 2.92 2.14 11.35 2.83 4.04 1.62 7.40 CSFR@128 ↑ 4.98 5.72 2.87 14.05 5.32 5.32 2.05 18.21 3.33 5.55 1.80 10.26 CSFR@64 ↑ 5.05 9.71 3.46 22.47 6.28 6.28 2.39 27.41 3.64 5.97 2.92 19.07 CSFR@32 ↑ 6.52 9.48 4.57 26.96 9.76 5.24 5.01 31.70 4.56 5.19 2.86 22.17 Avg. CSFR ↑ 5.15 7.03 3.16 17.77 6.56 4.94 2.90 22.17 3.59 5.19 2.30 14.73 VisPruner CSFR@192 ↑ 3.42 3.79 1.23 5.87 2.92 3.50 1.75 10.50 2.56 3.50 1.22 7.14 CSFR@128 ↑ 4.60 4.35 2.97 9.07 3.80 3.40 2.18 13.18 2.88 4.79 2.33 11.23 CSFR@64 ↑ 6.27 7.45 4.68 15.29 5.89 5.47 4.08 19.13 3.33 5.36 4.47 18.70 CSFR@32 ↑ 7.59 10.46 8.29 22.21 8.07 7.33 4.63 25.57 4.13 5.66 6.73 21.87 Avg. CSFR ↑ 5.47 6.51 4.29 13.11 5.17 4.93 3.16 17.09 3.22 4.83 3.69 14.73 PruMerge CSFR@192 ↑ 5.33 3.65 2.66 22.86 6.02 4.86 3.44 28.66 2.64 4.69 1.75 20.64 CSFR@128 ↑ 5.39 5.54 2.92 29.59 6.51 5.35 4.86 36.23 3.62 5.28 3.02 22.93 CSFR@64 ↑ 6.96 9.38 3.63 32.53 8.29 6.64 4.49 43.91 3.50 5.78 3.04 29.98 CSFR@32 ↑ 9.40 11.76 3.44 32.29 9.18 6.70 5.20 48.68 4.06 5.78 2.81 35.16 Avg. CSFR ↑ 6.77 7.58 3.16 29.32 7.50 5.89 4.50 39.37 3.45 5.38 2.66 27.18 FastV CSFR@192 ↑ 4.61 4.74 2.30 7.04 4.29 4.09 1.43 9.61 2.77 2.90 1.38 5.81 CSFR@128 ↑ 6.85 6.59 2.90 11.46 7.10 5.38 2.15 13.55 3.17 3.89 1.44 7.64 CSFR@64 ↑ 9.90 10.47 3.44 22.24 9.11 7.67 4.08 23.74 3.65 5.18 2.89 15.07 CSFR@32 ↑ 12.32 10.40 4.48 25.44 7.80 8.38 3.76 32.95 4.24 5.55 6.20 20.39 Avg. CSFR ↑ 8.42 8.05 3.28 16.55 7.08 6.38 2.85 19.96 3.46 4.38 2.98 12.23 Compression-specific selectivity. Across the four budgets and 12 dataset–compressor settings in Table 2, CIRA achieves a mean CSFR of 20.35%20.35\%, compared with 5.49%5.49\% for VEAttack and 5.92%5.92\% for CAGE. Its Full ASR is 6.92%6.92\%, substantially below 45.95%45.95\% and 50.20%50.20\%, respectively. CIRA therefore induces more compression-specific failures while causing much less full-token degradation than either downstream-agnostic baseline. Cross-configuration transfer. The same adversarial image is reused without re-optimization across selection-, pruning-, and merging-based compression rules and all evaluated budgets. CIRA remains effective across these configurations, reaching 31.96%31.96\% mean CSFR on PruMerge compared with 6.28%6.28\% for CAGE. Averaged equally over all 12 dataset–compressor settings, its CSFR increases from 12.04%12.04\% at K=192K=192 to 28.78%28.78\% at K=32K=32. Compression-selective behavior also extends to Qwen3-VL and InternVL (Table 10); qualitative cross-setting examples appear in Appendix I. Comparison with stronger-access CAA. CAA† optimizes with access to the downstream question and language model, whereas CIRA uses only the vision encoder. Despite its lower Full ASR (3.32%3.32\% for CAA† versus 6.92%6.92\% for CIRA), CAA† achieves only 3.24%3.24\% mean CSFR, well below CIRA’s 20.35%20.35\%. The contrast shows that preserving the full-token prediction is not sufficient to induce compression-specific failures; CIRA’s priority-reallocation objective addresses this distinct requirement under narrower access. ➪ Takeaway. CIRA induces compression-specific failures across heterogeneous compression rules and budgets while limiting full-token degradation. 5.3 Mechanistic Analysis To connect CIRA’s behavior to its objectives, we test whether it globally reallocates token priority while limiting representation drift in displaced evidence. Figure 3: CIRA-induced priority reallocation across the clean visual-token ranking at K=128K=128. Figure 4: Priority reallocation and representation drift across token groups. Priority reallocation. Figure 4 shows a near-monotonic priority reallocation across all three benchmarks: the highest-priority octile is demoted by 0.470.47–0.530.53 normalized-rank units, whereas the lowest-priority octile is promoted by 0.730.73–0.760.76, with the sign changing near the median. This cross-quantile pattern is not confined to a single retention boundary. Priority reallocation versus representation drift. Figure 4 shows that rank displacement and representation drift behave differently across clean-priority groups. Rank demotion peaks at 53.0%53.0\% and 53.7%53.7\% in the top 10%10\% and 1010–30%30\% groups, where representation drift is lowest at 12.4%12.4\% and 12.7%12.7\%. The lowest-priority 6060–100%100\% group shows the reverse pattern, with 7.8%7.8\% rank demotion and 26.1%26.1\% drift. This separation indicates that CIRA reallocates clean high-priority tokens while limiting their representation drift, consistent with the HEP objective. Component ablation. On VisionZip, Table 4 isolates the two objectives. Removing GSH reduces CSFR averaged across the three benchmarks and four budgets from 18.22%18.22\% to 3.68%3.68\%. Removing HEP leaves this mean nearly unchanged (17.04%17.04\%) but raises mean Full ASR from 6.92%6.92\% to 23.10%23.10\%; replacing it with global representation preservation yields 15.71%15.71\% mean CSFR and 8.66%8.66\% Full ASR. Global preservation yields lower CSFR and higher Full ASR than HEP, supporting preservation focused on displaced evidence rather than a uniform constraint. ➪ Takeaway. GSH drives compression-specific failure induction, while HEP limits representation drift in displaced evidence and substantially reduces full-token degradation. Table 3: CIRA ablation on VisionZip; arrows show changes from full CIRA. Full Selection Preservation Metric CIRA w/o Sel. w/o Pres. Global Pres. POPE Full ASR ↓ 4.96 2.01↓2.95 _2.95 15.72↑10.76 _10.76 4.73↓0.23 _0.23 CSFR@192 ↑ 7.61 1.60↓6.01 _6.01 9.94↑2.33 _2.33 4.79↓2.82 _2.82 CSFR@128 ↑ 14.05 1.99↓12.06 _12.06 15.42↑1.37 _1.37 9.95↓4.10 _4.10 CSFR@64 ↑ 22.47 4.39↓18.08 _18.08 22.07↓0.40 _0.40 19.95↓2.52 _2.52 CSFR@32 ↑ 26.96 5.63↓21.33 _21.33 26.81↓0.15 _0.15 26.37↓0.59 _0.59 Avg. CSFR ↑ 17.77 3.40↓14.37 _14.37 18.56↑0.79 _0.79 15.26↓2.51 _2.51 TextVQA Full ASR ↓ 10.61 6.91↓3.70 _3.70 36.00↑25.39 _25.39 15.82↑5.20 _5.20 CSFR@192 ↑ 11.35 3.90↓7.45 _7.45 9.94↓1.41 _1.41 8.58↓2.77 _2.77 CSFR@128 ↑ 18.21 2.05↓16.16 _16.16 15.20↓3.02 _3.02 12.53↓5.69 _5.69 CSFR@64 ↑ 27.41 5.43↓21.98 _21.98 23.70↓3.72 _3.72 23.04↓4.37 _4.37 CSFR@32 ↑ 31.70 6.44↓25.26 _25.26 26.73↓4.97 _4.97 36.28↑4.57 _4.57 Avg. CSFR ↑ 22.17 4.46↓17.71 _17.71 18.89↓3.28 _3.28 20.11↓2.06 _2.06 MME Full ASR ↓ 5.18 2.40↓2.77 _2.77 17.57↑12.40 _12.40 5.44↑0.26 _0.26 CSFR@192 ↑ 7.40 1.62↓5.79 _5.79 7.55↑0.14 _0.14 5.80↓1.61 _1.61 CSFR@128 ↑ 10.26 1.93↓8.33 _8.33 11.33↑1.06 _1.06 7.73↓2.53 _2.53 CSFR@64 ↑ 19.07 4.09↓14.98 _14.98 17.23↓1.84 _1.84 14.45↓4.62 _4.62 CSFR@32 ↑ 22.17 5.08↓17.09 _17.09 18.57↓3.60 _3.60 19.05↓3.12 _3.12 Avg. CSFR ↑ 14.73 3.18↓11.55 _11.55 13.67↓1.06 _1.06 11.76↓2.97 _2.97 Table 4: TCS evaluation on VisionZip; arrows denote changes from matched no-defense baselines. CIRA CAGE Adaptive CIRA Metric None TCS None TCS None TCS POPE Full ASR 4.96 39.48 3.78 CSFR@192 ↓ 7.61 2.60↓5.01 _5.01 3.19 4.34↑1.15 _1.15 6.99 6.57↓0.42 _0.42 CSFR@128 ↓ 14.05 3.40↓10.65 _10.65 5.72 4.79↓0.93 _0.93 10.07 10.71↑0.64 _0.64 CSFR@64 ↓ 22.47 2.84↓19.63 _19.63 9.71 5.68↓4.03 _4.03 20.08 13.65↓6.43 _6.43 CSFR@32 ↓ 26.96 4.21↓22.75 _22.75 9.48 5.52↓3.96 _3.96 28.00 21.92↓6.08 _6.08 Avg. CSFR ↓ 17.77 3.26↓14.51 _14.51 7.03 5.08↓1.95 _1.95 16.29 13.21↓3.08 _3.08 TextVQA Full ASR 10.61 71.35 11.45 CSFR@192 ↓ 11.35 2.39↓8.96 _8.96 2.92 1.59↓1.33 _1.33 8.38 8.17↓0.21 _0.21 CSFR@128 ↓ 18.21 2.07↓16.14 _16.14 5.32 2.48↓2.84 _2.84 11.09 9.30↓1.79 _1.79 CSFR@64 ↓ 27.41 4.50↓22.91 _22.91 6.28 3.43↓2.85 _2.85 26.09 14.13↓11.95 _11.95 CSFR@32 ↓ 31.70 3.09↓28.61 _28.61 5.24 4.99↓0.25 _0.25 38.90 25.65↓13.25 _13.25 Avg. CSFR ↓ 22.17 3.01↓19.16 _19.16 4.94 3.12↓1.82 _1.82 21.11 14.31↓6.80 _6.80 MME Full ASR 5.18 39.77 6.57 CSFR@192 ↓ 7.40 2.72↓4.68 _4.68 4.04 2.31↓1.73 _1.73 3.64 5.03↑1.39 _1.39 CSFR@128 ↓ 10.26 2.64↓7.62 _7.62 5.55 2.50↓3.05 _3.05 9.53 7.22↓2.31 _2.31 CSFR@64 ↓ 19.07 4.53↓14.54 _14.54 5.97 4.09↓1.88 _1.88 18.54 11.68↓6.86 _6.86 CSFR@32 ↓ 22.17 5.75↓16.42 _16.42 5.19 4.67↓0.52 _0.52 23.17 18.35↓4.82 _4.82 Avg. CSFR ↓ 14.73 3.91↓10.82 _10.82 5.19 3.39↓1.80 _1.80 13.72 10.57↓3.15 _3.15 6 Selection Stabilization Defense To test whether stabilizing token selection can suppress CIRA, TCS exploits cross-view priority stability: clean high-priority evidence tends to remain stable under small translations, whereas attack-induced replacements are more view-sensitive. At the LLaVA–VisionZip token-priority interface, TCS uses four translated views =T0,0,Td,0,T0,d,Td,dV=\T_0,0,T_d,0,T_0,d,T_d,d\ with d=7d=7 pixels. For view v, (v)s^(v) is its priority-score vector, vA_v maps the translated score grid back to the reference coordinates, and rank 11 denotes the highest priority. TCS selects the K tokens with highest aligned rank-quantile consensus: qi(v)=1−ri(v((v)))−1N−1,q¯i=1||∑v∈qi(v),ℐKTCS=TopK⁡(¯,K).q_i^(v)=1- r_i\! (A_v(s^(v)) )-1N-1, q_i= 1|V| _v q_i^(v), ^TCS_K=TopK( q,K). (11) Rank quantiles make priorities comparable across views despite differences in score scale. The selected indices are applied to the unshifted-view features, leaving aggregation and full-token inference unchanged. Construction and mechanism details appear in Appendix F; complementary utility results are reported in Appendix G. Defense effectiveness. Within the matched evaluation blocks in Table 4, TCS reduces standard CIRA CSFR averaged across the three datasets and four budgets from 18.22%18.22\% to 3.39%3.39\%, an 81.4%81.4\% relative reduction, compared with a 32.5%32.5\% reduction for CAGE. Suppression strengthens as the compression budget decreases, with the largest reductions under tighter compression. Adaptive stress test. Because TCS is deterministic and public, we also evaluate an adaptive attacker that optimizes the CIRA objectives over the same four views while sharing one image-space perturbation: ℒadapt(δ)=1||∑v∈[ℒGSH(v)(δ)+λℒHEP(v)(δ)].L_adapt(δ)= 1|V| _v [L_GSH^(v)(δ)+ _HEP^(v)(δ) ]. (12) Adaptive CIRA uses the same access assumptions and perturbation budget as standard CIRA. Under TCS, its mean CSFR rises from 3.39%3.39\% to 12.70%12.70\%, reaching 74.5%74.5\% of Adaptive CIRA’s matched undefended value (17.04%17.04\%). TCS therefore retains a smaller but nonzero effect against the adaptive attack. Additional optimization and mechanism diagnostics appear in Appendix F.3. 7 Conclusion Visual-token compression changes not only inference cost but also which visual evidence remains available after compression. By pairing full-token and compressed inference on the same adversarial input, we attribute failures specifically to the compression path. CIRA induces such failures across unknown compressor and budget settings by reallocating token priorities while limiting full-token degradation. Controlled diagnostics further show that retained-set allocation affects compressed correctness and that recovery is negatively associated with representation drift in displaced evidence. Selection stabilization substantially suppresses CIRA, although adaptive optimization partially restores its effectiveness. Together, these results support treating the compression boundary as a security-relevant component of LVLM deployment and motivate paired robustness evaluation for compressed LVLMs. AI Use Statement Generative AI tools were used in a limited supporting role for this work, including language editing and polishing, literature retrieval and discovery, research ideation and technical execution support, and drafting and revising parts of the manuscript. In particular, these tools were used to improve the clarity and fluency of writing, help identify related literature, provide technical suggestions for coding and experimental workflows, and assist in refining sections of the paper during revision. All AI-assisted content was independently checked, validated, and, where necessary, modified by the authors. The authors retain full responsibility for the scientific content, methodological choices, experimental evidence, interpretations, and final form of the manuscript. Reproducibility Statement We provide comprehensive details to facilitate the reproduction and verification of CIRA. The threat model, compression-specific failure formulation, and attack objectives are described in Sections 3 and 4, including the Global Selection Hijacking (GSH) and Hidden-Evidence Preservation (HEP) objectives and their joint optimization. The optimization procedure and experimental details are provided in Appendices A and B, respectively. Main results, mechanistic analyses, and defense evaluations are reported in Sections 5.2, 5.3, and 6. Additional analyses, including retained-set diagnostics, sensitivity studies, cross-model evaluation, task-utility results, limitations, and qualitative cases, are provided in Appendices C, D, E, G, H, and I. Code and scripts for reproducing the reported experiments are provided in the Github repository. Ethics Statement This work investigates the adversarial robustness of LVLMs under visual-token compression in a controlled research setting. Our experiments use publicly available benchmark datasets (e.g., POPE, TextVQA, and MME) and publicly released pretrained models (e.g., LLaVA-v1.5-7B, Qwen3-VL-8B-Instruct, and InternVL3.5-8B), involving no human subjects or personally identifiable information. We recognize the dual-use nature of adversarial robustness research. CIRA is developed to identify compression-specific vulnerabilities and support robustness evaluation and defense development. The proposed attack is evaluated under a restricted threat model with bounded image-space perturbations and vision-encoder-only access. We transparently report the attack assumptions and evaluation protocols to facilitate reproducibility and responsible security research. References Alayrac et al. (2022) J. Alayrac, J. Donahue, P. Luc, A. Miech, I. Barr, Y. Hasson, K. Lenc, A. Mensch, K. Millican, M. Reynolds, et al. Flamingo: a visual language model for few-shot learning. Advances in neural information processing systems 35, p. 23716–23736. Cited by: §1. Alvar et al. (2025) S. R. Alvar, G. Singh, M. Akbari, and Y. Zhang Divprune: diversity-based visual token pruning for large multimodal models. In 2025 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), p. 9392–9401. Cited by: §2. Bai et al. (2025) S. Bai, Y. Cai, R. Chen, K. Chen, X. Chen, Z. Cheng, L. Deng, W. Ding, C. Gao, C. Ge, et al. Qwen3-vl technical report. arXiv preprint arXiv:2511.21631. Cited by: §B.1, Appendix E, §5.1. Bulat et al. (2026) A. Bulat, Y. Ouali, and G. Tzimiropoulos Compress & cache: vision token compression for efficient generation and retrieval. Advances in Neural Information Processing Systems 38, p. 31943–31968. Cited by: Appendix H, §1. Chen et al. (2026) J. Chen, X. Liu, Z. Wen, Y. Wang, S. Huang, and H. Chen Variation-aware vision token dropping for faster large vision-language models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, p. 3489–3499. Cited by: Appendix H, §2. Chen et al. (2024) L. Chen, H. Zhao, T. Liu, S. Bai, J. Lin, C. Zhou, and B. Chang An image is worth 1/2 tokens after layer 2: plug-and-play inference acceleration for large vision-language models. In European Conference on Computer Vision, p. 19–35. Cited by: §B.3, §1, §5.1. Cho et al. (2026) H. Cho, D. Baek, Y. Kim, and B. Ham Improving visual token reduction via rectifying distortions for efficient multimodal llm inference. arXiv preprint arXiv:2606.01711. Cited by: Appendix H. Dhouib et al. (2025) M. Dhouib, D. Buscaldi, S. Vanier, and A. Shabou Pact: pruning and clustering-based token reduction for faster visual language models. In 2025 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), p. 14582–14592. Cited by: §2. Endo et al. (2025) M. Endo, X. Wang, and S. Yeung-Levy Feather the throttle: revisiting visual token pruning for vision-language model acceleration. In 2025 IEEE/CVF International Conference on Computer Vision (ICCV), p. 22826–22835. Cited by: §2. Fu et al. (2025) C. Fu, P. Chen, Y. Shen, Y. Qin, M. Zhang, X. Lin, J. Yang, X. Zheng, K. Li, X. Sun, Y. Wu, R. Ji, C. Shan, and R. He MME: a comprehensive evaluation benchmark for multimodal large language models. In Advances in Neural Information Processing Systems, D. Belgrave, C. Zhang, H. Lin, R. Pascanu, P. Koniusz, M. Ghassemi, and N. Chen (Eds.), Vol. 38, Main Conference, p. . External Links: Document, Link Cited by: §B.2, §5.1. Gao et al. (2026) T. Gao, S. Zhao, S. Fang, S. Zhu, and C. Ma QuietPrune: query-guided early token pruning for vision-language models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, p. 3553–3562. Cited by: §2. Gu et al. (2026) S. Gu, J. Cui, W. Hu, Z. Shi, Z. Hu, and R. Hong Visual token compression enhances robustness of mllms. arXiv preprint arXiv:2607.22716. Cited by: §1, §2. Hu et al. (2025) K. Hu, W. Yu, L. Zhang, A. Robey, A. Zou, C. Xu, H. Hu, and M. Fredrikson Transferable adversarial attacks on black-box vision-language models. arXiv preprint arXiv:2505.01050. Cited by: §2. Huang et al. (2026) Y. Huang, F. Ma, Y. Shao, J. Guo, Z. Yu, L. Cui, and Q. Tian N\ " uwa: mending the spatial integrity torn by vlm token pruning. arXiv preprint arXiv:2602.02951. Cited by: Appendix H. Li et al. (2026a) A. Li, Y. Duan, J. Zhang, C. Ma, Y. Xie, G. Carneiro, M. Yaqub, and H. Wang Transprune: token transition pruning for efficient large vision-language model. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, p. 39529–39538. Cited by: Appendix H, §2. Li et al. (2023a) J. Li, D. Li, S. Savarese, and S. Hoi Blip-2: bootstrapping language-image pre-training with frozen image encoders and large language models. In International conference on machine learning, p. 19730–19742. Cited by: §1. Li et al. (2025) W. Li, Y. Yuan, J. Liu, D. Tang, S. Wang, J. Qin, J. Zhu, and L. Zhang Tokenpacker: efficient visual projector for multimodal llm. International Journal of Computer Vision 133 (10), p. 6794–6812. Cited by: §1. Li et al. (2023b) Y. Li, Y. Du, K. Zhou, J. Wang, X. Zhao, and J. Wen Evaluating object hallucination in large vision-language models. In Proceedings of the 2023 conference on empirical methods in natural language processing, p. 292–305. Cited by: §B.2, §5.1. Li et al. (2026b) Z. Li, Z. Li, P. Wang, L. Zhang, and H. Ai Vista-llm: decoupled query-guided visual token pruning for efficient long-video large language models. In Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), p. 13171–13187. Cited by: Appendix H, §2. Liu et al. (2024a) D. Liu, M. Yang, X. Qu, P. Zhou, X. Fang, K. Tang, Y. Wan, and L. Sun Pandora’s box: towards building universal attackers against real-world large vision-language models. Advances in Neural Information Processing Systems 37, p. 52127–52158. Cited by: §2. Liu et al. (2024b) H. Liu, C. Li, Y. Li, and Y. J. Lee Improved baselines with visual instruction tuning. In 2024 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), p. 26286–26296. Cited by: §B.1, §1, §5.1. Liu et al. (2025) J. Liu, J. Du, X. Liu, P. Tiwari, and M. Xu Self-calibrated consistency can fight back for adversarial robustness in vision-language models. arXiv preprint arXiv:2510.22785. Cited by: §2. Lu et al. (2023) D. Lu, Z. Wang, T. Wang, W. Guan, H. Gao, and F. Zheng Set-level guidance attack: boosting adversarial transferability of vision-language pre-training models. In 2023 IEEE/CVF International Conference on Computer Vision (ICCV), p. 102–111. Cited by: §2. Luo et al. (2024) H. Luo, J. Gu, F. Liu, and P. Torr An image is worth 1000 lies: adversarial transferability across prompts on vision-language models. arXiv preprint arXiv:2403.09766. Cited by: §2. Ma et al. (2026) J. Ma, S. Jin, and L. Dang Visual token compression via run-length pruning in multimodal large language models. Pattern Recognition Letters. Cited by: Appendix H. Mei et al. (2026a) H. Mei, Z. Wang, C. Xu, J. Guo, and M. Dong PA-attack: guiding gray-box attacks on lvlm vision encoders with prototypes and attention. arXiv preprint arXiv:2602.19418. Cited by: §2. Mei et al. (2026b) H. Mei, Z. Wang, S. You, M. Dong, and C. Xu Veattack: downstream-agnostic vision encoder attack against large vision language models. In International Conference on Learning Representations, Vol. 2026, p. 18135–18161. Cited by: §B.4, §C.1, §2, §5.1. Qi et al. (2024) X. Qi, K. Huang, A. Panda, P. Henderson, M. Wang, and P. Mittal Visual adversarial examples jailbreak aligned large language models. In Proceedings of the AAAI conference on artificial intelligence, Vol. 38, p. 21527–21536. Cited by: Appendix H. Radford et al. (2021) A. Radford, J. W. Kim, C. Hallacy, A. Ramesh, G. Goh, S. Agarwal, G. Sastry, A. Askell, P. Mishkin, J. Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, p. 8748–8763. Cited by: §B.1. Schlarmann and Hein (2023) C. Schlarmann and M. Hein On the adversarial robustness of multi-modal foundation models. In 2023 IEEE/CVF International Conference on Computer Vision Workshops (ICCVW), p. 3679–3687. Cited by: §1. Schlarmann et al. (2024) C. Schlarmann, N. D. Singh, F. Croce, and M. Hein Robust clip: unsupervised adversarial fine-tuning of vision embeddings for robust large vision-language models. arXiv preprint arXiv:2402.12336. Cited by: §2. Shang et al. (2025) Y. Shang, M. Cai, B. Xu, Y. J. Lee, and Y. Yan Llava-prumerge: adaptive token reduction for efficient large multimodal models. In 2025 IEEE/CVF International Conference on Computer Vision (ICCV), p. 22857–22867. Cited by: §B.3, §1, §5.1. Shayegani et al. (2024) E. Shayegani, Y. Dong, and N. Abu-Ghazaleh Jailbreak in pieces: compositional adversarial attacks on multi-modal language models. In International conference on learning representations, Vol. 2024, p. 30853–30885. Cited by: Appendix H. Shen et al. (2024) X. Shen, Y. Xiong, C. Zhao, L. Wu, J. Chen, C. Zhu, Z. Liu, F. Xiao, B. Varadarajan, F. Bordes, et al. Longvu: spatiotemporal adaptive compression for long video-language understanding. arXiv preprint arXiv:2410.17434. Cited by: Appendix H, §2. Sheng et al. (2025) L. Sheng, J. Liang, Z. Wang, and R. He R-tpt: improving adversarial robustness of vision-language models through test-time prompt tuning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, p. 29958–29967. Cited by: §2. Singh et al. (2019) A. Singh, V. Natarajan, M. Shah, Y. Jiang, X. Chen, D. Batra, D. Parikh, and M. Rohrbach Towards vqa models that can read. In 2019 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), p. 8309–8318. Cited by: §B.2, §5.1. Sun et al. (2026) G. Sun, Y. Wang, S. Ma, Y. Xie, Y. Cheng, Z. Tao, and J. Wang IF-prune: information-flow guided token pruning for efficient vision-language models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, p. 3522–3531. Cited by: §2. Sun et al. (2025) Y. Sun, H. Li, C. Xu, H. Zhou, C. Lin, R. T. Batista-Navarro, and J. Sun Does acceleration cause hidden instability in vision language models? uncovering instance-level divergence through a large-scale empirical study. In Proceedings of the 2025 Conference on Empirical Methods in Natural Language Processing, p. 8453–8467. Cited by: §3.1. Takezoe et al. (2026) R. Takezoe, Y. Li, Z. Bo, A. Hou, M. Guang, and K. Long Learnpruner: rethinking attention-based token pruning in vision language models. In International Conference on Learning Representations, Vol. 2026, p. 66381–66400. Cited by: §2. Wang et al. (2025a) H. Wang, Y. Nie, Y. Ye, Y. Wang, S. Li, H. Yu, J. Lu, and C. Huang Dynamic-vlm: simple dynamic visual token compression for videollm. In 2025 IEEE/CVF International Conference on Computer Vision (ICCV), p. 20812–20823. Cited by: §2. Wang et al. (2026a) S. Wang, X. Lyu, S. Yuan, J. Song, H. T. Shen, and L. Gao Understanding and mitigating token-pruning-induced vulnerabilities in VLMs. In Forty-third International Conference on Machine Learning, External Links: Link Cited by: §1, §2. Wang et al. (2025b) W. Wang, Z. Gao, L. Gu, H. Pu, L. Cui, X. Wei, Z. Liu, L. Jing, S. Ye, J. Shao, Z. Wang, Z. Chen, H. Zhang, G. Yang, H. Wang, Q. Wei, J. Yin, W. Li, E. Cui, G. Chen, Z. Ding, C. Tian, Z. Wu, J. Xie, Z. Li, B. Yang, Y. Duan, X. Wang, Z. Hou, H. Hao, T. Zhang, S. Li, X. Zhao, H. Duan, N. Deng, B. Fu, Y. He, Y. Wang, C. He, B. Shi, J. He, Y. Xiong, H. Lv, L. Wu, W. Shao, K. Zhang, H. Deng, B. Qi, J. Ge, Q. Guo, W. Zhang, S. Zhang, M. Cao, J. Lin, K. Tang, J. Gao, H. Huang, Y. Gu, C. Lyu, H. Tang, R. Wang, H. Lv, W. Ouyang, L. Wang, M. Dou, X. Zhu, T. Lu, D. Lin, J. Dai, W. Su, B. Zhou, K. Chen, Y. Qiao, W. Wang, and G. Luo InternVL3.5: advancing open-source multimodal models in versatility, reasoning, and efficiency. External Links: 2508.18265, Link Cited by: §B.1, Appendix E, §5.1. Wang et al. (2026b) Y. Wang, J. Wu, Z. Ni, L. Yang, Y. Liu, C. Yang, Y. Wen, L. He, X. Tang, H. Liu, et al. When token pruning is worse than random: understanding visual token information in vllms. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, p. 31910–31919. Cited by: §G.1. Wang et al. (2026c) Y. Wang, H. Zhang, Q. Huang, A. Cao, G. Fang, W. Wang, X. Jin, J. Song, M. Song, and X. Wang Rethinking token reduction for large vision-language models. arXiv preprint arXiv:2603.21701. Cited by: Appendix H, §2. Wang et al. (2025c) Z. Wang, C. Xie, B. Bartoldson, and B. Kailkhura Double visual defense: adversarial pre-training and instruction tuning for improving vision-language model robustness. arXiv preprint arXiv:2501.09446. Cited by: §2. Xie et al. (2025) P. Xie, Y. Bie, J. Mao, Y. Song, Y. Wang, H. Chen, and K. Chen Chain of attack: on the robustness of vision-language models against transfer-based adversarial attacks. In 2025 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), p. 14679–14689. Cited by: §2. Xing et al. (2024) L. Xing, Q. Huang, X. Dong, J. Lu, P. Zhang, Y. Zang, Y. Cao, C. He, J. Wang, F. Wu, et al. Pyramiddrop: accelerating your large vision-language models via pyramid visual redundancy reduction. arXiv preprint arXiv:2410.17247. Cited by: §1. Xu et al. (2026) T. Xu, H. Shi, and X. Gao SCoRe: salience-coverage reduction for vision token pruning in vision-language models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, p. 24686–24695. Cited by: §2. Yang et al. (2026) C. Yang, S. Lo, and Y. Liu Reroute, don’t remove: recoverable visual token routing for vision-language models. arXiv preprint arXiv:2606.12412. Cited by: Appendix H. Yang et al. (2025a) L. Yang, D. Shen, C. Cai, K. Chen, F. Yang, T. Gao, D. Zhang, and X. Li Libra-merging: importance-redundancy and pruning-merging trade-off for acceleration plug-in in large vision-language model. In 2025 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), p. 9402–9412. Cited by: §2. Yang et al. (2025b) S. Yang, Y. Chen, Z. Tian, C. Wang, J. Li, B. Yu, and J. Jia Visionzip: longer is better but not necessary in vision language models. In 2025 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), p. 19792–19802. Cited by: §B.3, §1, §5.1. Ye et al. (2025) X. Ye, Y. Gan, Y. Ge, X. Zhang, and Y. Tang Atp-llava: adaptive token pruning for large vision language models. In 2025 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), p. 24972–24982. Cited by: Appendix H, §2. Yin et al. (2023) Z. Yin, M. Ye, T. Zhang, T. Du, J. Zhu, H. Liu, J. Chen, T. Wang, and F. Ma Vlattack: multimodal adversarial attacks on vision-language tasks via pre-trained models. Advances in Neural Information Processing Systems 36, p. 52936–52956. Cited by: §1. Yu et al. (2026) H. Yu, W. Li, X. Qu, S. Wang, J. Chen, and J. Zhu Visiontrim: unified vision token compression for training-free mllm acceleration. arXiv preprint arXiv:2601.22674. Cited by: §2. Zhang et al. (2025a) J. Zhang, J. Ye, X. Ma, Y. Li, Y. Yang, Y. Chen, J. Sang, and D. Yeung Anyattack: towards large-scale self-supervised adversarial attacks on vision-language models. In 2025 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), p. 19900–19909. Cited by: Appendix H, §2. Zhang et al. (2022) J. Zhang, Q. Yi, and J. Sang Towards adversarial attack on vision-language pre-training models. In Proceedings of the 30th ACM international conference on multimedia, p. 5005–5013. Cited by: §1. Zhang et al. (2025b) Q. Zhang, A. Cheng, M. Lu, R. Zhang, Z. Zhuo, J. Cao, S. Guo, Q. She, and S. Zhang Beyond text-visual attention: exploiting visual cues for effective token pruning in vlms. In 2025 IEEE/CVF International Conference on Computer Vision (ICCV), p. 20857–20867. Cited by: §B.3, §2, §5.1. Zhang et al. (2026a) X. Zhang, Z. Zhang, L. Y. Zhang, Y. Zhang, G. Tao, and S. Pan Less is more–until it breaks: security pitfalls of vision token compression in large vision-language models. arXiv preprint arXiv:2601.12042. Cited by: §B.4, §1, §2, §5.1. Zhang et al. (2026b) X. Zhang, L. Bai, T. Zhang, Y. Zhang, Q. Ye, Y. Zhao, R. Du, and H. Hu Grounding-driven attack: improving encoder-based adversarial transferability against large vision-language models. arXiv preprint arXiv:2602.09431. Cited by: §2. Zhang et al. (2026c) X. Zhang, H. Liu, L. Bai, H. Wang, Q. Ye, T. Zhang, and H. Hu On the adversarial robustness of large vision-language models under visual token compression. arXiv preprint arXiv:2601.21531. Cited by: §B.4, §1, §2, §5.1. Zhang et al. (2024) Y. Zhang, C. Fan, J. Ma, W. Zheng, T. Huang, K. Cheng, D. Gudovskiy, T. Okuno, Y. Nakata, K. Keutzer, et al. Sparsevlm: visual token sparsification for efficient vision-language model inference. arXiv preprint arXiv:2410.04417. Cited by: §1. Appendix Contents Appendix A. CIRA Optimization Procedure.A Appendix B. Detailed Experimental Setup.B Appendix C. Controlled Retained-Set Allocation Diagnostics.C Appendix D. Sensitivity Analyses.D Appendix E. Cross-Model Scope.E Appendix F. Selection Stabilization Defense.F Appendix G. Complementary Task-Utility Results.G Appendix H. Limitation Discussion and Future Work.H Appendix I. Qualitative Case Studies.I Appendix A CIRA Optimization Procedure Algorithm 1 gives the image-space optimization induced by CIRA’s two encoder-side objectives. Clean features and priority scores are cached once; adversarial scores, ranks, and displacement weights are refreshed at every step. The priority rule P maps encoder attention to the token-score vector used by GSH. Algorithm 1 Projected sign-gradient optimization of CIRA. 1: Image x; encoder ℰE; priority rule P 2: Candidate interval [Kmin,Kmax][K_ ,K_ ]; ϵ,ϵs,ϵhε, _s, _h 3: Step size α; steps T; preservation weight λ 4: Adversarial image xadvx^adv 5: ←Kmin,…,KmaxK←\K_ ,…,K_ \ 6: (c,c)←(ℰ⁡(x),⁡(x))(H^c,s^c)←(E(x),P(x)) 7: c←Rank↓⁡(c)r^c _ (s^c) 8: c←ℛ⁡(c)u^c (s^c) 9: Initialize δ(0)←δ^(0) 0 10: for τ=0,…,T−1τ=0,…,T-1 do 11: x(τ)←x+δ(τ)x^(τ)← x+δ^(τ) 12: (a,a)←(ℰ⁡(x(τ)),⁡(x(τ)))(H^a,s^a)←(E(x^(τ)),P(x^(τ))) 13: ℒGSH←Alignϵs⁡(a,c)L_GSH _ _s(s^a,u^c) 14: a←Rank↓⁡(a)r^a _ (s^a) 15: wi←sg(PrK∼Unif⁡()[ric≤K<ria])w_i \! ( _K (K)[r_i^c≤ K<r_i^a] ) 16: di←12(1−cos⁡(ic,ia))d_i← 12(1- (h_i^c,h_i^a)) 17: ℒHEP←−∑iwidimax⁡(∑iwi,ϵh)L_HEP←- _iw_id_i ( _iw_i, _h) 18: ℒCIRA←ℒGSH+λℒHEPL_CIRA _GSH+ _HEP 19: g←sign⁡(∇δℒCIRA)g ( _δL_CIRA) 20: δ(τ+1)←ΠΔϵ(x)(δ(τ)+αg)δ^(τ+1)← _ _ε(x)(δ^(τ)+α g) 21: end for 22: return xadv←x+δ(T)x^adv← x+δ^(T) Sorting remains outside of the gradient path, sg⁡(⋅)sg(·) blocks gradients through the discrete weights, and Δϵ(x)=δ:∥δ∥∞≤ϵ,x+δ∈[0,1]d _ε(x)=\δ:\|δ\|_∞≤ε,x+δ∈[0,1]^d\ is the feasible perturbation set. The update therefore uses only the vision encoder, priority-scoring rule, and candidate compression-budget interval. Appendix B Detailed Experimental Setup B.1 Models LLaVA-v1.5-7B. Our primary model is LLaVA-v1.5-7B (Liu et al., 2024b) with a CLIP ViT-L/14-336 vision encoder (Radford et al., 2021). The released model feeds 576 penultimate-layer patch tokens to its multimodal projector. We use LLaVA for the full benchmark–compressor matrix and for the mechanism, ablation, sensitivity, and defense studies. Additional LVLM families. We additionally evaluate Qwen3-VL-8B-Instruct (Bai et al., 2025) and InternVL3.5-8B (Wang et al., 2025b). Qwen3-VL uses its native dynamic resolution and spatial merger, so its visual-token count varies by image. InternVL uses one 448×448448× 448 tile, producing 256 tokens after spatial downsampling. Accordingly, we use family-specific evaluation budgets of 96,64,32\96,64,32\ for Qwen3-VL and 128,64,32\128,64,32\ for InternVL. CIRA is optimized separately for each model using its native vision encoder. B.2 Benchmarks We use fixed, randomly sampled 1,000-pair subsets from POPE (Li et al., 2023b), TextVQA (Singh et al., 2019), and MME (Fu et al., 2025). POPE evaluates object hallucination through balanced object-presence questions; TextVQA tests reading of text in natural images and provides ten reference answers per question; MME covers perception and cognition with binary questions. Within each model–benchmark setting, all attacks are evaluated on the same image–question pairs using a common prompt, decoding protocol, and answer evaluator. POPE and MME outputs are lowercased, stripped of punctuation, and matched by the first normalized yes/no token. TextVQA uses its EvalAI-style normalizer and accepts a match to any reference answer. These task-specific rules provide the per-example binary correctness required by the paired CSF definition. For TextVQA, we use a match to any normalized reference rather than the benchmark-level soft agreement score; MME is evaluated per question rather than by its aggregate category score. The same evaluators are used for clean, full-token, and compressed inference. B.3 Compression Mechanisms VisionZip. VisionZip (Yang et al., 2025b) retains high-attention tokens and merges the remainder around uniformly sampled contextual tokens. Compression budgets K=32,64,128,192K=32,64,128,192 use (dominant, contextual) counts (27,5)(27,5), (54,10)(54,10), (108,20)(108,20), and (162,30)(162,30), respectively. VisPruner. VisPruner (Zhang et al., 2025b) combines attention-based importance with feature diversity. Half of each budget is assigned to important tokens and the remainder to diverse tokens. PruMerge. PruMerge (Shang et al., 2025) selects attention-ranked representatives and merges nearby tokens by feature similarity. We use K−1K-1 representatives together with one attention-weighted residual aggregate, yielding exactly K output tokens. FastV. FastV (Chen et al., 2024) ranks visual tokens using language-model attention and removes low-ranked tokens after layer 2. These configurations are used for the primary LLaVA evaluation. For Qwen3-VL and InternVL, the corresponding compression rules are instantiated on their native visual-token interfaces while preserving each method’s selection and aggregation principle. Throughout the evaluation, compression budget K denotes the number of post-compression visual tokens passed to subsequent computation, whether obtained through selection, merging, or both. B.4 Implementation Inference. All experiments run on one NVIDIA GeForce RTX 4090 GPU. Decoding is deterministic (do_sample=False) with at most 64 new tokens, using each model’s native image processor and conversation template. Priority-score instantiation. CIRA derives token priorities from model-native late-layer visual attention. Let ℒscoreL_score denote the encoder layers used to compute token priority scores. We use the final, third-to-last, and fifth-to-last encoder blocks, corresponding to relative layer indices (−1,−3,−5)(-1,-3,-5). For the CLIP encoder in the primary setting, the score from Section 4.2 is si(x)=1|ℒscore|∑ℓ∈ℒscore∑h=1Hα0,iℓ,h(x),i=1,…,N,s_i(x)= 1|L_score| _ _score _h=1^H _0,i ,h(x), i=1,…,N, (13) where α0,iℓ,h _0,i ,h is attention from the class token to patch i at head h and visual layer ℓ , and H is the number of attention heads. InternVL derives token priorities from class-to-patch visual self-attention and aggregates them over its spatial-downsampling groups. For Qwen3-VL, which lacks a class-token routing interface, we use the mean incoming attention over visual queries and aggregate scores within its native spatial-merger groups. For both models, priorities are averaged over the same three relative encoder layers to produce one score per downstream visual token. Preservation features. For LLaVA, HEP measures tokenwise cosine distance between final CLIP patch features after post-layer normalization. For Qwen3-VL, it averages the tokenwise distances of the merged main visual features and the DeepStack features. For InternVL, it uses the visual tokens returned by the model’s feature-extraction module after pixel shuffle and MLP projection. Optimization. All downstream-agnostic attacks use an ℓ∞ _∞ perturbation budget of 4/2554/255 and 100 optimization steps. CIRA starts from the clean image and uses projected sign-gradient ascent with step size 1/2551/255, λ=0.8λ=0.8, and candidate budget range [Kmin,Kmax]=[32,192][K_ ,K_ ]=[32,192]. The remaining hyperparameters of VEAttack (Mei et al., 2026b) and CAGE (Zhang et al., 2026c) follow their released settings; CAGE uses its released budget interval [16,192][16,192]. CAA† (Zhang et al., 2026a) optimizes a question-conditioned objective at language-model layer 2 with ϵ=4/255ε=4/255 and 100 steps. One adversarial image is generated per image–question pair and then evaluated across compressors and budgets without re-optimization. Evaluation protocol. Each downstream-agnostic method generates one adversarial image per clean input. After optimization, the image is fixed and evaluated across all corresponding questions, compressors, and budgets. Clean eligibility is determined separately for each compressor–budget setting using KS_K in equation 1, while Full ASR is computed over fullS_full defined in Section 5.1. Appendix C Controlled Retained-Set Allocation Diagnostics The diagnostic in Section 3 isolates retained-set allocation through counterfactual exchanges at fixed first-observed CSF states. Answer-aware scores are used only for post-hoc exchange selection after the failure state is fixed and are not part of CIRA optimization. C.1 Diagnostic Cohort and Fixed Failure State Diagnostic setting. We use LLaVA–VisionZip on POPE, TextVQA, and MME at K∈32,64,128,192K∈\32,64,128,192\. VisionZip retains dominant patch tokens individually and aggregates the remainder into contextual tokens. We call the individually retained patches direct tokens; their capacity DKD_K is smaller than the total compressed budget K. Common clean cohort. Let j index an image–question observation, comprising image xjx_j, its associated question, and reference answer; index i is reserved for visual tokens. We retain only observations that are correct under full-token inference and under every evaluated compressed setting: cj(xj)=1,cj,K(xj)=1,∀K∈192,128,64,32,c_j(x_j)=1, c_j,K(x_j)=1, ∀ K∈\192,128,64,32\, (14) where cjc_j and cj,Kc_j,K denote full-token and compressed correctness for observation j. Every trajectory therefore begins from a common state without pre-existing compression errors in any evaluated path. Attack trajectory. We generate clean-initialized feature-objective VEAttack (Mei et al., 2026b) trajectories with ϵ=4/255ε=4/255, step size 1/2551/255, and at most 100100 steps, evaluating predictions every 1010 steps. For each observation and setting, we select the earliest evaluated checkpoint satisfying cj(xjadv)=1,cj,K(xjadv)=0.c_j(x_j^adv)=1, c_j,K(x_j^adv)=0. (15) This first-observed CSF supplies the fixed state for all subsequent counterfactuals. Balanced diagnostic cohort. Within each comparison, eligible images are ordered using a fixed sampling order and truncated to the smallest available image count across cells. All eligible questions associated with the retained images are then included. C.2 Controlled Counterfactual Reallocation At the fixed adversarial state, exchange identities are selected using retrospective answer support computed from the adversarial representations. For adversarial token i, we apply a multiplicative gate h~i=gihiadv h_i=g_ih_i^adv and define ai=−∂ℒNLL(y∣H~adv,q)∂gi|gi=1.a_i=- . _NLL(y H^adv,q)∂ g_i |_g_i=1. (16) Larger aia_i indicates that token i provides stronger support for the reference answer. Let RKR_K be the direct-token set and OKO_K its complement. We rank OKO_K in descending and RKR_K in ascending order of aia_i. For exchange size r, RK,rguide=(RK∖Lr)∪Hr,R_K,r^guide= (R_K L_r )∪ H_r, (17) where HrH_r contains the r highest-support omitted tokens and LrL_r the r lowest-support retained tokens. We evaluate the nominal schedule ρ∈0,5%,7.5%,10%,15%ρ∈\0,5\%,7.5\%,10\%,15\%\, with rK(ρ)=round⁡(ρDK),r_K(ρ)=round(ρ D_K), (18) where DKD_K is the number of direct slots. For ρ=(0,5%,7.5%,10%,15%)ρ=(0,5\%,7.5\%,10\%,15\%), this gives r32=(0,1,2,3,4)r_32=(0,1,2,3,4), r64=(0,3,4,5,8)r_64=(0,3,4,5,8), r128=(0,5,8,11,16)r_128=(0,5,8,11,16), and r192=(0,8,12,16,24)r_192=(0,8,12,16,24). The matched-random control independently permutes the retained and omitted pools. Each observation uses 40 nested paths, with larger exchanges extending smaller ones. VisionZip merging and aggregation are recomputed after every exchange. C.3 Retained-Set Allocation Sensitivity and Cumulative Recovery Let YjKz,p(ρ)∈0,1Y_jKz,p(ρ)∈\0,1\ indicate correctness for observation j under condition z, path p, and ratio ρ. Guided reallocation has one path; the random control has P=40P=40 nested paths. Because recovery need not persist under a larger exchange, curves report cumulative recovery: CjKz,p(ρ)=maxρ′≤ρ⁡YjKz,p(ρ′).C_jKz,p(ρ)= _ρ ≤ρY_jKz,p(ρ ). (19) Thus, CjKz,p(ρ)C_jKz,p(ρ) asks whether a CSF is corrected at any schedule point up to the displayed nominal ratio. Estimand and aggregation. Random paths are averaged within each observation before cell aggregation: C¯jK,rand(ρ)=1P∑p=1PCjK,rand,p(ρ), C_jK,rand(ρ)= 1P _p=1^PC_jK,rand,p(ρ), (20) and we set C¯jK,guide(ρ)=CjK,guide,1(ρ) C_jK,guide(ρ)=C_jK,guide,1(ρ). Let G denote the 12 dataset–budget cells and dKJ_dK their analyzed image–question observations. The reported equal-cell average is μ^z(ρ)=1||∑(d,K)∈[1|dK|∑j∈dKC¯jK,z(ρ)]. μ_z(ρ)= 1|G| _(d,K) [ 1|J_dK| _j _dK C_jK,z(ρ) ]. (21) The inner mean averages over image–question observations within each cell, while the outer mean assigns equal weight to every dataset–budget cell. The paired contrast is Δ^(ρ)=μ^guide(ρ)−μ^rand(ρ) (ρ)= μ_guide(ρ)- μ_rand(ρ). Table 5: Cumulative guided (G) and matched-random (R) recovery (%) by compression budget and exchange ratio, with Δ=G−R =G-R reported in percentage points and Avg. denoting the equal-cell average. K ρ=5%ρ=5\% ρ=7.5%ρ=7.5\% ρ=10%ρ=10\% ρ=15%ρ=15\% G R G R G R G R 32 29.1 12.3 +16.8+16.8 40.6 19.3 +21.4+21.4 45.5 23.6 +21.9+21.9 51.1 27.5 +23.6+23.6 64 39.1 16.5 +22.6+22.6 43.4 21.8 +21.6+21.6 44.1 25.9 +18.3+18.3 49.7 32.5 +17.2+17.2 128 35.8 13.8 +22.0+22.0 45.8 20.9 +24.9+24.9 49.4 25.4 +24.0+24.0 54.9 30.8 +24.1+24.1 192 44.4 15.6 +28.8+28.8 50.8 22.3 +28.4+28.4 53.6 27.0 +26.6+26.6 57.9 33.0 +24.9+24.9 Avg. 37.1 14.5 +22.6+22.6 45.2 21.1 +24.1+24.1 48.2 25.5 +22.7+22.7 53.4 31.0 +22.5+22.5 The average guided–random advantage ranges from 22.522.5 to 24.124.1 p across all nonzero exchange ratios. C.4 Factorial Decomposition of Recovery At ρ∗=7.5%ρ^*=7.5\%, we separate which omitted tokens enter from which retained tokens leave. Incoming tokens are high-support evidence or random omissions; outgoing tokens are low-support retained tokens or random ones: Random outLow-support outRandom inRRLEvidence inEREL array[]c|c&Random out&Low-support out\\ in&R&RL\\ Evidence in&ER&EL array Let QjK,z∈[0,1]Q_jK,z∈[0,1] denote exact recovery under arm z, averaged over random paths where applicable. We compute the two main effects and their interaction per observation before equal-cell aggregation: ejK e_jK =(QjK,ER−QjK,RR)+(QjK,EL−QjK,RL)2, = (Q_jK,ER-Q_jK,R)+(Q_jK,EL-Q_jK,RL)2, (22) ujK u_jK =(QjK,RL−QjK,RR)+(QjK,EL−QjK,ER)2, = (Q_jK,RL-Q_jK,R)+(Q_jK,EL-Q_jK,ER)2, (23) ηjK _jK =QjK,EL−QjK,ER−QjK,RL+QjK,RR. =Q_jK,EL-Q_jK,ER-Q_jK,RL+Q_jK,R. (24) The four arm rates and observation-level effects are aggregated with equation 21; Table 6 reports the resulting point estimates by budget. Table 6: Exact recovery and factorial effects at ρ∗=7.5%ρ^*=7.5\% by compression budget, with Avg. denoting the equal-cell average. K Exact Recovery (%) Factorial Effect (p) R ER RL EL Evidence Restoration Low-Support Removal Interaction 32 16.1 34.9 20.1 38.6 +18.6+18.6 +3.8+3.8 −0.3-0.3 64 18.4 34.0 25.4 38.5 +14.4+14.4 +5.7+5.7 −2.4-2.4 128 17.2 33.5 31.9 44.4 +14.4+14.4 +12.8+12.8 −3.9-3.9 192 17.9 31.1 34.5 46.4 +12.6+12.6 +15.9+15.9 −1.3-1.3 Avg. 17.4 33.4 28.0 42.0 +15.0+15.0 +9.6+9.6 −2.0-2.0 Evidence restoration is positive at every budget, while the removal effect increases with K. Interaction estimates are negative and smaller in magnitude than either main effect at every budget. C.5 Representation-Drift Moderation For a restored evidence token i, let hicleanh_i^clean and hiadvh_i^adv be its clean and adversarial encoder representations. We define didiag=1−cos⁡(hiclean,hiadv)2.d_i^diag= 1- (h_i^clean,h_i^adv )2. (25) Let EjKE_jK denote the high-support omitted tokens restored for observation j at budget K under the ρ∗=7.5%ρ^*=7.5\% intervention. We define their mean representation drift as DjK=1|EjK|∑i∈EjKdidiag.D_jK= 1|E_jK| _i∈ E_jKd_i^diag. (26) Clean representations are used only in this post-hoc diagnostic. Let D¯dK D_dK and e¯dK e_dK be within-cell means. We estimate the common within-cell association between drift and the evidence-restoration effect ejKe_jK from equation 24: β^=∑d,K∑j∈dK(DjK−D¯dK)(ejK−e¯dK)∑d,K∑j∈dK(DjK−D¯dK)2. β= _d,K _j _dK (D_jK- D_dK ) (e_jK- e_dK ) _d,K _j _dK(D_jK- D_dK)^2. (27) Thus, β β uses only within-cell variation. Drift tertiles provide a grouped summary, and the continuous slope summarizes the corresponding within-cell association. Table 7: Evidence-restoration effects across within-cell representation-drift tertiles and compression budgets, with mean drift defined as 100×100× half-cosine distance and Gap as the lower-minus-upper effect. K Lower Middle Upper Gap (p) Drift (%) Effect (p) Drift (%) Effect (p) Drift (%) Effect (p) 32 26.1 +28.5+28.5 38.0 +20.8+20.8 45.2 +6.1+6.1 +22.3+22.3 64 31.0 +18.2+18.2 39.3 +7.0+7.0 45.5 +18.0+18.0 +0.2+0.2 128 34.0 +18.9+18.9 40.4 +12.7+12.7 45.1 +11.5+11.5 +7.4+7.4 192 34.4 +16.6+16.6 40.2 +15.3+15.3 45.1 +5.4+5.4 +11.2+11.2 Avg. 31.4 +20.5+20.5 39.5 +14.0+14.0 45.2 +10.3+10.3 +10.3+10.3 The estimated within-cell association is −8.1-8.1 p per 0.10.1 increase in DjKD_jK. Budget-specific tertiles are not uniformly monotone and are therefore interpreted descriptively. The retained-set intervention establishes that changing token allocation can causally alter correctness within this fixed cohort, whereas the drift–recovery analysis supports an association rather than causal mediation. Appendix D Sensitivity Analyses We test whether CIRA’s selective operating point depends on individual design or optimization choices. Unless stated otherwise, each analysis varies one choice on LLaVA-v1.5-7B, POPE, and VisionZip while retaining the evaluation definitions in Section 5.1. D.1 Preservation Weight The relative loss weight λ in equation 9 controls the tradeoff between failure induction and full-token preservation. We summarize this operating tradeoff by Selective Gap (Avg. CSFR minus Full ASR), used only as a configuration score because the two metrics have different conditioning sets. Table 8 shows that increasing λ reduces Full ASR while retaining substantial CSFR. Selective Gap remains stable for λ∈[0.6,1.0]λ∈[0.6,1.0], with λ=0.8λ=0.8 attaining the highest observed value. We therefore use λ=0.8λ=0.8 throughout the evaluation. Table 8: Sensitivity of Full ASR, CSFR, and Selective Gap to the preservation weight λ on POPE. Best values in each row are shown in bold. Metric =0.2 λ=0.2 =0.4 λ=0.4 =0.6 λ=0.6 =0.8 λ=0.8 =1.0 λ=1.0 Full ASR↓ 9.57 6.03 5.79 4.96 4.61 CSFR@192↑ 8.59 9.08 9.20 7.61 6.50 CSFR@128↑ 15.67 15.30 15.55 14.05 13.68 CSFR@64↑ 22.07 22.07 22.47 22.47 21.94 CSFR@32↑ 27.56 27.26 25.93 26.96 27.41 Avg. CSFR↑ 18.47 18.43 18.29 17.77 17.38 Selective Gap↑ 8.90 12.40 12.50 12.81 12.77 D.2 Candidate Compression-Budget Interval The admissible interval [Kmin,Kmax][K_ ,K_ ] determines the budget-marginal displacement weights in equation 6. Table 9 varies this interval while keeping the evaluation budgets fixed at K∈192,128,64,32K∈\192,128,64,32\. Selectivity varies modestly across the six candidate intervals. We use [32,192][32,192] as the default because it matches the evaluation range; its Selective Gap is within 0.280.28 p of the best observed value. Table 9: Sensitivity of Full ASR, CSFR, and Selective Gap to the candidate compression-budget interval on POPE. Best values in each row are shown in bold. Metric Default Vary KminK_ Vary KmaxK_ [32,192] [32,192] [16,192] [16,192] [64,192] [64,192] [,] [32,64] [32,128] [32,128] [32,384] [32,384] Full ASR↓ 4.96 4.62 5.56 5.68 4.97 5.33 CSFR@192↑ 7.61 7.38 6.40 8.86 7.63 6.52 CSFR@128↑ 14.05 14.86 12.86 13.23 14.73 12.86 CSFR@64↑ 22.47 21.14 21.68 23.01 19.95 22.61 CSFR@32↑ 26.96 27.43 28.17 29.20 26.55 25.22 Avg. CSFR↑ 17.77 17.70 17.28 18.57 17.21 16.80 Selective Gap↑ 12.81 13.09 11.71 12.89 12.24 11.48 D.3 Perturbation Budget and Optimization Steps Figure 5: Sensitivity of Full ASR and CSFR to the perturbation budget and number of optimization steps on POPE. Figure 5 varies ϵ∈2,4,6,8/255ε∈\2,4,6,8\/255 and the number of optimization steps in 50,80,100,150\50,80,100,150\; the step size is set to ϵ/4ε/4. Across all 16 configurations, Full ASR remains between 4.14%4.14\% and 6.38%6.38\%, whereas Avg. CSFR ranges from 14.93%14.93\% to 20.82%20.82\%. Thus, substantial compression-specific failure rates persist while full-token degradation remains limited, and the default ϵ=4/255ε=4/255, 100-step setting lies within a stable selective region rather than at an isolated optimum. D.4 Scoring-Layer Configuration Study The primary score in equation 13 averages class-to-patch attention over the scoring-layer configuration ℒscoreL_score. We compare 14 single-layer, contiguous, and spaced multi-layer configurations using 100-example design subsets from POPE, TextVQA, and MME, evaluated with VisionZip, PruMerge, and VisPruner. We select the configuration with the highest mean Selective Gap across these nine dataset–compressor environments and use it throughout the reported evaluation. Figure 6: Selectivity and cross-layer exclusion stability across scoring-layer configurations. As shown in Figure 6, the selected spaced late-layer triplet −1,−3,−5\-1,-3,-5\ achieves 23.45%23.45\% Avg. CSFR with 5.59%5.59\% Full ASR, corresponding to a 17.8617.86 p Selective Gap. Its exclusion rate for clean high-priority tokens also varies less across observer layers than that of the single-layer −2\-2\ configuration (range 0.0550.055 versus 0.2420.242). The negative association between cross-layer variation and selectivity (ρs=−0.70 _s=-0.70) indicates that configurations with more stable exclusion behavior across depth tend to exhibit higher selectivity. This association is descriptive rather than causal. Appendix E Cross-Model Scope The main experiments establish CIRA across multiple compressors and budgets on LLaVA. We next test whether paired selectivity persists when the vision encoder, multimodal interface, and language model change together, using Qwen3-VL-8B-Instruct (Bai et al., 2025) and InternVL3.5-8B (Wang et al., 2025b). Here K denotes the number of compressed visual tokens passed to subsequent computation in each model’s native interface. For compact reporting, we use K1/64/32K_1/64/32, where K1=96K_1=96 for Qwen3-VL and K1=128K_1=128 for InternVL. Avg. CSFR is the arithmetic mean over the three budgets within each model family. Table 10 reports the corresponding CSFR and Full ASR results, while clean accuracy is provided in Table 11. Table 10: CSFR and Full ASR on Qwen3-VL-8B-Instruct and InternVL3.5-8B. Compressor entries report CSFR at K1/64/32K_1/64/32 followed by their average. Bold and underlined values denote the best and second-best downstream-agnostic results, respectively; CAA† is shown as a stronger-access reference and excluded from these rankings. Dataset Attack Full ASR (%)↓ VisionZip K1/64/32/Avg.K_1/64/32/Avg.↑ VisPruner K1/64/32/Avg.K_1/64/32/Avg.↑ PruMerge K1/64/32/Avg.K_1/64/32/Avg.↑ FastV K1/64/32/Avg.K_1/64/32/Avg.↑ Qwen3-VL-8B-Instruct (K1=96K_1=96) POPE VEAttack 45.93 4.63 / 5.83 / 10.06 / 6.84 2.15 / 2.78 / 4.67 / 3.20 3.91 / 5.01 / 7.49 / 5.47 2.32 / 3.95 / 4.60 / 3.62 CAGE 46.47 7.13 / 8.13 / 10.19 / 8.48 4.65 / 5.68 / 8.32 / 6.22 7.23 / 8.23 / 8.95 / 8.14 4.65 / 6.25 / 11.49 / 7.46 CAA† 4.40 4.03 / 5.82 / 11.31 / 7.05 2.61 / 4.86 / 8.70 / 5.39 2.47 / 3.33 / 7.06 / 4.29 8.44 / 11.61 / 14.80 / 11.61 CIRA 10.66 7.59 / 9.82 / 14.14 / 10.52 4.98 / 6.56 / 11.22 / 7.59 3.65 / 4.39 / 7.94 / 5.33 8.56 / 10.08 / 13.07 / 10.57 TextVQA VEAttack 80.14 10.11 / 10.03 / 3.58 / 7.91 10.14 / 7.75 / 1.98 / 6.62 10.22 / 7.87 / 3.34 / 7.14 3.54 / 3.58 / 3.65 / 3.59 CAGE 66.03 23.96 / 22.56 / 15.22 / 20.58 19.59 / 16.71 / 10.62 / 15.64 21.54 / 18.12 / 16.56 / 18.74 12.38 / 11.00 / 12.41 / 11.93 CAA† 8.24 46.37 / 42.86 / 35.22 / 41.48 29.28 / 31.23 / 32.84 / 31.12 31.69 / 30.19 / 31.91 / 31.27 43.81 / 42.20 / 38.69 / 41.57 CIRA 16.37 51.65 / 52.13 / 45.37 / 49.72 30.63 / 28.81 / 34.57 / 31.34 34.88 / 32.13 / 35.20 / 34.07 38.31 / 36.57 / 34.31 / 36.40 MME VEAttack 44.11 7.14 / 6.59 / 9.29 / 7.67 5.95 / 6.51 / 6.29 / 6.25 7.60 / 8.31 / 9.69 / 8.54 2.58 / 3.84 / 6.58 / 4.33 CAGE 41.60 11.67 / 7.78 / 10.19 / 9.88 11.41 / 7.83 / 9.37 / 9.54 12.47 / 11.87 / 15.09 / 13.14 4.42 / 6.64 / 8.53 / 6.53 CAA† 4.76 8.27 / 9.96 / 12.35 / 10.19 7.37 / 7.72 / 10.34 / 8.48 9.71 / 9.72 / 12.08 / 10.51 9.16 / 11.71 / 15.66 / 12.18 CIRA 12.43 8.34 / 7.79 / 12.79 / 9.64 8.82 / 7.37 / 9.91 / 8.70 7.14 / 5.26 / 7.05 / 6.49 5.28 / 6.77 / 10.21 / 7.42 InternVL3.5-8B (K1=128K_1=128) POPE VEAttack 51.46 6.01 / 8.17 / 10.35 / 8.18 4.20 / 7.44 / 7.91 / 6.52 5.28 / 7.56 / 9.03 / 7.29 7.13 / 11.29 / 12.99 / 10.47 CAGE 45.01 7.67 / 13.25 / 17.45 / 12.79 3.82 / 5.61 / 8.19 / 5.87 3.48 / 5.33 / 8.49 / 5.77 9.68 / 13.31 / 16.35 / 13.11 CAA† 9.85 4.35 / 7.50 / 8.78 / 6.88 1.90 / 4.14 / 6.66 / 4.23 0.77 / 2.21 / 2.65 / 1.88 7.26 / 20.03 / 28.32 / 18.54 CIRA 11.68 4.87 / 7.90 / 11.65 / 8.14 3.04 / 4.01 / 8.42 / 5.16 2.83 / 3.26 / 4.75 / 3.61 3.82 / 8.60 / 12.26 / 8.23 TextVQA VEAttack 56.92 7.50 / 12.77 / 14.81 / 11.69 3.83 / 7.17 / 11.67 / 7.56 7.52 / 10.50 / 11.86 / 9.96 5.24 / 7.37 / 12.68 / 8.43 CAGE 55.24 11.00 / 16.23 / 16.24 / 14.49 6.20 / 10.99 / 11.41 / 9.53 6.64 / 10.37 / 13.50 / 10.17 4.94 / 9.89 / 14.83 / 9.89 CAA† 5.03 10.58 / 22.73 / 23.58 / 18.96 13.64 / 15.66 / 23.68 / 17.66 5.25 / 9.51 / 10.53 / 8.43 10.18 / 33.63 / 45.69 / 29.84 CIRA 12.03 33.17 / 36.80 / 36.93 / 35.63 10.55 / 17.00 / 21.58 / 16.38 17.35 / 21.20 / 15.79 / 18.11 10.03 / 25.36 / 38.28 / 24.56 MME VEAttack 40.63 2.57 / 4.44 / 4.91 / 3.98 2.27 / 4.05 / 4.44 / 3.58 3.28 / 3.93 / 5.71 / 4.31 1.72 / 3.27 / 5.35 / 3.45 CAGE 37.25 7.13 / 9.14 / 9.16 / 8.48 4.07 / 4.93 / 6.32 / 5.10 4.38 / 5.73 / 7.82 / 5.98 4.70 / 7.02 / 8.43 / 6.72 CAA† 5.53 2.92 / 3.58 / 8.11 / 4.87 2.41 / 6.19 / 8.31 / 5.64 2.81 / 2.81 / 5.04 / 3.55 3.44 / 13.68 / 18.88 / 12.00 CIRA 5.76 6.88 / 10.63 / 12.50 / 10.00 3.98 / 7.45 / 8.72 / 6.72 5.37 / 5.61 / 7.17 / 6.05 4.01 / 6.54 / 11.38 / 7.31 Across both additional model families, CIRA continues to induce compression-specific failures while keeping Full ASR substantially below those of VEAttack and CAGE. The pattern is strongest on TextVQA, where CIRA achieves the highest Avg. CSFR among downstream-agnostic attacks for every compressor on both model families. Results on POPE and MME are more heterogeneous across compressors, but compression-specific failure induction remains observable under target-encoder-only access. Together, these results show that CIRA’s compression-selective behavior extends beyond LLaVA to distinct vision encoders and native visual-token interfaces. Appendix F Selection Stabilization Defense Translation-Consensus Selection (TCS) stabilizes priority rankings by aggregating aligned scores across spatially translated views. F.1 Translation-Consensus Selection For pixel displacement d, define =T0,0,Td,0,T0,d,Td,d.V=\T_0,0,T_d,0,T_0,d,T_d,d\. (28) We set d=7d=7 pixels and construct each view by reflection-padding and cropping to the original size. The four views form one batched vision-encoder input. Score alignment and consensus. For view v∈v , let (v)∈ℝNs^(v) ^N be its encoder-side priority-score vector. Operator vA_v inverse-aligns the score grid to T0,0T_0,0 by bilinear sampling with reflection padding. With patch size p=14p=14, the offset is d/p=0.5d/p=0.5 patch: ~(v)=v((v)). s^(v)=A_v\! (s^(v) ). (29) Equation 11 converts the aligned scores to descending rank quantiles, so each view contributes a priority ranking rather than a score scale. Selection interface. At compression budget K, TCS replaces the original priority ranking with the cross-view consensus ranking. The unshifted view supplies token features and the compressor’s key-similarity metric, while translated views contribute aligned priority scores. Token counts, aggregation, and language-model input length remain unchanged. Matched evaluation and cost. We evaluate None and TCS on matched adversarial images, questions, references, eligibility sets, and compression budgets. Full-token inference is unchanged, so Full ASR is shared within each matched pair in Table 4. The four views are processed by the vision encoder in one batch, with no additional language-model inference. F.2 Cross-View Support Mechanism Cross-view support characterizes the contrast between stable clean evidence and view-specific adversarial replacements. At each compression budget K, the canonical clean Top-K set serves as the reference, while CIRA replacements are tokens that enter the canonical adversarial Top-K set from outside this reference. A candidate’s view support is the number of aligned views in which it remains within the Top-K set. Figure 7: Cross-view support distributions of clean Top-K tokens and CIRA replacement tokens across compression budgets. Figure 7 shows that CIRA replacements are predominantly view-specific. As K increases from 32 to 192, the one-view share decreases from 88.0%88.0\% to 63.9%63.9\%, while fewer than 2.5%2.5\% are supported by all four views. Clean Top-K tokens show the opposite pattern: their four-view share increases from 20.0%20.0\% to 44.2%44.2\%. Averaging aligned rank quantiles therefore downweights isolated replacement spikes while favoring evidence supported across translations. F.3 Adaptive Evaluation Standard CIRA is optimized on the unshifted view, with TCS applied only at evaluation. Adaptive CIRA instead optimizes equation 12 over all four public transformations, using the view-specific encoder objectives as differentiable surrogates for rank conversion and Top-K selection. For each view, we cache clean features and ranks and compute GSH and HEP with view-specific hidden-evidence weights over [Kmin,Kmax][K_ ,K_ ]. The averaged gradient updates one shared perturbation using the original ϵ=4/255ε=4/255, step size 1/2551/255, 100 steps, and λ=0.8λ=0.8. Adaptive CIRA retains the same target-encoder access as standard CIRA. Selection and rank response. For each condition, clean Top-K retention measures the fraction of tokens in the clean Top-K set that remain in the adversarial Top-K set under the corresponding ranking rule. Signed normalized rank change is (ria−ric)/(N−1)(r_i^a-r_i^c)/(N-1), where ricr_i^c and riar_i^a denote clean and adversarial ranks under the same ranking rule, and positive values indicate demotion. Retention is averaged per image. Figure 8: Clean Top-K retention and priority-reallocation profiles under CIRA, CIRA + TCS, and Adaptive CIRA + TCS. Across the four compression budgets, Figure 8(a) shows that TCS raises clean Top-K retention under CIRA from 0.30.3–12.3%12.3\% to 62.962.9–70.8%70.8\%. Adaptive CIRA reduces this retention under TCS to 6.56.5–34.7%34.7\%. Figure 8(b) shows the corresponding priority reallocation: TCS attenuates both the demotion of clean high-priority tokens and the promotion of initially low-priority tokens, whereas Adaptive CIRA restores much of this signed reallocation. Together, the cross-view support patterns and the selection responses under TCS show that TCS suppresses view-fragile priority reallocation. Corresponding task-utility results are reported in Appendix G. Appendix G Complementary Task-Utility Results CSFR is the primary clean-conditioned metric for compression-specific failure. We complement it with accuracy-based results that characterize clean utility under compression, post-attack performance, and task-utility recovery under TCS. G.1 Clean Utility under Compression Table 11: Clean full-token and compressed accuracy (%) across model families, datasets, compressors, and family-specific compression budgets. The final value in each compressed cell reports the average across the listed budgets. Dataset Full ACC VisionZip VisPruner PruMerge FastV LLaVA-v1.5-7B (K=192/128/64/32/Avg.K=192/128/64/32/Avg.) POPE 84.6 84.3/83.5/79.6/73.0/80.1 84.1/82.9/80.5/75.3/80.7 75.6/74.1/72.0/70.1/73.0 81.1/80.0/74.7/68.0/76.0 TextVQA 60.3 54.3/51.6/49.5/45.1/50.1 58.4/58.0/56.8/51.9/56.3 51.9/51.6/51.7/49.4/51.2 51.6/49.4/44.7/37.8/45.9 MME 79.2 77.2/76.3/73.5/69.0/74.0 76.9/76.5/74.0/71.7/74.8 72.9/70.9/71.1/70.8/71.4 76.1/73.9/70.8/67.4/72.1 Qwen3-VL-8B-Instruct (K=96/64/32/Avg.K=96/64/32/Avg.) POPE 86.3 86.1/84.1/80.7/83.6 86.6/85.1/82.3/84.7 86.4/86.1/83.8/85.4 85.1/82.4/74.4/80.6 TextVQA 88.6 46.7/41.1/34.1/40.6 45.8/42.6/41.5/43.3 33.9/31.8/31.2/32.3 52.6/40.6/28.1/40.4 MME 90.0 88.2/88.6/83.4/86.7 88.4/88.1/83.5/86.7 86.7/86.7/83.3/85.6 88.3/85.7/79.4/84.5 InternVL3.5-8B (K=128/64/32/Avg.K=128/64/32/Avg.) POPE 82.2 81.4/80.8/77.1/79.8 80.4/80.5/78.3/79.7 79.0/79.0/78.8/78.9 81.0/79.5/74.9/78.5 TextVQA 71.5 64.7/48.6/37.5/50.3 57.7/47.0/40.4/48.4 46.8/37.7/32.8/39.1 68.5/57.6/44.3/56.8 MME 88.6 87.5/83.8/78.2/83.2 84.8/82.2/77.0/81.3 84.0/80.8/78.5/81.1 88.2/85.2/77.8/83.7 Table 11 shows that high-budget settings preserve most POPE and MME accuracy across model families, while TextVQA generally exhibits larger losses under compression. Related analyses also find task-dependent visual-token requirements, with OCR tasks relying on visual information deeper into the decoder (Wang et al., 2026b). These values provide the clean reference for the post-attack comparisons below. G.2 Post-Attack Task Utility Unlike CSFR, adversarial accuracy is computed over the complete evaluation set and therefore reflects both pre-existing compression errors and attack-induced failures. It provides a complementary view of overall task degradation under CIRA. Table 12: Full-token clean-to-adversarial accuracy and compressed adversarial accuracy (%) under CIRA across model families, datasets, compressors, and family-specific budgets. The final value in each compressed cell reports the budget average. Dataset Full ACC VisionZip VisPruner PruMerge FastV LLaVA-v1.5-7B (compressed Adv. ACC: K=192/128/64/32/Avg.K=192/128/64/32/Avg.) POPE 84.6→82.684.6→ 82.6 78.4/74.0/64.6/58.0/68.878.4/74.0/64.6/58.0/68.8 80.0/77.7/71.7/64.1/73.480.0/77.7/71.7/64.1/73.4 63.3/55.9/51.2/50.3/55.263.3/55.9/51.2/50.3/55.2 76.8/71.4/61.0/53.4/65.776.8/71.4/61.0/53.4/65.7 TextVQA 60.3→57.160.3→ 57.1 46.7/40.0/32.2/26.1/36.346.7/40.0/32.2/26.1/36.3 51.3/49.5/44.6/38.6/46.051.3/49.5/44.6/38.6/46.0 36.9/32.2/27.4/25.0/30.436.9/32.2/27.4/25.0/30.4 45.5/41.9/33.5/25.7/36.745.5/41.9/33.5/25.7/36.7 MME 79.2→77.279.2→ 77.2 73.2/70.7/65.5/59.5/67.273.2/70.7/65.5/59.5/67.2 73.2/70.9/64.1/60.5/67.273.2/70.9/64.1/60.5/67.2 62.2/58.9/55.4/51.5/57.062.2/58.9/55.4/51.5/57.0 73.2/71.6/63.7/58.8/66.873.2/71.6/63.7/58.8/66.8 Qwen3-VL-8B-Instruct (compressed Adv. ACC: K=96/64/32/Avg.K=96/64/32/Avg.) POPE 86.3→81.386.3→ 81.3 75.9/74.2/71.2/73.875.9/74.2/71.2/73.8 78.1/76.7/73.5/76.178.1/76.7/73.5/76.1 79.8/79.8/76.0/78.579.8/79.8/76.0/78.5 76.0/73.9/67.7/72.576.0/73.9/67.7/72.5 TextVQA 88.6→75.788.6→ 75.7 21.1/20.5/21.4/21.021.1/20.5/21.4/21.0 31.9/31.9/31.0/31.631.9/31.9/31.0/31.6 22.2/22.4/22.0/22.222.2/22.4/22.0/22.2 30.7/25.6/20.6/25.630.7/25.6/20.6/25.6 MME 90.0→81.590.0→ 81.5 74.9/76.1/69.8/73.674.9/76.1/69.8/73.6 73.9/76.5/72.5/74.373.9/76.5/72.5/74.3 74.3/77.1/74.6/75.374.3/77.1/74.6/75.3 79.9/77.5/70.9/76.179.9/77.5/70.9/76.1 InternVL3.5-8B (compressed Adv. ACC: K=128/64/32/Avg.K=128/64/32/Avg.) POPE 82.2→76.182.2→ 76.1 75.5/75.4/71.8/74.275.5/75.4/71.8/74.2 75.8/75.9/73.4/75.075.8/75.9/73.4/75.0 75.7/75.0/75.4/75.475.7/75.0/75.4/75.4 77.3/73.1/70.2/73.577.3/73.1/70.2/73.5 TextVQA 71.5→65.971.5→ 65.9 42.3/32.3/27.1/33.942.3/32.3/27.1/33.9 53.9/43.6/36.1/44.553.9/43.6/36.1/44.5 39.5/31.8/29.5/33.639.5/31.8/29.5/33.6 58.4/42.7/29.6/43.658.4/42.7/29.6/43.6 MME 88.6→85.388.6→ 85.3 81.4/75.7/71.5/76.281.4/75.7/71.5/76.2 81.3/77.8/74.9/78.081.3/77.8/74.9/78.0 80.2/78.3/74.7/77.780.2/78.3/74.7/77.7 83.8/79.8/72.9/78.883.8/79.8/72.9/78.8 Table 12 shows that CIRA reduces budget-averaged compressed accuracy by 3.33.3–20.820.8 p across the evaluated model–dataset–compressor settings. After averaging compressors within each model–dataset pair and weighting the nine pairs equally, the mean compressed accuracy drop is 9.69.6 p, compared with 5.45.4 p under full-token inference. This aggregate view complements the paired CSFR analysis by quantifying end-task degradation under compressed inference. Table 13: Matched compressed accuracy (%) without and with TCS on LLaVA-v1.5-7B with VisionZip under clean, CIRA, and Adaptive CIRA conditions across datasets and compression budgets. Entries show None→TCSNone\!→\!TCS, with annotations reporting the corresponding change. Evaluation K=192K=192 K=128K=128 K=64K=64 K=32K=32 Avg. POPE Clean →83.1084.30\!→\!83.10 ↓ 1.20 →82.5083.50\!→\!82.50 ↓ 1.00 →78.4079.60\!→\!78.40 ↓ 1.20 →74.3073.00\!→\!74.30 ↑ 1.30 →79.5880.10\!→\!79.58 ↓ 0.52 CIRA →81.3078.40\!→\!81.30 ↑ 2.90 →80.9074.00\!→\!80.90 ↑ 6.90 →78.3064.60\!→\!78.30 ↑ 13.70 →75.4058.00\!→\!75.40 ↑ 17.40 →78.9868.75\!→\!78.98 ↑ 10.23 Adaptive CIRA →78.6078.80\!→\!78.60 ↓ 0.20 →76.0076.40\!→\!76.00 ↓ 0.40 →71.6066.80\!→\!71.60 ↑ 4.80 →62.2056.30\!→\!62.20 ↑ 5.90 →72.1069.58\!→\!72.10 ↑ 2.52 TextVQA Clean →53.2054.30\!→\!53.20 ↓ 1.10 →51.3051.60\!→\!51.30 ↓ 0.30 →49.4049.50\!→\!49.40 ↓ 0.10 →45.4045.10\!→\!45.40 ↑ 0.30 →49.8350.13\!→\!49.83 ↓ 0.30 CIRA →52.4046.70\!→\!52.40 ↑ 5.70 →50.9040.00\!→\!50.90 ↑ 10.90 →49.0032.20\!→\!49.00 ↑ 16.80 →45.5026.10\!→\!45.50 ↑ 19.40 →49.4536.25\!→\!49.45 ↑ 13.20 Adaptive CIRA →49.0048.30\!→\!49.00 ↑ 0.70 →46.9045.10\!→\!46.90 ↑ 1.80 →42.1036.00\!→\!42.10 ↑ 6.10 →33.0026.70\!→\!33.00 ↑ 6.30 →42.7539.03\!→\!42.75 ↑ 3.72 MME Clean →76.5077.20\!→\!76.50 ↓ 0.70 →76.0076.30\!→\!76.00 ↓ 0.30 →74.3073.50\!→\!74.30 ↑ 0.80 →70.6069.00\!→\!70.60 ↑ 1.60 →74.3574.00\!→\!74.35 ↑ 0.35 CIRA →75.0073.20\!→\!75.00 ↑ 1.80 →75.8070.70\!→\!75.80 ↑ 5.10 →74.0065.50\!→\!74.00 ↑ 8.50 →71.2059.50\!→\!71.20 ↑ 11.70 →74.0067.23\!→\!74.00 ↑ 6.77 Adaptive CIRA →73.6074.50\!→\!73.60 ↓ 0.90 →72.0070.90\!→\!72.00 ↑ 1.10 →68.4062.90\!→\!68.40 ↑ 5.50 →62.0058.50\!→\!62.00 ↑ 3.50 →69.0066.70\!→\!69.00 ↑ 2.30 G.3 Task-Utility Recovery under TCS Table 13 reports matched task accuracy without and with TCS. Each comparison fixes the input image, question, and compression budget, with TCS applied only to compressed inference. Across the three benchmarks, TCS changes clean Avg. ACC by at most 0.520.52 p in magnitude, while recovering 6.776.77–13.2013.20 p under CIRA, with larger gains at tighter budgets. Under Adaptive CIRA, TCS recovers less utility, consistent with the attack partially restoring the priority reallocation suppressed by TCS. Appendix H Limitations and Future Work Temporal and contextual dependencies. The present evaluation is restricted to single-image inference. In video, multi-image, and multi-turn settings, evidence retention also depends on temporal redundancy and evolving context (Shen et al., 2024; Li et al., 2026b; Wang et al., 2026c). Run-Length Pruning, for instance, combines temporal redundancy removal with token distillation (Ma et al., 2026). Whether encoder-only priority manipulation remains compression-selective when evidence is distributed across frames or dialogue turns remains an open question. Compression mechanisms beyond token selection. Our diagnostics examine retained-set allocation and representation drift, but do not exhaust the mechanisms underlying compression-induced errors. Adaptive pruning changes allocation across inputs or layers (Ye et al., 2025; Chen et al., 2026; Li et al., 2026a); learned summarization transforms the token representation (Bulat et al., 2026); and recoverable routing permits deferred tokens to re-enter subsequent selection stages (Yang et al., 2026). These mechanisms complicate a description based on a single token ranking. Moreover, spatial disruption (Huang et al., 2026) and positional or attentional distortion (Cho et al., 2026) are not separately identified by our retained-set interventions. Paired CSF evaluation remains applicable, but attributing failures within these compression mechanisms requires additional diagnostics. Beyond task correctness. CSF is defined through task-level correctness rather than response safety or targeted attack success. Accordingly, preserving full-token correctness does not establish safety preservation, and compressed-path errors do not necessarily constitute safety-alignment failures. Targeted manipulation and multimodal jailbreaks (Zhang et al., 2025a; Qi et al., 2024; Shayegani et al., 2024) provide distinct settings for paired evaluation, requiring outcome criteria tailored to the corresponding security objective. Appendix I Qualitative Case Studies Figure 9: Qualitative examples of compression-specific failures on LLaVA-v1.5-7B across visual-token compressors and retention budgets. Figure 10: Qualitative examples of compression-specific failures on Qwen3-VL-8B-Instruct across visual-token compressors and retention budgets. Figure 11: Qualitative examples of compression-specific failures on InternVL3.5-8B across visual-token compressors and retention budgets.

Instant research discovery