Knowledge editing aims to efficiently modify the internal knowledge of large language models (LLMs) without compromising their other capabilities. The prevailing editing paradigm, which appends an update matrix to the original parameter matrix, has been shown by some studies to damage key numerical stability indicators (such as condition number and norm), thereby reducing editing performance and general abilities, especially in sequential editing scenario. Although subsequent methods have made some improvements, they remain within the additive framework and have not fundamentally addressed this limitation. To solve this problem, we analyze it from both statistical and mathematical perspectives and conclude that multiplying the original matrix by an orthogonal matrix does not change the numerical stability of the matrix. Inspired by this, different from the previous additive editing paradigm, a multiplicative editing paradigm termed Multiplicative Orthogonal Sequential Editing (MOSE) is proposed. Specifically, we first derive the matrix update in the multiplicative form, the new knowledge is then incorporated into an orthogonal matrix, which is multiplied by the original parameter matrix. In this way, the numerical stability of the edited matrix is unchanged, thereby maintaining editing performance and general abilities. We compared MOSE with several current knowledge editing methods, systematically evaluating their impact on both editing performance and the general abilities across three different LLMs. Experimental results show that MOSE effectively limits deviations in the edited parameter matrix and maintains its numerical stability. Compared to current methods, MOSE achieves a 12.08% improvement in sequential editing performance, while retaining 95.73% of general abilities across downstream tasks.

Vision-Language Models (VLMs) like CLIP have advanced image representation through open-vocabulary semantic alignment. Yet, existing few-shot transfer learning methods largely overlook the intrinsic interdependencies between text and image embeddings, limiting their ability to fully transfer CLIP’s pretrained capabilities. To address this gap, we propose Hyperspherical Interpolation Variational Encoding (HIVE), a novel method for few-shot image classification. Our core idea is to shift away from directly training feature extraction capabilities for downstream tasks, and instead focus on exploring the semantic transformation relationships between upstream and downstream tasks. By modeling semantics from coarse to fine granularity, HIVE enables the transfer of original feature extraction and modality alignment capabilities to downstream tasks. Extensive experiments on eight established benchmarks, including CUB and EuroSAT, validate HIVE’s efficacy, achieving up to 46.2% and 80.0% improvements over the original CLIP in 1-shot and 16-shot classification tasks, respectively. Our work underscores the importance of preserving pretrained geometric constraints while exploiting semantic hierarchies for effective few-shot adaptation, providing a principled approach for vision-language model customization.

Large language models (LLMs) often suffer from outdated or incorrect knowledge, prompting ongoing research into efficient model editing . Existing methods, however, mainly target individual knowledge facts. When multiple facts need to be edited in a coherent sequence, they frequently lead to deviations or even breakdowns in model’s general abilities. This problem intensifies in batch-sequential editing, where multiple facts are updated simultaneously, compared to single-sequential editing. In this work, by analyzing the parameter matrix, we identify that the degradation stems from unintended modifications that should ideally remain unaffected. These changes accumulate with the number and batch size of edits, ultimately harming editing performance and general abilities. To address this, we propose Batch-Aware Editing Anchor Compression (B-EAC) , a framework tailored for sequential model editing. B-EAC dynamically selects essential anchors for each edit while compressing the influence on nearby parameters. It adopts a layer-wise anchor selection strategy to prevent anchor conflicts during concurrent edits and introduces a rolling anchor refresh mechanism to enhance adaptability across batches. Experiments conducted on three LLMs across four tasks demonstrate that B-EAC effectively suppresses deviation during model editing, achieving a 36.54% performance improvement compared to the case without it. Our work offers a practical and theoretically grounded framework for updating LLMs efficiently, paving the way for continual knowledge refinement in real-world applications.

The presence of shadows degrades the performance of various multimedia tasks. Image shadow removal aims at restoring the background of shadow regions, which is generally an open challenge. Unlike most existing deep learning-based methods that focus on restoring such degradations in the spatial domain, we introduce a novel shadow removal method that also exploits frequency domain information. Specifically, we firstly revisit the frequency characteristics of shadow images via Fourier transform, where amplitude components contain most lightness information and phase components are related to structure information. To this end, we propose a two-stage deep Fourier shadow removal network (FSR-Net) to enhance the brightness of shadow regions, and correspondingly improve the shadow removal performance of whole images. For each stage, it consists of an amplitude recovery network and a phase recovery network to progressively reconstruct the lightness and structure components. To facilitate the learning of these two representations, we introduce the frequency and spatial interaction blocks to process the local spatial features and the global frequency information separately. Extensive experiments demonstrate that FSR-Net achieves superior results than other approaches with fewer parameters. For example, our method obtains a 1.05dB improvement on ISTD[34] dataset over the previous state-of-the-art method [43] with 0.30M parameters.