PyTorchKR Papers

최신 논문

총 568,830편
태그·기간 등 상세 필터

카테고리 탐색

arXiv:2611.10578v1cs.HC

EgoVeil: Personalized Privacy Reasoning for Egocentric Video

Recent advances in vision-language models (VLMs) enable AI systems to interpret egocentric video. Yet continuous visual capture raises privacy concerns because sensitive content is pervasive and varies across users and contexts. We present EgoVeil, a personalized privacy reasoning system that determines which visual content to obscure based on individual preferences while retaining context useful for downstream assistance. It converts localized visual regions into structured semantic observations and combines a privacy reasoner with an explicit user-specific memory for personalized privacy reasoning. This design supports personalization and continual adaptation as preference evidence accumulates. We evaluate on in-the-wild data from 10 smart-glasses users under a cross-user chronological protocol. EgoVeil achieves 91.93% balanced accuracy and 84.42% Mask recall, outperforming direct VLM baselines and other controlled privacy reasoning baselines. Our results show that personalized, adaptive privacy reasoning better aligns privacy preservation with individual preferences while retaining the visual context needed for effective smart-glasses assistance.

arXiv:2611.10577v1cs.CL

Global Context Guidance for Diffusion Large Language Models

Diffusion large language models (LLMs) have recently emerged as a promising alternative to autoregressive LLMs, enabling parallel generation through iterative denoising. Recent guidance methods improve diffusion LLMs by introducing a weak model to guide the prediction of the base model during denoising. However, the influence of the weak model during denoising remains less clear, and computing its prediction requires an additional forward pass at each denoising step. We present in this paper an in-depth analysis of how diffusion LLMs use contextual information during denoising. Our analysis reveals that diffusion LLMs rely more heavily on local context while often underutilizing global context. Based on this observation, we introduce Global Context Guidance (GCG), a feature-space guidance method that enhances global-context utilization without constructing an explicit weak model. Extensive experiments on LLaDA-8B, Dream-7B, and LLaDA-1.5 across diverse benchmarks demonstrate that GCG consistently outperforms existing guidance methods, while incurring much lower inference overhead.

arXiv:2611.10575v1cs.CV

ROLECUE: Diagnosing Cross-Lingual Role-Binding Failures in Vision-Language Models

Vision-language models (VLMs) can recognize events in images and parse sentences, but it remains unclear when they bind visual Agent-Patient structure to linguistic semantic roles across languages whose role cues depend on morphology, word order or construction. We introduce ROLECUE, a controlled diagnostic that pairs 192 validated two-person action images with German, Turkish, Japanese, English, Korean and Bangla sentences in canonical and patient-first orders. The evaluation separates image-only role recognition, text-only role parsing, direct image-sentence matching, response-compliance behavior and an external modular composition of the model's own component predictions. Across eight recent open-weight VLMs, ROLECUE reveals a structured spectrum: some models fail visual or linguistic component tests, some largely solve the diagnostic, some collapse to constant semantic assent and several show component-joint gaps in which reliable image-only and text-only role predictions do not survive direct joint inference. The largest non-degenerate gaps reach 44.9 percentage points, with final bootstrap intervals confirming that the effect is not a point-estimate artifact. These results reveal multilingual component-joint performance gaps and acceptance biases hidden by aggregate VLM benchmarks and show why component diagnostics are necessary before interpreting joint multimodal errors.

arXiv:2611.10571v1cs.CV

Understanding the Effect of Visual Attention in CNNs: A Decision-Space Alignment Perspective

Visual Attention enhances convolutional neural networks (CNNs) by modulating intermediate features, yet their effects on downstream decisions remain less understood. We present a decision-space framework that models multiplicative attention as an intermediate feature perturbation and derives a first-order prediction of its effect on the target-versus-competitor classification margin. We further introduce the Decision Alignment Score (DAS) to quantify the directional alignment between the perturbation and the local margin gradient. We evaluate the framework on CIFAR-100 using SimAM and CBAM across five popular CNN backbones, ResNet-18, ResNet-20, ResNet-50, ResNet-56, and MobileNetV2. The analysis examines prediction agreement, DAS margin correlations, and downstream activation-state changes. Across most final-attention configurations, predicted and actual margin changes show good agreement, while DAS exhibits positive correlations with actual margin changes. Single-attention experiments further show that prediction agreement and DAS margin correlations depend on attention placement, with stronger associations at later stages.

arXiv:2611.10570v1stat.ML

Deep backward multistep high-order regression methods for high dimensional nonlinear parabolic partial differential equations

We propose a deep backward multistep high-order regression (MHDBR) method for semilinear parabolic partial differential equations through their backward stochastic differential equation representations. The method combines multistep time discretization with regression targets formed by conditional expectations. Under explicit regularity, starting-value, forward-approximation and network-moment assumptions, we prove an error bound separating time discretization, conditional Monte Carlo sampling and neural-network approximation. The analysis assumes exact population minimization; finite-training-sample generalization and optimization errors are not included. For generators independent of the gradient variable, suitably constrained GroupSort networks and sufficiently large conditional sample sizes give order $J$ in the stated nodal $L^2$ error norm. The required network sizes depend on dimension, so this result does not establish dimension-independent complexity. We present a reproducible time-discretization test satisfying the analytical assumptions and three broader neural-network benchmarks, whose relation to the assumptions is discussed explicitly.

arXiv:2611.10565v1q-bio.NC

Pixels over Symbols: Sensory Realism Improves Behavioral Alignment in Models of Cognition

Computational models of cognition have traditionally relied on abstract symbolic inputs that approximate sensory experiences with simplified vector representations. This simplification reflects historical computational constraints and a long-held assumption that the details of sensory realism are incidental to the study of cognitive processes. Recent advances have enabled the development of sensory-realistic models capable of performing diverse cognitive tasks, yet whether sensory realism meaningfully shapes cognition has not been formally established. We address this gap through a large-scale empirical comparison of abstract and sensory-realistic neural network models trained on a battery of memory-dependent decision-making tasks. Our results show that sensory-realistic models mirror human behavioral response patterns significantly more than their abstract counterparts, though a substantial gap to human internal behavioral consistency remains even for the best tested models. By analyzing the internal dynamics of these systems, we find that the learned recurrent weights are not the main source of this divergence. In fact, frozen abstract-trained weights driven by sensory-realistic inputs recover the human consistency of sensory-realistic models. The resulting dynamics are also within the seed-to-seed variability of sensory-realistic models. Together, our findings argue that across a broad set of visual working-memory tasks, sensory realism accounts for a substantial portion of the model-human behavioral agreement, more than the choice of sequence architecture and comparable to the choice of vision encoder. These results suggest that findings obtained from abstract task models in neuroscience should be more carefully interpreted in light of their potential behavioral limitations.

arXiv:2611.10563v1cs.LG

Confronting Uncertainty in Mechanism-AI Coupled Dynamical Systems via Optimal Transport and Information Geometry: Part I. Distribution Learning

Mechanism-AI coupled dynamical systems are increasingly used to study complex biological systems. However, uncertainty associated with noisy, sparse, and partially observed data limits their reliability. This paper presents a systematic mathematical formulation for treating these uncertainties from the perspective of optimal transport and information geometry. We formulate the uncertainty quantification problem as the minimization of an energy functional on a probability density space equipped with a Riemannian metric, with the loss evaluated through a differential equation, and give a unified description of the gradient flows of the energy functional under specific metrics, such as the Wasserstein and Fisher-Rao metrics. Furthermore, we present corresponding numerical methods for solving these gradient flows using abstract adjoint sensitivity analysis. The formulation connects a range of Bayesian inference methods for dynamical systems; numerical experiments compare posterior approximation and predictive uncertainty under a common computational budget. Moreover, it also provides a geometric perspective on dynamical modeling and the study of heterogeneity evolution.

arXiv:2611.10562v1cs.CV

ZARA: Zero-shot Animatable and Relightable 3D Avatar reconstruction from a single image

Recent large human reconstruction models create remarkable animatable 3D avatars from a single image, but their appearance remains entangled with the illumination and cannot be relighted. In contrast, existing 3D relightable-avatar reconstruction methods typically require person-specific optimization. We present ZARA, a zero-shot feed-forward framework that reconstructs an animatable 3D Gaussian avatar together with its intrinsic attributes from a single image. ZARA couples a pretrained large-scale human reconstruction model with a Material Transformer to inherit geometry priors while learning material-aware Gaussian representations. To preserve localized details, we introduce Gaussian-aligned material feature fusion strategy, which injects multi-scale material features onto individual Gaussians and combines them with geometry tokens. As most human-avatar datasets lack ground-truth material maps, we use pseudo-material labels as conditioning inputs and supervision targets during training. We further introduce MatFace regularization to preserve facial structure and color in albedo prediction under unreliable facial pseudo-material supervision. Extensive experiments demonstrate that ZARA reconstructs animatable and relightable humans within tens of seconds without person-specific optimization or post-processing, outperforming existing baselines in both relighting quality and materials generalization ability. As the ability of recovering the physically-based-rendering attributes of the human, ZARA also enables direct material editing of the avatar's appearance.

arXiv:2611.10560v1cs.CL

Large Language Models without QK-Normalization

Query-Key Normalization (QK-Norm) is widely adopted in modern large language models to control the growth of attention logits and improve training stability. However, query and key norms are not merely sources of instability. The query norm controls the sharpness of the attention distribution, while the key norm determines how strongly a token attracts different queries. These norms therefore carry token-specific information about attention selectivity and token importance. By normalizing both, QK-Norm suppresses this information. This raises a natural question: can we control query-key representations without discarding their sensitivity to magnitude? To address this question, we propose QK-MAP, a Magnitude-Aware Projection for constructing queries and keys. Rather than removing magnitude through normalization, QK-MAP transforms the input representation using multiple learnable frequencies, so that changes in magnitude are encoded as bounded phase patterns. This allows magnitude to be represented through phase and direction rather than vector length alone. Its periodic features have a fixed norm while remaining sensitive to magnitude changes, separating magnitude encoding from norm growth. Multiple frequencies further help distinguish magnitudes and reduce periodic ambiguity. Experiments across model scales from 190M to 7B parameters show that QK-MAP achieves lower training and validation loss than both Standard attention and QK-Norm. Further analyses show that QK-MAP preserves sensitivity to input scale, assigns less attention to the first token, and produces stronger attention contrast for relevant evidence in synthetic long-context tests. These results demonstrate that query-key scale can be effectively controlled without sacrificing token-specific magnitude information.

arXiv:2611.10559v1cs.CV

CompositionEvidence: Controlling Geometric Evidence for Cinematic Composition Reasoning with Frozen Vision-Language Models

Cinematic composition questions depend on in-frame geometry that vision-language models (VLMs) use inconsistently. External perception tools can expose this geometry, but naive prompt injection can overwrite useful visual predictions. On ShotBench composition, pre-question geometry reduces Qwen2.5-VL-7B from 49.5% to 46.3%; corrupted measurements lower it to 33.6%. CompositionEvidence derives nine reproducible features from frozen detection, depth, and line experts and supplies their shared 24-dimensional representation through two interfaces. A training-free channel appends gated neutral evidence after the question, reaching 56.2% (+6.7, exact McNemar p=0.0018) and 48.4% under corrupted evidence. A minimal-training channel maps the same representation to four soft tokens with a 7.37M-parameter projector while leaving the VLM frozen; it reaches 61.0% (+11.5). Across nine VLMs, including five commercial models queried through official provider APIs, zero-shot accuracy ranges from 43.8% to 63.0%, and the text channel changes it by +6.7 to -6.9 points; the highest baseline is among the regressions. The effect of explicit geometry depends on measurement reliability and on how the receiving model integrates it.

arXiv:2611.10553v1cs.LG

Meta-Reinforcement Learning with Zero-Shot Reinforcement Learning

Meta-reinforcement learning (meta-RL) trains agents to adapt to novel tasks from test-time experience, but requires diverse training environments and reward functions. While environments are becoming cheaper to generate, it remains difficult to anticipate every objective an agent may be asked to pursue at test time. Behavior Foundation Models (BFMs) learn policies from reward-free data and support zero-shot RL (ZSRL): adaptation to new objectives at test time using a large reward-labeled dataset. Recent work has shown that BFM task inference can instead be performed online without prior task-specific data. We note that this capability positions BFMs and meta-RL as alternative approaches for online adaptation to new reward functions in a fixed environment. We then ask whether we can extend BFMs to compete with meta-RL more broadly, including in novel environments, and identify four challenges: environment identification, online data collection, test-time exploration, and learning from mixed reward supervision. We study each individually, then propose MetaBFM, a hybrid method that addresses all four. During training, it combines supervised reward-following with unsupervised learning; at test time, it can interpolate between exploration, exploitation, and ZSRL-style reward inference. We demonstrate that MetaBFM can recover the in-distribution performance of supervised meta-RL while generalizing to reward functions outside the training set. Our work suggests that BFMs and meta-RL are complementary research directions whose combined strengths allow adaptive agents to learn more diverse behaviors from a fixed set of supervised objectives.

arXiv:2611.10550v1cs.LG

Dual-Granularity Learning for Regression with Continuous Noisy Labels

Learning with noisy labels has been extensively studied for classification, yet its regression counterpart remains relatively underexplored. Unlike classification, regression involves continuous-valued targets where label noise exhibits heterogeneous magnitudes across instances, which poses conflicting requirements on the model: correcting small-to-moderate deviations requires sensitivity to label values for instance-level refinement, whereas large deviations caused by corrupted or outlying annotations render exact label values unreliable, necessitating reduced reliance on exact labels and instead leveraging ordinal structure among samples. These requirements call for methods that capture both global structure and sample-specific deviations in a unified manner. To this end, we propose GRAIN, a dual-granularity framework for robust noisy regression. At the fine-grained level, we derive an explicit residual learning mechanism from a robust loss formulation, introducing instance-specific auxiliary variables to decouple label noise from predictive signals. At the coarse-grained level, we introduce a binning-based contrastive regularization that enforces an order-preserving feature structure aligned with label proximity, addressing noise beyond the capacity of local residual refinement. The two levels are coupled via a histogram-based bridging mechanism that enables bidirectional information flow: coarse-grained structural guidance establishes a well-ordered feature space that anchors residual learning, while fine-grained residual refinement corrects instance-level deviations that structural regularization alone cannot resolve. Experimental results show that GRAIN consistently improves robustness and generalization under diverse noise conditions.

arXiv:2611.10547v1cs.RO

HSGNav: Hypothesis Scene Graphs for Multi-Scale Lifelong Navigation under Object Relocation

In real homes, assistive robots must repeatedly locate targets from large furniture to tiny tabletop items, as people move objects between household surfaces. Uniformly instantiating detections at different scales as 3D objects can be brittle for small on-surface items; without explicit object-state validation, stale entries may persist after scene changes. Our key insight is to separate persistent 3D furniture entities from revisable surface-object hypotheses, retaining their visual evidence without requiring independent object geometry. We formalize this design as the Hypothesis Scene Graph (HSG), which preserves small-object evidence and supports local revision without rebuilding the stable entity topology. Building on HSG, HSGNav uses this separation for multi-scale lifelong navigation: retrieval selects entity or hypothesis evidence according to target scale, exploration acquires additional surface observations, and verification confirms target presence or invalidates contradicted hypothesis records. Because existing lifelong benchmarks such as GOAT-Bench rarely require mid-episode memory invalidation or multi-scale search, we introduce FAROS-Bench (Find And Re-find Objects across Scales) for mid-episode relocation of surface objects and multi-scale find goals. Experiments show that HSGNav performs competitively on GOAT-Bench while substantially outperforming prior methods on FAROS-Bench.

arXiv:2611.10544v1quant-ph

Compact Quantum Channel Attention for Static ASL Recognition: A Controlled Study with Parameter-Matched Baselines

Hybrid quantum-classical models often report good accuracy on image tasks, but it is rarely clear whether the quantum part is responsible. This paper proposes QCA-DenseNet, a static American Sign Language (ASL) recognition pipeline in which an eight-qubit variational circuit acts as a channel-attention gate on DenseNet121 features. The quantum channel attention (QCA) gate has 17,464 parameters, 7.51 times fewer than a squeeze-and-excitation (SE) block. The pipeline is tested on 1,815 MU HandImages images by holding out one of five signers at a time, and a frozen-backbone ablation uses small parts of the Kaggle ASL Alphabet data. On unseen signers, QCA-DenseNet reaches 94.50% mean accuracy, above SE (93.72%) and within 0.28 points of the best model. A matched classical MLP and a model without attention perform about the same, and no difference is significant. QCA gives a much smaller attention gate without a detectable loss in accuracy, but no quantum advantage.

arXiv:2611.10543v1cs.LG

Intra-Option Fitted Q-Evaluation: Evaluating Hierarchical Policies from Non-Hierarchical Data

Off-policy evaluation (OPE) estimates the expected return of a target policy from previously collected data without additional environment interaction. While OPE methods for flat Markovian policies are well studied, little work has addressed evaluating hierarchical policies in which a high-level policy selects temporally extended options that generate primitive actions until a termination condition is met. The primary existing approach applies per-decision importance sampling at the option level, but requires option-annotated trajectories and exhibits variance that still grows with the horizon. For non-hierarchical policies, fitted Q-evaluation (FQE) typically achieves lower error than importance sampling by leveraging the classic Bellman equation for policy evaluation; however, a naive extension of FQE to hierarchical policies results in a biased policy value estimate. In this paper, we introduce Intra-Option Fitted Q-Evaluation (IO-FQE), which estimates a value function in the augmented state-option space; this approach enables OPE of hierarchical policies even when we lack annotations of what option was run in the data. We develop two continuous-action instantiations of IO-FQE and show on long-horizon hierarchical continuous control tasks (AntMaze navigation and OGBench Puzzle manipulation) that IO-FQE eliminates the bias incurred by a naive application of FQE to hierarchical policies and substantially lowers mean squared-error compared to option-level importance sampling.

arXiv:2611.10542v1cs.CR

When Experience Becomes a Backdoor: Poisoning Self-Evolving LLM Agents via Failure-Aware Task Optimization

Large language model (LLM) agents increasingly improve themselves through self-evolution, distilling execution experience into persistent, reusable skills. However, this persistence creates a new attack surface: malicious rules implanted into skills can compromise subsequent agent behavior. Existing skill backdoor attacks typically assume direct manipulation of the victim's skills, which is impractical when skills are constructed internally by a trusted evolver. In this paper, we study a novel problem of task-level backdoor attacks, where the attacker can only inject poisoned evolution tasks, inducing the agent to implant backdoors into its own skills. Since the attacker controls neither skill induction nor downstream execution, a natural approach is to iteratively optimize poisoned tasks based on attack success. However, our preliminary analysis reveals two key challenges: (i) failures at different stages of skill evolution and downstream execution can produce indistinguishable outcomes, providing little guidance for targeted poisoned-task updates; and (ii) unsuccessful updates may undermine earlier progress, while simply discarding them may lead to repeated ineffective updates. To address these challenges, we propose BadEvolve, a failure-aware task-poisoning framework against self-evolving LLM agents. BadEvolve diagnoses attack failures across four stages, i.e., Induction, Activation, Following, and Effect, and directs updates toward the failing stage. It further employs rollback to preserve earlier progress and failure memory to avoid repeating ineffective updates. Extensive experiments across diverse self-evolution frameworks, benchmarks, and backbone LLMs demonstrate the effectiveness of BadEvolve. Our code is available at https://github.com/ventr1c/BadEvolve.

arXiv:2611.10541v1cs.CL

AFVE-Bench: Measuring Abandoned-Frame Verbalization Errors in Multi-Turn LLM Revision

Multi-turn revision often asks a model to replace an earlier organizing frame while preserving still-valid facts and artifact requirements. Existing instruction-following evaluations mainly assess the latest target and may therefore score an artifact as successful even when it adopts the new frame but exposes the superseded one. We call this failure an abandoned-frame verbalization error (AFVE). AFVE-Bench evaluates silent frame replacement in English and Chinese using 192 matched four-turn dialogues across eight writing domains. Each seed has zero, one, or two pre-target user requests to generate under the source frame. AFVE eligibility depends on the user-request history and applies to the one- and two-request conditions; the zero-request condition serves as a frame-bearing baseline. Eleven model configurations yielded 4,224 expected post-target response units. Two annotators independently scored 4,204 scorable responses for target adoption, source-frame verbalization, and direct usability; model and condition metadata were hidden, and disagreements were adjudicated. AFVE occurs in 25.54% (715/2,799) of history-conditioned responses. Of 2,388 target-passing history-conditioned responses, 472 (19.77%) still expose the abandoned source frame, most often through contrastive framing (423/472). For these standalone revision tasks, target adoption does not ensure source-frame silence.

arXiv:2611.10538v1cs.DS

Close and Sort: Exact Maximum-Likelihood Aggregation of Cyclic Rank-Two Preferences

Pairwise probabilities arise in rank aggregation, recommendation, and human or LLM evaluation, yet converting them into a maximum-likelihood list is a weighted Kemeny problem and is NP-hard in general. Known exact algorithms for rank-two tournaments use only comparison signs and do not optimize this likelihood. We show that the weighted problem is exactly solvable when the skew-symmetric log-odds matrix $W_{ij}=\log(p_{ij}/p_{ji})$ has rank two, including when the induced preferences contain cycles. Writing $W_{ij}=\det(h_i,h_j)$ and adding the closing vector $h_0=-\sum_i h_i$ turns the objective into twice the signed area of a closed polygon. This geometry yields CloseSort, which polar-sorts the $n+1$ edge vectors, cuts after $h_0$, and returns a global maximum-likelihood order in $O(n\log n)$ time from the factors. The same geometry identifies the runner-up and gives a stability certificate beyond exact rank two. When the smallest determinant $\delta$ of two consecutive edges in the closed angular cycle is positive, it is exactly the best-to-runner-up log-likelihood gap. For $W=W_0+E$, the CloseSort order of $W_0$ has regret at most $\sum_{i<j}|E_{ij}|$ on $W$ and remains optimal whenever this sum is below $\delta$, a worst-case tight threshold. In 5,120 determinant-logit simulations, increasing the observations per pair from 4 to 64 raises the fraction on which rank-two projection outperforms the empirical plug-in optimum from 68.6% to 85.7%. On 3,600 Arena-Hard LLM-judge profiles drawn from 63 systems and ten judges, the leading rank-two component captures 89.2% of the log-odds energy. On profiles containing a directed majority triangle, CloseSort reduces normalized regret by 44% relative to Borda and improves the mean normalized objective over each of eight baselines. With factors supplied, it is 8,407 times faster than exact dynamic programming on 18-item rankings.

arXiv:2611.10536v1cs.AI

In Search of Lost Time: Temporal Competence in Large Language Models

Today's large language models (LLMs) increasingly interact with time: AI agents schedule tasks, coordinate with people and other AI systems, and even decide how long to think. Yet LLMs only ever encounter time secondhand---in descriptions rather than experience. Does this hinder their reasoning and decisions that involve time? Our paper traces this question from a model's knowledge to its actions through four successive evaluations. We find that models know the rules of time, which can be read, but struggle to predict durations, which must be lived; while they are highly accurate (92.0-97.5\%) on a collection of questions built from the theorems of various time-related literatures, they underestimate their own completion times by up to 99.5\% and fail to identify which tasks will take them longer to run. And while they are better at estimating durations of human tasks, their duration judgments change based on how an event is worded, magnifying similar human biases by up to 1.3$\times$. Moving from estimation to action, we find that some LLMs disregard time costs completely when answering user queries, ignoring a 1,200-fold increase in wait time entirely. Together, our experiments reveal new challenges posed to LLMs as they navigate a dimension they cannot directly perceive, and highlight how their blind spots can even be simple cognitive abilities that humans take for granted.

arXiv:2611.10528v1cs.CV

PrivMark: Pixel-Space Steering against Privacy Leakage in Vision-Language Models

Vision-language models (VLMs) routinely disclose soft-biometric attributes such as age, gender, and race in their text outputs, exposing demographic information from uploaded images to any downstream API user. Existing privacy defenses do not fit this setting: face de-identification leaves distributed soft-biometric cues intact, classifier-targeted adversarial perturbations do not transfer to open-ended generation, and stronger VLM-specific methods require retraining or runtime activation access unavailable to downstream APIs. We propose PrivMark, a framework that suppresses biometric leakage at the image level. By analyzing VLM activations, we identify two weakly aligned steering vector directions in hidden-state space, each tied to a distinct leakage behavior --- one corresponding to incidental attribute disclosure under open-ended descriptions, the other to compliance with direct biometric probes instead of refusal --- and encode both directions into a single bounded pixel perturbation, so the protected image alone enforces privacy with no invasive access to model internals at deployment. To support evaluation, we assemble BioleakBench, $1{,}600$ VQA pairs covering biometric privacy attributes under both free-form description and direct-probe protocols. Across three white-box source models, PrivMark achieves up to $85.9\%$ biometric suppression while retaining $75\%$ of clean OK-VQA utility, up to $1.4\times$ the privacy-utility tradeoff of the strongest pixel-only baseline. The same one-shot perturbations transfer consistently to seven unseen VLMs including closed-source APIs.