Large language models (LLMs) have been increasingly applied across a wide range of domains. However, recent studies have identified the presence of certain glitch tokens in their vocabularies, which can trigger hallucinations and lead to unpredictable or even harmful outputs. While various methods have been proposed to detect such tokens, effectively repairing them remains a key challenge for ensuring the reliability of LLMs. In this work, we propose GlitchCleaner, a lightweight yet effective approach to mitigate the adverse effects caused by glitch tokens. GlitchCleaner introduces auxiliary branches into specific components within selected layers of the model, enabling efficient and targeted token repair. These branches are implemented using the low-rank adaptation (LoRA) technique, adding less than 0.1% additional parameters to the original model. Furthermore, a gating mechanism dynamically controls the activation of these branches based on the model’s input, ensuring precise intervention without disrupting normal inference behavior. Experimental results across multiple mainstream models demonstrate that our method achieves an average repair rate of 86.88%, representing an improvement of over 30% compared to existing approaches, while ensuring lossless preservation of the model’s baseline capabilities and causing negligible impact on inference speed.

Large Vision-Language Models (LVLMs) rely on attention-based retrieval of safety instructions to maintain alignment during generation.Existing attacks typically optimize image perturbations to maximize harmful output likelihood, but suffer from slow convergence due to gradient conflict between adversarial objectives and the model's safety-retrieval mechanism.We propose Attention-Guided Visual Jailbreaking, which circumvents rather than overpowers safety alignment by directly manipulating attention patterns.Our method introduces two simple auxiliary objectives: (1) suppressing attention to alignment-relevant prefix tokens and (2) anchoring generation on adversarial image features.This simple yet effective push-pull formulation reduces gradient conflict by 45% and achieves 94.4% attack success rate on Qwen-VL (vs.68.8% baseline) with 40% fewer iterations.At tighter perturbation budgets (ϵ=8/255), we maintain 59.0% ASR compared to 45.7% for standard methods.Mechanistic analysis reveals a failure mode we term safety blindness: successful attacks suppress system-prompt attention by 80%, causing models to generate harmful content not by overriding safety rules, but by failing to retrieve them.

Large Vision-Language Models (LVLMs) have achieved remarkable success in video understanding tasks, yet they suffer from substantial computational overhead due to the large number of visual tokens generated from video frames. A key observation is that consecutive video frames often contain significant temporal redundancy, especially in static or slowly-changing scenes. In this paper, we propose MotionKV, a training-free, motion-aware KV cache compression framework that selectively retains key-value pairs corresponding to high-motion frames while compressing those from static regions. Our method computes per-frame motion scores using pixel-level difference analysis, then applies a hybrid selection strategy that combines motion-based importance with background-aware penalties to guide token-level retention. Extensive experiments on three video question-answering benchmarks demonstrate that MotionKV retains 12.5% of visual tokens while maintaining 98.8% of full-cache accuracy on VideoMME (Qwen2-VL-7B), achieving 3.1 × inference speedup with 68% reduction in peak memory. The approach is model-agnostic and requires no fine-tuning, enabling seamless integration into existing LVLMs.

Magnetotelluric (MT) and teleseismic tomography are critical techniques in deep Earth exploration for geodynamic studies, internal energy circulation, plate tectonics, volcanic systems and so on. However, due to the uneven data distribution, the inherent non-uniqueness of inversion, and differing sensitivities to subsurface media, the velocity and resistivity structures derived from independent inversions often exhibit substantial discrepancies and contradict conventional geological understanding. To address these issues, we propose a novel three-dimensional (3D) joint inversion method for MT and teleseismic traveltime data, applicable to different types of grid discretization. The method first establishes a parameter mapping for different grid types. Then, a joint inversion framework based on the local Pearson correlation constraints (LPCC) is developed using a virtual grid technique. The inversion alternately updates the resistivity and velocity models by ensuring structural similarity between them. Numerical experiments demonstrate that the proposed method can effectively integrate the advantages of both techniques, enhance resolution and stability. Additionally, the tests on hyperparameter selection provides guidance for optimizing joint inversion parameters. Finally, the developed joint inversion method is applied to MT and teleseismic traveltime data in southeastern Australia and yields a constrained 3D resistivity and velocity model of the region. The results offer significant theoretical and practical insights into the lithospheric structure, the magma transport pathways, and the geodynamic processes in this area.

The globalization of the IC supply chain has increased the possibility for adversaries to insert hardware Trojans (HTs), which are always inserted into nodes with high concealment. Although several methods have been introduced to detect HTs, they do not specifically focus on node concealment and are not accurate enough. In this paper, 11 new features are proposed reflecting the node concealment, i.e, the controllable difficulty and observable difficulty of nodes, for machine-learning-based HT detection. The proposed features comprehensively consider the circuit structure information and signal propagation characteristics, and quantify the possibility of attacks. After combining the proposed features with existing ones for classification, the true positive rate (TPR) is 35 points higher than that without using the proposed features. Compared with the state-of-the-art works, our TPR and true negative rate (TNR) have increased by 15.6% and 2.1%, respectively, especially in large-scale circuits. It shows that our method has a great advantage in detection accuracy.

Recommendation system provides prediction of users' preference and guidance for users' purchase. Probabilistic matrix factorization (PMF) as a widely-used collaborative filtering approach so far, generally has excellent performance for personalised recommendation. Nevertheless, it still suffers from data sparsity problem. In this paper, we present a novel recommendation algorithm CR-IIA which combines matrix factorization and association rule mining. Our method employs ItemCorrelation to measure the latent correlation among items given by association rule mining. And ItemCorrelations are integrated into PMF model to optimize feature vectors of items. The factor of ItemCorrelations makes rating prediction more accurate and realistic. Compared with social trust model and context-based model, our approach improves the accuracy of ratings prediction without aggravating data sparsity problem caused by additional information. Moreover, our framework also allows additional dimensionality to be considered simultaneously, which benefits subsequent research in personalised recommendation based on matrix factorization. We conduct experiments on a real- world data set MovieLens, and the results show that our model outperforms the item-oriented approaches.

This paper proposes an optimal locating and sizing method of distributed generations (DGs) considering local consumption and power export potential. Based on the analysis of seasonal and temporal characteristics of different types of DGs and load, the proposed DGs locating and sizing model, is to minimize the total costs of DGs depreciation, network power losses, and redundant power. Under the premise of load demand locally satisfied, the optimal location and size of DGs is obtained by reducing the redundant power of DGs. Finally, the IEEE 33-bus test system is applied to validate the effectiveness and practicability of the proposed method.

Identifying key nodes is an important research issue in social networks. Most of current social networks are weighted networks as well as consist of multiple communities. A suitable centrality measure for weighted social networks should be capable of finding most important nodes in each community. However, based on the existing centrality measure for weighted social networks, the most influential nodes are closely gathered in one community or distribute in a portion of all communities. In this paper, we propose a Tie Strength Matrix based Principal Component Centrality (TSM-based PCC), which extends PCC, a centrality measure for unweighted networks, to weighted social networks. Experiment results show that, based on TSM-based PCC influential nodes can be picked out accurately in real social network datasets. Furthermore, TSM-based PCC outperforms other centrality measures in identifying important nodes in each community. Hence the proposed TSM- based PCC is feasible and effective in weighted social networks.

In this paper, an intelligent genetic algorithm (IGA) and a double layer neural network (NN) are integrated into a hybrid intelligent algorithm for solving the quadratic bilevel programming problem. The intelligent genetic algorithm is used to select a set of potential solution combinations from the entire generated combinations of the upper level. Then a meta-controlled Boltzmann machine, which is formulated by comprising the Hopfield model (HM) and the Boltzmann machine (BM), is used to effectively and efficiently determine the optimal solution of the lower level. Numerical experiments on examples show that the genetic algorithm based double layer neural network enables us to efficiently and effectively solve quadratic bilevel programming problems.