Large Language Models (LLMs) have emerged as powerful tools for diverse applications. However, their uniform token processing paradigm introduces critical vulnerabilities in instruction handling, particularly when exposed to adversarial scenarios. In this work, we identify and propose a novel class of vulnerabilities, termed Tool-Completion Attack (TCA), which exploits function-calling mechanisms to subvert model behavior. To evaluate LLM robustness against such threats, we introduce the Tool-Completion benchmark, a comprehensive security assessment framework, which reveals that even state-of-the-art models remain susceptible to TCA, with surprisingly high attack success rates. To address these vulnerabilities, we introduce Context-Aware Hierarchical Learning (CAHL), a sophisticated mechanism that dynamically equilibrates semantic comprehension with role-specific instruction constraints. CAHL leverages the contextual correlations between different instruction segments to establish a robust, context-aware instruction hierarchy. Extensive experiments demonstrate that CAHL significantly enhances LLM robustness against both conventional attacks and the proposed TCA, exhibiting strong generalization capabilities in zero-shot evaluations while still preserving model performance on generic tasks. Our code is available at https://github.com/S2AILab/CAHL.

The Metaverse, a convergence of cutting-edge technologies such as Virtual Reality (VR), Augmented Reality (AR), Artificial Intelligence (AI), and blockchain, represents the next frontier in digital society’s evolution. Meanwhile, as societal emphasis on education intensifies, a multitude of novel technologies are being integrated into the educational landscape with the aim of improving overall educational outcomes. This trend has resulted in the conceptualization of the Educational Metaverse as an extension of Metaverse applications in this domain. Notwithstanding its advent, there exists a notable dearth of scholarly efforts dedicated to meticulously summarizing and analyzing the most up-to-date research findings on the Educational Metaverse. Thus, this study systematically reviews recent literature to examine the Metaverse’s role in education, a field that stands to benefit significantly from the integration of these technologies. From an interdisciplinary perspective, guided by the insights derived from clustering algorithms, approximately 90 recent publications were meticulously analysed to explore and synthesise various aspects pertaining to the Educational Metaverse. This comprehensive review meticulously explores its primary characteristics, progression of foundatonal technologies, and multiplicity of practical implementations, thereby unravelling the transformative potential it holds for revolutionising pedagogical practices. The findings underscore the Educational Metaverse’s capacity to enhance learner engagement and motivation through immersive experiences, while also addressing the educational needs of individuals with special requirements. The study highlights the importance of optimizing the Educational Metaverse’s technical architecture and developing pedagogical strategies that are inclusive and effective. Furthermore, it emphasizes the need to ensure educational equity and accessibility, allowing all learners to harness the benefits of this technological innovation. As the Metaverse emerges as a pivotal force in educational modernization, it is anticipated that its applications will continue to proliferate. The research further indicates that the future trajectory of the Educational Metaverse is poised towards comprehensiveness and integration, with technology advancements emphasizing real-time capabilities and security enhancements. It is anticipated to extend its reach into a broader array of scenarios, thereby enabling a wider cross-section of disciplines and populations to reap the benefits it affords.

Given an untrimmed video, we consider localizing specific temporal segments according to the text description. Existing methods employ the computationally-expensive sliding window fashion to generate proposals and use simple operations which are not expressive enough to combine the video and language information. To address these two issues, we propose a novel model, namely Bidirectional Single-Stream Temporal Localization (BSSTL) which effectively and effi-ciently localizes language queries in videos. We make two contributions: (1) Our BSSTL can run continuously in a video stream, without dividing input videos into short overlapping temporal windows for batch processing. (2) We propose a novel Attentive Cross-modal Fusion approach to jointly project sentence queries and video segments. For evaluation, We evaluate BSSTL on ActivityNet Captions [1] and TACoS datasets [2] and our model achieves state-of-the-art performance.

Head pose estimation has attracted many research interest in recent years. With the advent of deep learning, it is possible to predict the head pose accurately from the RGB images without the help of facial landmarks or depth information. However, existing head pose datasets often lack large pose head images, which extremely limits the development of head pose estimation algorithms. In this paper, we build the largescale head pose dataset (LHPD) including more than 140,000 images with the diverse and accurate head poses. The LHPD dataset includes the head images recorded from different shooting angles between the camera and the human body for the first time, which greatly expands the range of head pose compared to previous datasets. Therefore, the range of head pose can cover +/-90° for each Euler angle. The accurate and reliable head pose annotation is labeled by the motion capture system and careful calibration procedures. We then propose a head pose estimation method through fine-tuning the ResNet on the LHPD dataset when using the Euclidean distance of quaternions as the loss function. The results show that our method achieves better performance than current state-of-theart algorithms.