With the rapid advancement of technology, large language models (LLMs) have become key content generators that shape social discourse and exert far-reaching influence. However, the ability of these models to produce potentially harmful or inappropriate content poses a significant threat to the health and harmony of the online environment. In response, researchers have actively worked to guide models in generating content that aligns with universally accepted societal values, aiming to curb the spread of malicious information. Despite these efforts, the issue of “jailbreak” attacks remains a critical challenge. This not only tests the technological boundaries of LLMs but also raises significant concerns regarding the ethical application and social responsibility of such models. This paper explores the diversity of jailbreak attacks and the complexity of defense algorithms from both white-box and black-box perspectives. Additionally, key evaluation metrics such as attack success rate, robustness, efficiency, and portability are discussed to provide a comprehensive framework for assessing the effectiveness of offensive and defensive strategies. In conclusion, this paper offers a holistic view of LLM jailbreak attacks and defenses, underscoring the importance of vigilance and proactive exploration of effective defense strategies as LLMs continue to be widely deployed.
In recent years, with the rapid advancement of diffusion-based image generation and virtual try-on technologies, novel research methodologies integrating multimodal inputs such as text and images have emerged. These approaches have become a significant direction in fashion image editing. However, most existing datasets only provide images with basic attribute annotations, lacking systematic support for explicit structural constraints such as sketches and structure maps. This makes it difficult to ensure structural consistency and controllability during the editing and generation processes. Simultaneously, existing textual descriptions often remain confined to superficial semantics like color and category, lacking multidimensional annotation information regarding material properties and design styles. This limits the model’s ability to understand and generalize complex instructions and authentic fashion contexts. To address this, we propose constructing a novel benchmark dataset: FashionAtlas. This dataset provides more refined sketches as structural constraints to enhance the controllability of the generated results. Simultaneously, we introduce more detailed and layered prompts generated by large language models (LLMs), integrating structured clothing attribute ontologies with specialized terminology systems. This enhances the model’s comprehension of fashion semantics and styling logic, thereby better supporting complex editing and generation requirements. Experimental results demonstrate that FashionAtlas outperforms traditional image-annotation-only baseline dataset in terms of generation accuracy, semantic consistency, detail fidelity, and user controllability. This benchmark dataset aims to fill the current gap in combining structured constraints with fine-grained semantic prompts, providing a new research standard for multimodal control editing and in-context generation tasks in the fashion domain.
As specialized chips designed to accelerate artificial intelligence (AI) algorithms, with core capability in enabling large-scale parallel computing, AI chips have been widely adopted in diverse fields, including smart home appliances, robotics, mobile terminals, and automotive manufacturing. This extensive application has further drawn industry attention to verifying whether their computing performance meets the demands of practical scenarios. However, existing AI chip evaluation systems still lack systematic integration, and conventional quality assurance methodologies cannot fully address the uniqueness of AI chips. As an emerging and indispensable process, AI chip evaluation plays a pivotal role in guiding the selection and practical deployment of AI chips. This work systematically sorts out the current status, inherent characteristics, and relevant standards of domestic and international existing AI chip evaluation systems. Furthermore, it puts forward targeted suggestions for future research directions of AI chip evaluation methods, aiming to provide a reference for optimizing AI chip evaluation systems and promoting the healthy development of the AI chip industry.
This study investigates the International Civil Aviation Organization’s (ICAO) methodology for calculating aviation carbon emissions, focusing on empirical validation and bias analysis of its assumptions regarding flight distance and payload. Using full-sample operational data from January 1, 2023, to July 31, 2024, the research examines routes between Kunming airport (KMG) and Mangshi airport (LUM), Shuangliu airport (CTU), Xianyang airport (XIY), and Daxing airport (PKX), analyzing the impact of directional, seasonal, and airline-specific operational factors on load factors, baggage weights, and cargo strategies. Findings reveal that ICAO’s current method—relying on a fixed assumption of 100kg per passenger, overlooking return-trip load imbalances, and ignoring variations in airline cargo strategies—introduces systematic biases in emission attribution. The study concludes that the static, average-based framework lacks scientific robustness under dynamic operational conditions, and calls for methodological adjustments to address multi-dimensional uncertainties.
With the resurgence of lunar exploration, sustainable communication and navigation infrastructure has become critical. This paper addresses the trade-off between miniaturized and large-platform architectures for lunar constellations by proposing a systematic evaluation framework. The core idea of this framework is to separate design constraints from performance metrics: first, engineering models are used to place different schemes under unified equivalent resource constraints, ensuring a fair comparison; then, a hierarchical threshold-based (“pyramid”) evaluation method is employed. This model sequentially filters schemes based on (1) a qualification threshold (basic service availability), (2) a reliability threshold (N-1 fault tolerance), and finally (3) ranks geometric dilution of precision (GDOP) and daily data throughput in a performance optimization layer. This framework is applied to evaluate typical configurations with varying satellite numbers and individual satellite masses. Simulation results validate the effectiveness of the evaluation framework, revealing potential limitations of the “large-platform” scheme in basic availability, while the all-miniaturized configuration demonstrates robust availability and reliability. The research findings not only provide a systematic evaluation framework and quantitative basis for the engineering design of lunar positioning, navigation, and timing (PNT) and communication systems but also lay a theoretical foundation for the evolution of advanced technologies such as the global coverage extension of lunar communication and navigation constellations.
