Large language models (LLMs) excel in diverse applications but face dual challenges: generating harmful content under jailbreak attacks and over-refusing benign queries due to rigid safety mechanisms.These issues severely affect the application of LLMs.Existing approaches can be divided into three types: contrastive decoding, activation manipulation, and prompting strategies.However, all these approaches face challenges like inefficiency, fragility, or architectural constraints, ultimately failing to strike a balance between safety and usability.These problems are more obvious in multimodal large language models (MLLMs), especially in terms of heightened over-refusal in cross-modal tasks and new security risks arising from expanded attack surfaces.We propose Magic Image 1 , an optimization-driven visual prompt framework that enhances security and reduces over-refusal at the same time.The Magic Image is optimized using gradients derived from harmful/benign training samples.Using the magic image can modify the model's original safety alignment, maintaining robust safety while reducing unnecessary denials.Experiments demonstrate its effectiveness in preserving model performance and improving safety-responsiveness balance across datasets, including unseen data, offering a practical solution for reliable MLLM deployment.

ABSTRACT Camouflaged object detection (COD) aims to identify and segment objects that closely resemble and are seamlessly integrated into their surrounding environments, making it a challenging task in computer vision. COD is constrained by the limited availability of training data and annotated samples, and most carefully designed COD models exhibit diminished performance under low‐data conditions. In recent years, there has been increasing interest in leveraging foundation models, which have demonstrated robust general capabilities and superior generalisation performance, to address COD challenges. This work proposes a knowledge‐guided domain adaptation (KGDA) approach to tackle the data scarcity problem in COD. The method utilises the knowledge descriptions generated by multimodal large language models (MLLMs) for camouflaged images, aiming to enhance the model's comprehension of semantic objects and camouflaged scenes through highly abstract and generalised knowledge representations. To resolve ambiguities and errors in the generated text descriptions, a multi‐level knowledge aggregation (MLKG) module is devised. This module consolidates consistent semantic knowledge and forms multi‐level semantic knowledge features. To incorporate semantic knowledge into the visual foundation model, the authors introduce a knowledge‐guided semantic enhancement adaptor (KSEA) that integrates the semantic knowledge of camouflaged objects while preserving the original knowledge of the foundation model. Extensive experiments demonstrate that our method surpasses 19 state‐of‐the‐art approaches and exhibits strong generalisation capabilities even with limited annotated data.

Current multi-illumination color constancy methods typically estimate illumination for each pixel directly. However, according to the multi-illumination imaging equation, the color of each pixel is determined by various components, including the innate color of the scene content, the colors of multiple illuminations, and the weightings of these illuminations. Failing to distinguish between these components results in color coupling. On the one hand, there is color coupling between illumination and scene content, where estimations are easily misled by the colors of the content, and the distribution of the estimated illuminations is relatively scattered. On the other hand, there is color coupling between illuminations, where estimations are susceptible to interference from high-frequency and heterogeneous illumination colors, and the local contrast is low. To address color coupling, we propose a Color Decoupling Network (CDNet) that includes a Content Color Awareness Module (CCAM) and a Contrast HArmonization Module (CHAM). CCAM learns scene content color priors, decoupling the colors of content and illuminations by providing the model with the color features of the content, thereby reducing out-of-gamut estimations and enhancing consistency. CHAM constrains feature representation, decoupling illuminants by mutual calibration between adjacent features. CHAM utilizes spatial correlation to make the model more sensitive to the relationships between neighboring features and utilizes illumination disparity degree to guide feature classification. By enhancing the uniqueness of homogeneous illumination features and the distinctiveness of heterogeneous illumination features, CHAM improves local edge contrast. Additionally, by allocating fine-grained margin coefficients to emphasize the soft distinctiveness of similar illumination features, further enhancing local contrast. Extensive experiments on single- and multi-illumination benchmark datasets demonstrate that the proposed method achieves superior performance.

Semi-supervised semantic segmentation has gained considerable attention due to its ability to leverage large amounts of unlabeled data to enhance model generalization. Although increasing the amount of unlabeled data ideally enhances performance, reality contradicts this notion. Research indicates that the reasons for performance degradation span the entire semi-supervised learning process. These include the inconsistent distribution of unlabeled and labeled data at the data level, the susceptibility to interference from unknown categories at the model level, and the semantic drift problem caused by the training strategy. This article refers to these challenges as the “full-cycle” challenge. To address these issues, we propose a full-cycle framework. At the data level, we develop a consistency-based method, beginning with unlabeled data screening, to prioritize unlabeled data that closely aligns with the labeled data distribution in training. At the model level, we introduce a consistency-based semi-supervised semantic segmentation model capable of identifying unknown category regions and distancing their features from known category features, reducing the interference of unknown classes. At the training strategy level, we adopt a progressive approach that gradually incorporates data from simpler to more complex cases. This strategy improves the model’s adaptability to potential data changes, ensures stability, and improves overall generalization performance. Extensive experiments demonstrate that our method significantly outperforms existing state-of-the-art approaches, validating its effectiveness in addressing the full-cycle challenge.

The clustering assumption is widely adopted in semi-supervised semantic segmentation methods. However, this assumption heavily relies on high-quality feature representation, leading to learning and cognitive biases if it does not hold. Learning bias entails the potential overfitting of labeled data, leading to capturing local features and consequently misclassifying semantic categories. Cognitive bias indicates the network’s susceptibility to interference features. Recently, a consistency-based mechanism has been proposed to address these biases. By subjecting unlabeled data to diverse weak perturbations, it breaks original clustering features, compelling the model to learn more robust and generalized representations. However, when processing complex remote sensing images, these weak perturbations often prove ineffective in the later stages of training. To address this, we propose a bias correction framework (BCF). The BCF begins with a feature consistency enhancement module (CEM) that guides the student model to learn feature representations with greater generalization capability. In the feature decoding stage, we introduce a multidecoder structure with weakly orthogonal weights to maximize feature differences, thereby further reducing learning and cognitive biases. In addition, to improve the confidence of pseudolabels and enhance consistency learning, we design a multidecoder teacher model based on symmetrical knowledge transfer, allowing the diverse and multiangle information learned by the student model to be transferred to the teacher model. Extensive experimental results show that our method significantly outperforms state-of-the-art methods on the ISPRS Potsdam and Vaihingen datasets.

Semantic segmentation of high-resolution remote-sensing images (HR-RSIs) focuses on classifying each pixel of input images. Recent methods have incorporated a downscaled global image as supplementary input to alleviate global context loss from cropping. Nonetheless, these methods encounter two key challenges: diminished detail in features due to down-sampling of the global auxiliary image, and noise from the same image that reduces the network’s discriminability of useful and useless information. To overcome these challenges, we propose a truncation compensation network (TCNet) for HR-RSI semantic segmentation. TCNet features three pivotal modules: the guidance feature extraction module (GFM), the related-category semantic enhancement module (RSEM), and the global-local contextual cross-fusion module (CFM). GFM focuses on compensating for truncated features in the local image and minimizing noise to emphasize learning of useful information. RSEM enhances discernment of global semantic information by predicting spatial positions of related categories and establishing spatial mappings for each. CFM facilitates local image semantic segmentation with extensive contextual information by transferring information from global to local feature maps. Extensive testing on the ISPRS, BLU, and GID datasets confirms the superior efficiency of TCNet over other approaches.

Color transfer in image processing usually suffers from misleading color mapping and loss of details. This paper presents a novel directive local color transfer method based on dynamic look-up table (D-DLT) to solve these problems in two steps. First, a directive mapping between the source and the reference image is established based on the salient detection and the color clusters to obtain directive color transfer intention. Then, dynamic look-up tables are created according to the color clusters to preserve the details, which can suppress pseudo contours and avoid detail loss. Subjective and objective assessments are presented to verify the feasibility and the availability of the proposed approach. Experimental results demonstrate that our proposed method has better performance on natural color images than classical color transfer algorithms. Furthermore, the reference image can be extended to color blocks instead of images.

Computing a maximum (acyclic) agreement forest (M(A)AF) of a pair of phylogenetic trees is known to be fixed-parameter tractable; the two main techniques are kernelization and depth-bounded search. In theory, kernelization-based algorithms for this problem are not competitive, but they perform remarkably well in practice. We shed light on why this is the case. Our results show that, probably unsurprisingly, the kernel is often much smaller in practice than the theoretical worst case, but not small enough to fully explain the good performance of these algorithms. The key to performance is cluster partitioning, a technique used in almost all fast M(A)AF algorithms. In theory, cluster partitioning does not help: some instances are highly clusterable, others not at all. However, our experiments show that cluster partitioning leads to substantial performance improvements for kernelization-based M(A)AF algorithms. In contrast, kernelizing the individual clusters before solving them using exponential search yields only very modest performance improvements or even hurts performance; for the vast majority of inputs, kernelization leads to no reduction in the maximal cluster size at all. The choice of the algorithm applied to solve individual clusters also significantly impacts performance, even though our limited experiment to evaluate this produced no clear winner; depth-bounded search, exponential search interleaved with kernelization, and an ILP-based algorithm all achieved competitive performance.