arXiv:2605.25629v2 Announce Type: replace
Abstract: Weak-to-strong (W2S) generalization is a promising framework for scalable oversight, yet existing evaluations often test students under matched train-test distributions. Therefore, we study W2S preference learning under zero-shot distribution shift and find that strong students trained on weak preference labels can appear successful in-distribution while failing to transfer across preference datasets. We provide evidence for a representational failure mode in which weak-supervised fine-tuning can pull the strong model toward source-domain features instead of maintaining broadly transferable preference representations. To mitigate this, we propose Representation Anchoring (Anchor), a simple yet effective regularizer that constrains excessive drift from the pretrained strong model's representation space during fine-tuning, while still allowing task-relevant adaptation. Across preference domains, datasets, and model families, Anchor consistently improves out-of-distribution transfer while maintaining competitive in-distribution performance. Together, our evaluation protocol, transfer-aware metrics, and method expose hidden brittleness in current W2S reward modeling and provide a practical path toward more robust preference transfer.
Science Journals
arXiv:2511.16449v5 Announce Type: replace
Abstract: Vision-Language-Action (VLA) models have shown great potential for embodied AI by integrating visual perception, language understanding, and action execution. In real-time deployment, these models must process continuous visual streams, incurring substantial computational overhead. Visual token pruning -- a mainstream technique for accelerating Vision-Language Models (VLMs) by retaining salient tokens while discarding redundant ones -- offers a natural candidate solution to this challenge. However, directly applying VLM-oriented pruning methods to VLA inference can cause severe degradation in manipulation performance. Our analysis attributes this degradation to a key mismatch: VLA inference exhibits distinct attention patterns between the vision-language prefill stage and the action-decode stage, so pruning based only on context-prefill semantic salience is biased toward semantic cues and may remove action-critical visual tokens. Motivated by this observation, we propose VLA-Pruner, an effective plug-and-play token pruning method grounded in the visual requirements of VLA inference, further exploiting the temporal continuity of robot manipulation. Specifically, VLA-Pruner estimates visual-token importance from both semantic prefilling and temporally smoothed action relevance, and then applies a Combine-then-Filter strategy to retain compact, non-redundant tokens under the compute budget. Experiments show that VLA-Pruner outperforms state-of-the-art approaches across multiple VLA architectures, achieving up to 1.99x speedup with comparable manipulation quality.
arXiv:2511.19741v4 Announce Type: replace
Abstract: Optimal Transport (OT) offers a powerful framework for finding correspondences between distributions and addressing matching and alignment problems in various areas of computer vision, including shape analysis, image generation, and multimodal tasks. The computation cost of OT, however, hinders its scalability. Slice-based transport plans have recently shown promise for reducing the computational cost by leveraging the closed-form solutions of 1D OT problems. These methods optimize a one-dimensional projection (slice) to obtain a conditional transport plan that minimizes the transport cost in the ambient space. While efficient, these methods leave open the question of whether learned optimal slicers can transfer to new distribution pairs under distributional shift. Understanding this transferability is crucial in settings with evolving data or repeated OT computations across closely related distributions. In this paper, we study the min-Sliced Transport Plan (min-STP) framework and investigate the transferability of optimized slicers: can a slicer trained on one distribution pair yield effective transport plans for new, unseen pairs? Theoretically, we show that optimized slicers remain close under slight perturbations of the data distributions, enabling efficient transfer across related tasks. To further improve scalability, we introduce a minibatch formulation of min-STP and provide statistical guarantees on its accuracy. Empirically, we demonstrate that the transferable min-STP achieves strong one-shot matching performance and facilitates amortized training for point cloud alignment and flow-based generative modeling.
arXiv:2605.24456v2 Announce Type: replace
Abstract: Humans constantly reason about 3D proximity, the relations between their body and surrounding objects, to guide perception and action in daily life. Whether multimodal large language models (MLLMs) can perform such embodied 3D reasoning remains unclear. To this end, we introduce EgoProx, a benchmark for egocentric 3D proximity reasoning. We organize our tasks along a cognitive chain, covering intention, exploration, exploitation, and chain-of-actions reasoning. We also design an agent based data engine that produces diverse and consistent QA pairs at scale. We benchmark prevailing MLLMs on EgoProx and conduct additional analyses with dataset specific and task specific instruction tuning. We observe large cross-domain gains, indicating that current MLLMs contain some spatial knowledge; however, they still struggle to effectively leverage it for spatial reasoning VQA.
arXiv:2605.24001v2 Announce Type: replace
Abstract: Recent advances in one-step text-to-image generation have enabled real-time synthesis with remarkable efficiency and quality. Previous reinforcement learning methods for one-step generators combine image-space reward optimization with diffusion noisy-space distribution matching. This paradigm brings challenges due to a mismatch between terminal reward optimization and the underlying generative dynamics. As a result, optimization tends to exploit stochastic degrees of freedom, often improving reward at the expense of image fidelity. To address this issue, we propose Diff-Instruct with Diffused Reward (DIDR), a data-free trajectory-level alignment framework derived from Integral KL minimization. DIDR propagates the RLHF-optimal reward-tilted clean-image distribution across all noise levels along the diffusion trajectory. We show that this objective admits the same minimizer as clean-image RLHF, while naturally inducing the Diffused Reward Score (DRS), which acts as a reward-driven correction to the reference score function. To make this practical, we further introduce the Diffused Reward Proxy (DRP), an efficient estimator of DRS based on differentiable short-step denoising. Extensive experiments demonstrate that DIDR consistently Pareto-dominates existing one-step SDXL baselines. Moreover, when transferred to a 6B DiT backbone (Z-Image), DIDR surpasses its 50-step teacher in preference alignment while requiring only a single generation step.
arXiv:2605.23651v2 Announce Type: replace
Abstract: While factual correctness and task-performance have been in focus of Large Language Model (LLM) research for a long time, the fundamental question of how human-like generated texts are on a linguistic level has been underexplored. From a corpus-linguistic perspective, language production is inherently context-dependent, with distinct communicative contexts giving rise to differences in frequencies and co-occurrence patterns of linguistic features. A text failing to adhere to these patterns can be content-wise correct, but still be unfavorable to human readers. In this work, we propose a context-aware evaluation framework in which human-likeness is assessed using a two-sample problem between the linguistic feature distribution of a human reference corpus for a given register and a corresponding LLM-generated corpus. We implement this framework using the Maximum Mean Discrepancy (MMD) and the 67 lexico-grammatical features introduced by Biber, which are commonly applied in corpus linguistics. In our experiments, we compare seven instruction-tuned, open-source models across five English-language datasets spanning distinct registers against a human baseline. While across all tested setups, LLMs deviate from the human baseline, which models are closest to human language depends on the register and is not dictated by model size.
arXiv:2605.21385v2 Announce Type: replace
Abstract: Many digital systems are designed as collections of asynchronous processes orchestrated by a domain-specific scheduler. The verification of such scheduler-restricted asynchronous systems (SRA) is challenging due to process-process and process-scheduler interactions. In this paper, we tackle the problem of verifying configurable SRA. A configurable SRA describes an unbounded family of possible SRA, each resulting from an instantiation satisfying given configuration constraints; our goal is proving at once that every legal instantiation of a configurable SRA is correct. We propose a contract-based, deductive verification approach that combines (i) compositional proof rules that abstract the scheduler to prove top-level invariant properties, (ii) automatic summarizations of the methods invoked by the scheduler, (iii) simplification with respect to the nature of the space of configurations. The approach is grounded in (object-oriented) first order logic, requires reasoning over quantified statements, and leverages the Dafny software verifier as a backend. An experimental evaluation on industrial case studies demonstrates that the framework scales effectively and enables practical reasoning about complex parameterized behaviors.
arXiv:2605.14799v2 Announce Type: replace
Abstract: In recent years, computer vision has witnessed remarkable progress, fueled by the development of innovative architectures such as Convolutional Neural Networks (CNNs), Generative Adversarial Networks (GANs), diffusion-based architectures, Vision Transformers (ViTs), and, more recently, Vision-Language Models (VLMs). This progress has undeniably contributed to creating increasingly realistic and diverse visual content. However, such advancements in image generation also raise concerns about potential misuse in areas such as misinformation, identity theft, and threats to privacy and security. In parallel, Mamba-based architectures have emerged as versatile tools for a range of image analysis tasks, including classification, segmentation, medical imaging, object detection, and image restoration, in this rapidly evolving field. However, their potential for identifying AI-generated images remains relatively unexplored compared to established techniques. This study provides a systematic evaluation and comparative analysis of Vision Mamba models for AI-generated image detection. We benchmark multiple Vision Mamba variants against representative CNNs, ViTs, and VLM-based detectors across diverse datasets and synthetic image sources, focusing on key metrics such as accuracy, efficiency, and generalizability across diverse image types and generative models. Through this comprehensive analysis, we aim to elucidate Vision Mamba's strengths and limitations relative to established methodologies in terms of applicability, accuracy, and efficiency in detecting AI-generated images. Overall, our findings highlight both the promise and current limitations of Vision Mamba as a component in systems designed to distinguish authentic from AI-generated visual content. This research is crucial for enhancing detection in an age where distinguishing between real and AI-generated content is a major challenge.
arXiv:2512.17366v2 Announce Type: replace
Abstract: Scanning transmission X-ray microscopy (STXM) is a nanoscale imaging technique that can utilize several powerful contrast mechanisms for the quantitative mapping of chemical and physical materials properties. Spatial resolutions down to 7~nm at the soft X-ray energy range have been demonstrated. A limiting factor for high-resolution STXM imaging is given by the positioning precision of the sample with respect to the focusing optic, with the current state-of-the-art leading to significant overheads, especially at low pixel dwell times, and being vulnerable to unavoidable external vibrations sources. In this work, we present a method, called supersampled scanning microscopy, that allows for a significant reduction of overhead times while simultaneously removing the effects of vibrational noise by sampling the position of the sample at a rate significantly higher than the vibration spectrum and reconstructing the sample transmission image from the recorded list of positions and detector counts. We demonstrate the performance of the technique with a set of proof-of-concept high-resolution imaging experiments.
arXiv:2512.04472v2 Announce Type: replace
Abstract: The aggregate flexibility region of distributed energy resources (DERs) quantifies the aggregate power shaping capabilities of DERs. It characterizes the distribution network's potential for wholesale market participation and grid service provision at the transmission level. To enhance flexibility and fully exploit the potential of DERs, this paper proposes a method to optimize the aggregate flexibility region through distribution network reconfiguration. First, we formulate the ellipsoidal aggregate flexibility region characterization problem as a two-stage adaptive robust optimization problem and derive an exact convex reformulation with a large number of second-order cone constraints. By exploiting the problem structure, we propose a scalable Benders decomposition algorithm with provable finite convergence to the optimal solution. Finally, we propose an optimal reconfiguration problem for aggregate flexibility region optimization and solve it using the custom Benders decomposition. Numerical simulations on the IEEE 123-bus test feeder demonstrate that, compared to existing approaches, substantial improvements in the aggregate flexibility region can be achieved over multiple scenarios with the optimized topology.
arXiv:2511.22882v3 Announce Type: replace
Abstract: We introduce boundary quotients and present a framework for learning densities on manifolds that arise as boundary quotients of simpler domains. We show that this framework can be used to construct normalizing flows on quotient manifolds $N/G$, where a discrete group $G$ acts on $N$. We instantiate this construction for genus-$g$ surfaces $\Sigma_g$. When $G$ is finite, we show applicability to symmetry aware learning; we demonstrate this on cyclic quotients of the 3-sphere. Experiments on lens spaces show that simple pre-quotient RealNVP models can achieve strong results while being substantially cheaper to evaluate.
arXiv:2512.04085v2 Announce Type: replace
Abstract: We introduce the "single-life" learning paradigm, where we train a distinct vision model exclusively on egocentric videos captured by one individual. We leverage the multiple viewpoints naturally captured within a single life to learn a visual encoder in a self-supervised manner. Our experiments demonstrate three key findings. First, models trained independently on different lives develop a highly aligned geometric understanding. We demonstrate this by training visual encoders on distinct datasets each capturing a different life, both indoors and outdoors, as well as introducing a novel cross-attention-based metric to quantify the functional alignment of the internal representations developed by different models. Second, we show that single-life models learn generalizable geometric representations that effectively transfer to downstream tasks, such as depth estimation, in unseen environments. Third, we demonstrate that training on up to 30 hours from one week of the same person's life leads to comparable performance to training on 30 hours of diverse web data, highlighting the strength of single-life representation learning. Overall, our results establish that the shared structure of the world, both leads to consistency in models trained on individual lives, and provides a powerful signal for visual representation learning.
arXiv:2601.05028v2 Announce Type: replace
Abstract: Equivariance is a powerful inductive bias in neural networks, improving generalisation and physical consistency. Recently, however, non-equivariant models have regained attention, due to their better runtime performance and imperfect symmetries that might arise in real-world applications. This has motivated the development of approximately equivariant models that strike a middle ground between respecting symmetries and fitting the data distribution. Existing approaches in this field usually apply sample-based regularisers which depend on data augmentation at training time, incurring a high sample complexity, in particular for continuous groups such as $SO(3)$. This work instead approaches approximate equivariance via a projection-based regulariser which leverages the orthogonal decomposition of linear layers into equivariant and non-equivariant components. In contrast to existing methods, this penalises non-equivariance at an operator level across the full group orbit, rather than point-wise. We present a mathematical framework for computing the non-equivariance penalty exactly and efficiently in both the spatial and spectral domain. In our experiments, our method consistently outperforms prior approximate equivariance approaches in both model performance and efficiency, achieving substantial runtime gains over sample-based regularisers.
arXiv:2511.17852v2 Announce Type: replace
Abstract: Transformers can acquire Chain-of-Thought (CoT) capabilities to solve complex reasoning tasks through fine-tuning. Reinforcement learning (RL) and supervised fine-tuning (SFT) are two primary approaches to this end. In this work, we specifically examine RL with process rewards and SFT for learning $k$-sparse Boolean functions with a one-layer transformer through intermediate reasoning steps akin to CoT. In particular, we consider $k$-sparse Boolean functions that can be recursively decomposed into fixed 2-sparse Boolean functions. We first analyze the learning dynamics of RL fine-tuning with process reward and SFT in a unified way. This allows us to identify sufficient conditions under which the transformer provably learns these sparse Boolean functions. We then verify that these conditions hold for three basic examples, including $k$-PARITY, $k$-AND, and $k$-OR, thus demonstrating their learnability via both RL and SFT. Notably, we reveal that RL and SFT exhibit distinct learning behaviors: RL learns the whole CoT chain simultaneously, whereas SFT naturally learns the CoT chain step by step. Overall, our findings provide insights on the mechanisms underlying RL and SFT and how they differ in triggering the CoT capabilities of transformers, and suggest that the comparison between RL and SFT may need to consider the reward design and the use of teacher forcing.
arXiv:2601.14702v2 Announce Type: replace
Abstract: Autonomous driving requires reliable perception and safe decision-making in complex scenarios. Recent vision-language models (VLMs) demonstrate reasoning and generalization abilities, opening new possibilities for autonomous driving; however, existing benchmarks often evaluate perception and decision-making separately, limit failure analysis with choice-only formats, or introduce evaluation bias through LLM-scored long-form outputs. To address these issues, we present Drive-P2D, a progressive perception-to-decision benchmark with 6,650 questions across Object, Scene, and Decision levels. Drive-P2D adopts a separated reasoning-and-answer protocol: final answers are scored objectively, while reasoning is analyzed to identify error modes exposed along the progressive perception-to-decision chain. We evaluate mainstream VLMs across all and high-risk scenarios, and further characterize the perception-to-decision capability boundary through correlation analysis and similar-scene robustness testing. Reasoning further exposes failure modes such as logical reasoning errors and semantic feature omissions, and we train a lightweight analyzer model to automate large-scale error-mode annotation of reasoning. Together, these designs provide practical insights for building safer and more reliable VLMs for real-world autonomous driving.
arXiv:2601.15925v2 Announce Type: replace
Abstract: Optical absorption is a primary, label-defining contrast across biology, pathology, and materials science, yet three-dimensional quantitative absorption imaging has remained largely inaccessible to the brightfield microscopes used in everyday practice. We introduce quantitative absorption tomography (QAT), which recovers volumetric distributions of the extinction coefficient by treating brightfield image formation as a linear inverse problem in logarithmic intensity space and inverting a three-dimensional absorption optical transfer function. Under weak-scattering conditions, QAT yields spectrally resolved, three-dimensional absorption maps from through-focus image stacks acquired on standard brightfield platforms, without interferometry, coherent illumination, or sample rotation. We use QAT to track melanin dynamics in living melanoma cells without exogenous labels, image pigment organization in intact Petunia hybrida petals in vivo, and reconstruct chromogenic contrast across large H&E-stained human tissue volumes. By establishing absorption as a directly measurable volumetric quantity within standard brightfield workflows, QAT positions chromogenic contrast as a quantitative axis alongside fluorescence- and refractive-index-based imaging.
arXiv:2511.16870v3 Announce Type: replace
Abstract: Enforcing alignment between the internal representations of diffusion or flow-based generative models and those of pretrained self-supervised encoders has recently been shown to provide a powerful inductive bias, improving both convergence and sample quality. In this work, we extend this idea to inverse problems, where pretrained generative models are employed as priors. We propose applying representation alignment (REPA) between diffusion or flow-based models and a DINOv2 visual encoder, to guide the reconstruction process at inference time. Although ground-truth signals are unavailable in inverse problems, we empirically show that aligning model representations of approximate target features can substantially enhance reconstruction quality and perceptual realism. We provide theoretical results showing (a) that REPA regularization can be viewed as a variational approach for minimizing a divergence measure in the DINOv2 embedding space, and (b) how under certain regularity assumptions REPA updates steer the latent diffusion states toward those of the clean image. These results offer insights into the role of REPA in improving perceptual fidelity. Finally, we demonstrate the generality of our approach by We integrate REPA into multiple state-of-the-art inverse problem solvers, and provide extensive experiments on super-resolution, box inpainting, Gaussian deblurring, and motion deblurring confirming that our method consistently improves reconstruction quality, while also providing efficiency gains reducing the number of required discretization steps.
From Per-Image Low-Rank to Encoding Mismatch: Rethinking Feature Distillation in Vision Transformers
arXiv:2511.15572v3 Announce Type: replace
Abstract: Feature-map knowledge distillation (KD) transfers internal representations well between comparably sized Vision Transformers (ViTs), but it often fails in compression. We revisit this failure and uncover a paradox. Sample-wise SVD shows that each image is highly compressible, which seems to suggest that a narrow student with a linear projector should match the teacher "in principle". However, a dataset-level view contradicts this intuition: PCA shows that the teacher is a union of low-rank subspaces with significant subspace rotation across inputs. We further introduce token-level Spectral Energy Patterns (SEP) and find an architecture-invariant encoding law: tokens spread energy broadly across channel modes even when they live in low-rank subspace, creating a bandwidth mismatch. We refer to this combined phenomenon as an encoding mismatch. We propose two minimal remedies, Lift or WideLast: (i) Lift retains a lightweight lifting projector at inference to provide wider channel, or (ii) WideLast widens only the student's last block, enabling an input-dependent expansion. On ImageNet-1K, these fixes revive feature KD for ViT compression, improving DeiT-Tiny distilled from CaiT-S24 from 74.86% to 77.53%/78.23% top-1 accuracy, and they also strengthen students trained without distillation. Our analyses clarify when and why feature-map KD fails and then how to fix it. Code and raw data are provided in https://github.com/thy960112/From-Per-Image-Low-Rank-to-Encoding-Mismatch.
arXiv:2510.14542v2 Announce Type: replace
Abstract: We study deep state-space models (Deep SSMs) that contain linear quadratic-output (LQO) systems as internal blocks and present a compression method with a provable output error guarantee. We first derive an upper bound on the output error between two Deep SSMs and show that the bound can be expressed in terms of the $h^2$-error norms between the layerwise LQO systems. In particular, we show that reducing the $h^2$ approximation errors of the LQO systems placed in shallow layers is effective in reducing the derived upper bound on the output error. Next, we formulate an optimization problem for the derived upper bound and develop a gradient-based MOR method. In the numerical experiments, using the IMDb task from the LRA benchmark, we demonstrate the effectiveness of the proposed upper-bound-based compression method. In particular, we show that the number of trainable parameters can be reduced by approximately 60\% without retraining while maintaining the performance of the original model.
arXiv:2510.07478v2 Announce Type: replace
Abstract: We study group fairness in the context of feedback loops induced by meritocratic selection into programs that themselves confer additional advantage, like college admissions. We introduce a stylized, yet novel inter-generational model for the setting and analyze it in situations where there are no underlying differences between two populations. When the benefit of the program (or the harm of not getting into it) is completely symmetric, we show that disparities between the two populations will vanish on average in the long term, although in the short term disparities will continue to arise and dissipate cyclically. Further, the time an accumulated advantage takes to dissipate can be significant, and increases as a function of the relative importance of the program in conveying benefits. Interestingly, significant disparities can arise purely due to randomness even from completely symmetric initial conditions, especially when populations are small. The introduction of even a slight asymmetry, where the group that has accumulated an advantage becomes slightly preferred, leads to a completely different outcome. In these instances, starting from completely symmetric initial conditions, disparities between groups arise stochastically and then persist over time, yielding a permanent advantage for one group. Our analysis precisely characterizes conditions under which disparities persist or diminish, with a particular focus on the role of the scarcity of available spots in the program and its effectiveness. We also present extensive simulations in a richer model that further support our theoretical results in the simpler, stylized model. Our findings are relevant for the design and implementation of algorithmic fairness interventions in similar selection processes.
arXiv:2605.26634v1 Announce Type: new
Abstract: Practical ISAC is constrained by static clutter and NLoS multipath, which obscure target-coupled echoes and induce spurious peaks for beam alignment. Existing receiver-side methods largely model targets as passive scatterers, limiting the structural separability of target echoes from the environment. This paper establishes a structural correspondence between these limitations and target-side Backscatter-MIMO responses: reflection modulation enables waveform-domain separation from unmodulated clutter, while retro-directional passive beamforming concentrates the tagged echo toward the BS-facing direction and suppresses NLoS-induced false-peak locking. To operationalize this correspondence, dual-end spatial locking is required to overcome cascaded backscatter loss and provide beam-domain angular information. We propose a downlink-triggered blind dual-end alignment protocol that jointly selects the BS and Backscatter-MIMO codeword indices from the tagged echo observed at the BS, without pilots, CSI feedback, or target synchronization. We further derive a clutter-aware remodulation waveform robust to fractional timing offsets and construct adjustable-width BS/Backscatter-MIMO codebooks via quadratic phase spoiling. For reliability characterization, we derive closed-form expressions for the coherence-averaged end-to-end success probability. The analysis shows that beam narrowing is not universally beneficial: in NLoS-dominated regimes, enlarging the array aperture may degrade alignment reliability. The optimal beamwidth is instead governed by cross-phase competition between discovery and alignment, yielding a nontrivial feasible region with an analytically characterized boundary. Simulations validate the analysis and demonstrate improved reliability-gated locked-link performance under strong clutter, severe NLoS multipath, and finite coherence time.
arXiv:2605.26635v1 Announce Type: new
Abstract: Stream-based monitoring is a runtime verification approach where a monitor aggregates streams of input data from sensors and other sources to give real-time statistics and assessments of a system's health. One of the central challenges in designing reliable stream-based monitors is to deal with the asynchronous nature of data streams: in concrete applications, the different sensors being monitored produce values at different speeds, and it is the monitor's responsibility to correctly react to the asynchronous arrival of different streams of values. To ease this process, modern frameworks for stream-based monitoring such as RTLola enable users to finely specify data synchronization policies via a system of pacing annotations. While this feature simplifies the design of monitors, it can also lead users to write inconsistent policies, where synchronization between two streams is explicitly requested via annotations, but cannot always be achieved. To mitigate this issue, this paper presents pacing types, a novel type system implemented in RTLola to ensure that monitors for asynchronous streams are free of timing inconsistencies. We give a formal semantics to pacing annotations for a core fragment of RTLola, and present a soundness proof of the pacing type system. For an additional level of guarantees, we machine-checked the soundness proof using the Rocq proof assistant.
arXiv:2605.26285v1 Announce Type: new
Abstract: This work addresses the problem of learning the dynamics of high-dimensional probability densities over time using unlabeled samples, without assuming access to trajectory information. We introduce two-parameter flows that learn only sampling-time transports from a base distribution to each marginal and then extract a physics-time velocity by regressing on coupled synthetic trajectories. We prove that the resulting physics-time dynamics are unique and inherit regularity from the sampling-time transports. Because we can build on standard, well-developed conditional flow matching techniques for learning the base-to-marginal transports, our approach scales to high dimensions and avoids per-step optimal-transport couplings, while allowing admissible non-gradient dynamics that can naturally explain rotational or circulating physics phenomena.
arXiv:2605.26287v1 Announce Type: new
Abstract: Masked autoencoders (MAE) have shown great promise in medical image classification. However, the random masking strategy employed by traditional MAEs may overlook critical areas in medical images, where even subtle changes can indicate disease. To address this limitation, we propose a novel approach that utilizes a multifractal measure (Renyi entropy) to optimize the masking strategy. Our method, termed Multifractal-Optimized Masked Autoencoder (MO-MAE), employs a multifractal analysis to identify regions of high complexity and information content. By focusing the masking process on these areas, MO-MAE ensures that the model learns to reconstruct the most diagnostically relevant features. This approach is particularly beneficial for medical imaging, where fine-grained inspection of tissue structures is crucial for accurate diagnosis. We evaluate MO-MAE on several medical datasets covering various diseases, including MedMNIST and COVID-CT. Our results demonstrate that MO-MAE achieves promising performance, surpassing other basiline and state-of-the-art models. The proposed method also adds minimum computational overhead as the computation of the proposed measure is straightforward. Our findings suggest that the multifractal-optimized masking strategy enhances the model's ability to capture and reconstruct complex tissue structures, leading to more accurate and efficient medical image representation. The proposed MO-MAE framework offers a promising direction for improving the accuracy and efficiency of deep learning models in medical image analysis, potentially advancing the field of computer-aided diagnosis.
arXiv:2605.26636v1 Announce Type: new
Abstract: We introduce JetViT, a novel family of hybrid-architecture Vision Transformer (ViT) models that match the accuracy of state-of-the-art full-attention vision foundation models while achieving substantially higher inference efficiency on high-resolution images. At the core of our approach is Post-Training Attention Search, a post-training acceleration framework that converts pre-trained full-attention ViTs into efficient hybrid-attention variants by identifying and replacing redundant full-attention blocks with linear or window-attention blocks. By inheriting the MLP and attention weights from the base model, Post-Training Attention Search efficiently explores the architectural design space through three key steps: (1) optimizing the linear-attention block design; (2) finding the best combination of linear-attention and window-attention blocks; and (3) identifying and preserving critical full-attention blocks. We evaluate JetViT on two representative high-resolution vision foundation models, DINOv3 and DepthAnythingV2. On the NVIDIA H100 GPU, JetViT achieves up to 1.79x higher throughput and up to 44.81% lower latency without sacrificing accuracy. We will release our code and accelerated ViT models soon.