With the rapid advancement of retrieval-augmented vision-language models, multimodal medical retrieval-augmented generation (MMed-RAG) systems are increasingly adopted in clinical decision support. These systems enhance medical applications by performing cross-modal retrieval to integrate relevant visual and textual evidence for tasks, e.g., report generation and disease diagnosis. However, their complex architecture also introduces underexplored adversarial vulnerabilities, particularly via visual input perturbations. In this paper, we propose Medusa, a novel framework for crafting cross-modal transferable adversarial attacks on MMed-RAG systems under a black-box setting. Specifically, Medusa formulates the attack as a perturbation optimization problem, leveraging a multi-positive InfoNCE loss (MPIL) to align adversarial visual embeddings with medically plausible but malicious textual targets, thereby hijacking the retrieval process. To enhance transferability, we adopt a surrogate model ensemble and design a dual-loop optimization strategy augmented with invariant risk minimization (IRM). Extensive experiments on two real-world medical tasks, including medical report generation and disease diagnosis, demonstrate that Medusa achieves over 90% average attack success rate across various generation models and retrievers under appropriate parameter configuration, while remaining robust against four mainstream defenses, outperforming state-of-the-art baselines. Our results reveal critical vulnerabilities in the MMed-RAG systems and highlight the necessity of robustness benchmarking in safety-critical medical applications. The code and data are available at https://github.com/yiliucs/MMed-RAG-Attack. © 2026 Owner/Author.

We tackle the problem of quantifying the number of objects by a generative text-to-image model. Rather than retraining such a model for each new image domain of interest, which leads to high computational costs and limited scalability, we are the first to consider this problem from a domain-agnostic perspective. We propose QUOTA, an optimization framework for text-to-image models that enables effective object quantification across unseen domains without retraining. It leverages a dual-loop meta-learning strategy to optimize a domain-invariant prompt. Further, by integrating prompt learning with learnable counting and domain tokens, our method captures stylistic variations and maintains accuracy, even for object classes not encountered during training. For evaluation, we adopt a new benchmark QUANT-Bench specifically designed for object quantification in domain generalization, enabling rigorous assessment of object quantification accuracy and adaptability across unseen domains in text-to-image generation. Extensive experiments demonstrate that QUOTA outperforms conventional models in both object quantification accuracy and semantic consistency, setting a new benchmark for efficient and scalable text-to-image generation for any domain.

Multispectral filter array (MSFA) imaging enables low-cost, real-time, and compact spectral sensing, making it highly suitable for IoT applications such as smartphone imaging, smart agriculture, UAV-based remote sensing, and wearable diagnostics. However, due to the strong coupling among spectral response functions (SRFs), MSFA pattern, and reconstruction algorithms, achieving high-fidelity reconstruction remains a significant challenge. Conventional independent optimization of each module often leads to suboptimal results, highlighting the necessity of a co-optimized design method. To this end, we propose an end-to-end spectral reconstruction framework that jointly optimizes the SRF, MSFA pattern, and reconstruction network. In our method, the SRFs are parameterized as Gaussian functions defined by the center wavelength and full width at half maximum (FWHM). The MSFA pattern is designed using the optimal sphere-packing (OSP) method to enhance sampling uniformity and reduce artifacts. In addition, a dual-stage reconstruction pipeline is introduced: the first stage employs a Res2FFT-based U-Net with multi-scale gated attention (MGA) for spatial demosaicing, and the second stage uses an enhanced deep residual network (EDSR) for spectral super-resolution. All components are jointly trained under a hybrid loss with supervision from both multispectral and hyperspectral reconstruction. Experiments on three benchmark datasets demonstrate that our method achieves superior spatial and spectral fidelity compared to state-of-the-art methods. Furthermore, the model has been successfully deployed on an Nvidia Jetson-based IoT device, achieving real-time inference at 45 ms per frame, validating its potential for real-time spectral imaging in IoT systems.

Compared to conventional RGB images, hyperspectral images offer a more comprehensive range of spectral information, encompassing both visible and infrared bands. This enhanced spectral information facilitates trackers in effectively differentiating the target object from background clutter, realizing more robust recognition. However, hyperspectral video cameras exhibit variability in terms of spectral ranges and band numbers, producing distinct modalities. Developing specialized networks for each modality proves to be inefficient and time-consuming, while networks trained for one modality struggle to perform well on others, generating unsatisfactory outcomes. To confront these challenges encountered in various tracking scenarios, HyperTrack is proposed as a unified object tracking network tailored for hyperspectral videos. The proposed HyperTrack can be employed individually for single object tracking across three different modalities of hyperspectral videos: near infrared (NIR), visible (VIS), and red-to-near infrared (RedNIR). Specifically, since hyperspectral data have multiple bands, a band gate module is introduced into the network to enable it to select bands from hyperspectral images, thereby reducing the dimensionality for hyperspectral images. Furthermore, in order to effectively utilize the variability amongst the three different modal data, a computationally simple band embedding module is introduced to improve the object tracking performance of different modal hyperspectral videos. Additionally, a hybrid attention module is devised to efficiently extract and interact features between the template and search at each stage. As a unified network, HyperTrack achieves state-of-the-art comprehensive results across three different types of hyperspectral videos, particularly excelling with VIS and NIR type data. The code and models are publicly available at https://github.com/supertyd/HyperTrack.

In this paper, we re-introduce the Curse of Depth, a concept that re-introduces, explains, and addresses the recent observation in modern Large Language Models (LLMs) where deeper layers are much less effective than expected. We first confirm the wide existence of this phenomenon across the most popular families of LLMs, such as Llama, Mistral, DeepSeek, and Qwen. Our analysis, theoretically and empirically, identifies that the underlying reason for the ineffectiveness of deep layers in LLMs is the widespread usage of Pre-Layer Normalization (Pre-LN). While Pre-LN stabilizes the training of Transformer LLMs, its output variance exponentially grows with the model depth, which undesirably causes the derivative of the deep Transformer blocks to be an identity matrix, and therefore barely contributes to the training. To resolve this training pitfall, we propose LayerNorm Scaling, which scales the variance of output of the layer normalization inversely by the square root of its depth. This simple modification mitigates the output variance explosion of deeper Transformer layers, improving their contribution. Our experimental results, spanning model sizes from 130M to 7B, demonstrate that \ours significantly enhances LLM pre-training performance compared to Pre-LN. Moreover, this improvement seamlessly carries over to supervised fine-tuning. All these gains can be attributed to the fact that LayerNorm Scaling enables deeper layers to contribute more effectively during training.

Hyperspectral videos contain richer spectral and physical features than RGB videos and thus have greater potential for use in object tracking. The mainstream hyperspectral object tracking approach involves the integration of multiple RGB-based video tracking models. Although ensembles of multiple models can effectively utilize spectral information and improve tracker performance, this approach has high computational complexity, making it difficult to meet the real-time requirements of video object tracking. To bridge the gap, we propose a new hyperspectral object tracking framework (HotMoE) based on Mixture-of-Experts (MoE). HotMoE leverages a divide-and-conquer strategy, where only a subset of expert models is computed for each input, reducing computational complexity while maintaining performance. In this paper, we first design a splitter to group multiple spectral bands into multiple false-color images based on spectral correlations. Then, we design a hyperspectral MoE router that can adaptively learn to aggregate spectral image feature information and route it to suitable experts. Different experts can handle various scenarios, and HotMoE effectively utilizes the capabilities of different experts to obtain better overall performance. Compared with previous state-of-the-art hyperspectral object tracking networks, our model has significantly reduced inference time and performs well, with a processing speed of 43.7 FPS and an AUC of 0.704 with the HOT2022 dataset.

Panoptic Scene Graph (PSG) generation aims to generate a scene graph representing pairwise relationship between objects within an image. Its use of pixel-wise segmentation mask and inclusion of background regions in relationship inference make it quickly become a popular approach. However, it has an intrinsic challenge that the trained relationship predictors are either of low value or of low quality due to the long-tail distribution of typical datasets. Inspired by how humans use prior knowledge to greatly simplify this problem, we bring in two novel designs, using a pre-trained vision-language model to correct the data skewness, and using conditional prior distribution on contexts to further refine the prediction quality. Specifically, the approach named CKT-RCM first exploits relation-associated visual features from the image encoder and constructs a relation classifier by extracting text embeddings for all relationships from the text encoder of the vision-language model. It also utilizes rich relational context from subject-object pairs to facilitate informative relation predictions via a cross-attention mechanism. We conduct comprehensive experiments on the OpenPSG dataset and achieve state-of-the-art performance.

Pre-trained vision-language models like CLIP have remarkably adapted to various downstream tasks. Nonetheless, their performance heavily depends on the specificity of the input text prompts, which requires skillful prompt template engineering. Instead, current approaches to prompt optimization learn the prompts through gradient descent, where the prompts are treated as adjustable parameters. However, these methods tend to lead to overfitting of the base classes seen during training and produce prompts that are no longer understandable by humans. This paper introduces a simple but interpretable prompt optimizer (IPO), that utilizes large language models (LLMs) to generate textual prompts dynamically. We introduce a Prompt Optimization Prompt that not only guides LLMs in creating effective prompts but also stores past prompts with their performance metrics, providing rich in-context information. Additionally, we incorporate a large multimodal model (LMM) to condition on visual content by generating image descriptions, which enhance the interaction between textual and visual modalities. This allows for the creation of dataset-specific prompts that improve generalization performance, while maintaining human comprehension. Extensive testing across 11 datasets reveals that IPO not only improves the accuracy of existing gradient-descent-based prompt learning methods but also considerably enhances the interpretability of the generated prompts. By leveraging the strengths of LLMs, our approach ensures that the prompts remain human-understandable, thereby facilitating better transparency and oversight for vision-language models.

Meta-learning algorithms are able to learn a new task using previously learned knowledge, but they often require a large number of meta-training tasks which may not be readily available. To address this issue, we propose a method for few-shot learning with fewer tasks, which we call MetaModulation. The key idea is to use a neural network to increase the density of the meta-training tasks by modulating batch normalization parameters during meta-training. Additionally, we modify parameters at various neural network levels, rather than just a single layer, to increase task diversity. To account for the uncertainty caused by the reduced number of training tasks, we propose a variational MetaModulation where the modulation parameters are treated as latent variables. We also introduce learning variational feature hierarchies by the variational MetaModulation, which modulates features at all layers and can take into account task uncertainty and generate more diverse tasks. The ablation studies illustrate the advantages of utilizing a learnable task modulation at different levels and demonstrate the benefit of incorporating probabilistic variants in few-task meta-learning. Our MetaModulation and its variational variants consistently outperform state-of-the-art alternatives on four few-task meta-learning benchmarks.

Single-sensor multispectral cameras generally employ a multispectral filter array (MSFA) to rapidly acquire spatial-spectral information. However, MSFA cameras capture only one spectral band’s information at each pixel location. Thus, to fully utilize the spatial-spectral information recorded by MSFA cameras, demosaicing methods are necessary. In this paper, we propose a novel two-stage demosaicing method that combines the traditional method and the deep learning method. We first use a weighted bilinear interpolation convolution filter to preliminarily interpolate the raw MSFA image. And then a multi-scale dense connections large kernel attention method (MDLKA) is proposed to further reconstruct the full-resolution spectral images. MDLKA can effectively increase the image receptive field and capture long-range dependence. Through dense connections, attention maps of different scales are fused to enrich image features. Finally, compared with current advanced demosaicing methods on the dataset provided by NTIRE 2022, we exceed traditional interpolation methods by 11–15 dB in PSNR. Compared with the mosaic convolution-attention network (MCAN) and Res2Unet network based on deep learning methods, we exceed them by 5.4 dB and 1.95 dB in PSNR, respectively.

Spatiotemporal fusion is commonly used in remote sensing to generate fine spatial–temporal resolution images. Among most present spatiotemporal fusion methods, the registration error results in a mismatch in the reflectance and class abundance information within the MODIS pixel to be analyzed during the spatiotemporal fusion. In this article, a new spatiotemporal fusion model, robust flexible spatiotemporal data fusion (RFSDAF), is proposed; in other words, this method is robust to registration errors. The RFSDAF method uses a multiscale fusion strategy that is adaptive to different degrees of coregistration error. It incorporates multiscale information, which extends the analysis of the reflectance and class fraction from per-MODIS pixel to inter-MODIS pixels, and it is robust to coregistration errors. This method does not require high-accuracy preprocession coregistration algorithms and can automatically reduce the effect of the registration error to a great extent. The RFSDAF method is compared with four spatiotemporal image fusion algorithms, and the effectiveness of the method in resolving the registration error is demonstrated using both a simulated dataset and two actual satellite datasets. The experimental results show that the RFSDAF method can better reduce the impact of the registration error than the spatial and temporal adaptive reflectance fusion model (STARFM) and FSDAF, which adopts the phase cross-correlation algorithm to preprocess and coregister the MODIS–Landsat image through real image experiments. Consequently, the RFSDAF method has a great robustness for real spatiotemporal image fusion with registration errors and has the potential for monitoring land surface dynamics.