Fine-tuning large language models (LLMs) improves performance but introduces critical safety vulnerabilities: even minimal harmful data can severely compromise safety measures. We observe that perturbations orthogonal to the alignment direction—defined by weight differences between aligned (safe) and unaligned models—rapidly compromise model safety. In contrast, updates along the alignment direction largely preserve it, revealing the parameter space as a "narrow safety basin". To address this, we propose AsFT (Anchoring Safety in Fine-Tuning) to maintain safety by explicitly constraining update directions during fine-tuning. By penalizing updates orthogonal to the alignment direction, AsFT effectively constrains the model within the "narrow safety basin," thus preserving its inherent safety. Extensive experiments on multiple datasets and models show that AsFT reduces harmful behaviors by up to 7.60%, improves task performance by 3.44%, and consistently outperforms existing methods across multiple tasks.

ABSTRACT Multivariate time series forecasting (MTSF) involves predicting future values of multiple interrelated variables based on historical observations. While existing models often struggle to capture complex temporal multi‐scale dependencies and simultaneously modelling intraperiod and interperiod variations, thereby limiting their predictive accuracy. To address these limitations, we introduce a novel forecasting model named MTSPnet. This model employs a two‐dimensional (2D) temporal multi‐scale patching strategy, which converts one‐dimensional (1D) time series data into 2D multi‐scale Patch across different time periods. Additionally, MTSPnet incorporates two complementary modules: an interactive multilayer perceptron (MLPmix) module and a dynamic depthwise separable convolution (DDSC) module. These modules enable MTSPnet to efficiently extract both local and global temporal features, further enhancing its ability to model multi‐scale dependencies and periodicity variations. Experimental evaluations on seven real‐world datasets demonstrate that MTSPnet achieves superior performance in long‐term forecasting, proving its effectiveness as a robust and efficient solution for accurate time series prediction.

Recently, reconfigurable holographic surfaces (RHSs) have been proposed as a cost-efficient approach for achieving ultra-massive multiple-input multiple-output to satisfy the growing data rate requirements. Unlike existing antennas, the RHS follows a serial transmission characteristic where the electromagnetic (EM) wave propagates along the RHS element and is radiated outward one after another, concurrently with the gradual decrease of the EM wave energy. Thus, there exists a minimum RHS size that satisfies the data rate requirement in the RHS-enabled communication system without excessive antenna elements. In this paper, we first propose a serial propagation model of RHS to depict the coupling effect among RHS elements that is induced by the excitation of EM waves carrying signals one after another. Based on such a model, we derive the closed-form expression of the minimum required RHS size by estimating the upper and lower bounds of the achievable rate. Simulation results verify the theoretical analysis and reveal the relationship between the radiation characteristics of the RHS and the required minimum size.

Histopathological image retrieval aims to search histopathological images sharing similar content with the query image, which could provide pathologists with an approach to easily obtain similar diagnostic cases for reference. Recent histopathological image retrieval methods are usually based on CNN feature extractors, which require a large amount of annotated data for training. Besides, most of existing methods could not define a reasonable similarity metric for histopathological images. In this paper, we apply deep metric learning to liver histopathological image retrieval. We construct a model based on mixed attention mechanism and train the model with a modified version of multi-similarity loss, which enables embedding vectors of similar images in the given metric space to be closer and dissimilar ones to be far from each other. Additionally, our model can be well fitted with limited data. Finally, we evaluate the proposed method with our own established liver histopathological image dataset. Compared with several published methods, our model shows higher performance.

Disk additions to a RAID-4 storage system can increase the I/O parallelism and expand the storage capacity simultaneously. To regain load balance among all disks including old and new, RAID-4 scaling requires moving certain data blocks onto newly added disks. Existing data redistribution approaches to RAID-4 scaling, restricted by preserving a round-robin data distribution, require migrating all the data, which results in an expensive cost for RAID-4 scaling. In this paper, we propose McPod-a new data redistribution approach to accelerating RAID-4 scaling. McPod minimizes the number of data blocks to be moved while maintaining a uniform data distribution across all data disks. McPod also optimizes data migration with four techniques. First, it coalesces multiple accesses to physically successive blocks into a single I/O. Second, it piggybacks parity updates during data migration to reduce the cost of maintaining consistent parities. Third, it outsources all parity updates brought by RAID scaling to a surrogate disk. Fourth, it delays recording data migration on disks to minimize the number of metadata writes without compromising data reliability. We implement McPod in Linux Kernel 2.6.32.9, and evaluate its performance by replaying three real-system traces. The results demonstrate that McPod outperforms the existing “moving-everything” approach by 67.78-79.64 percent in redistribution time and by 14.24-27.16 percent in user response time. The experiments also illustrate that the performance of the RAID scaled using McPod is almost identical to that of the round-robin RAID.

In this work we introduce a probabilistic model that utilizes spatial contextual information to aid recognition when dealing with ambiguous segmentations of handwritten patterns. The recognition problem is formulated as an optimization problem in a Bayesian framework by explicitly conditioning on the spatial configuration of the letters. As a consequence, and in contrast to HMMs, the proposed model can handle duration modeling without an increase in computational complexity. We test the model on a real-world handwriting dataset and discuss several factors that affect the recognition performance.

In this work, we present a context-based model for tracking object features. More specifically, the context is defined as a collection of features within the local region that surrounds the feature that is being tracked. The model does not rely on any knowledge about the object, and therefore the collection of contextual features in one frame is just a hypothesis that is then reconfirmed or rejected in subsequent frames. We applied the model to tracking horizontal and vertical edges extracted from real-time video streams that contain moving vehicles. Our results show that the performance of the system improves as the number of the contextual features increases.