Wenjuan Li

Background: Fine‐tuning has become a core mechanism for adapting pretrained Large Language Models (LLMs) to downstream tasks. However, the dependence of fine‐tuning on training data, parameter update mechanisms, and reusable components provides entry points for attackers. Related threats have evolved from data poisoning and weight tampering to agent behavior manipulation and interface exploitation, while defensive research has expanded from preimmunization to post‐hoc remediation. However, existing reviews do not yet provide a unified framework that organizes attack and defense methods across the complete fine‐tuning lifecycle. Objective: This paper serves as a systematic survey of LLM security in fine‐tuning scenarios and establishes a lifecycle‐based framework for comparing attack and defense methods, complemented by unified empirical evaluation. Methods: Fine‐tuning‐related attack and defense mechanisms are divided into three phases according to the timing of intervention: the pre‐tuning, during‐tuning, and post‐tuning phases. Within each phase, attack and defense strategies are reviewed and contrasted to expose their evolutionary relationships and limitations. Representative methods from each phase are then evaluated under a unified model selection, hardware setup, and evaluation protocol, with additional cross‐phase experiments pairing attacks and defenses from different phases. Results: Unified evaluation reveals that attack effectiveness is highly model‐dependent and nonmonotonic with scale: weight‐editing attacks that succeed on earlier models lose impact on modern open‐source LLMs; cross‐lingual backdoor transfer, reported as near‐perfect at larger scales, fails entirely on tested 1B‐4B models; and purely benign fine‐tuning samples can compromise safety alignment in instruction‐tuned models. Cross‐phase experiments further show that single‐phase defenses rarely generalize to attacks from other phases, and that defense effectiveness depends jointly on model architecture and alignment state. Conclusion: Based on the survey and experimental findings, this paper identifies key open problems, including configuration‐robust defense, cross‐phase defense composition, and embedding‐space attacks beyond behavioral assumptions‐and proposes concrete directions for future research.

Humanoid robotics represents a rapidly evolving research domain that integrates artificial intelligence and robotics. Despite significant advances, existing reviews have predominantly focused on narrow technical aspects and lack comprehensive analysis from academic and industrial perspectives. This paper presents a systematic dual-perspective survey, in which academic literature, commercial products, and industry reports are extensively analyzed. A comprehensive taxonomic framework and systematic review of key enabling technologies are established, including ontological structures, perception systems, locomotion control, intelligent decision-making algorithms, foundation model integration, and human-robot interaction (HRI) technologies. From academic and industrial perspectives, research progress across diverse applications is examined, and a detailed comparative analysis of commercial products from leading companies, including Tesla, Boston Dynamics, and UBTECH, is performed. Six major challenge categories are identified: hardware design limitations, control system complexities, perception constraints, HRI difficulties, application-specific requirements, and ethical considerations. In addition, the transformative impact and integration challenges of large language models are particularly discussed. Seven promising research directions are outlined, and a systematic academic-industrial gap analysis is conducted. Consequently, significant disparities and technology transfer bottlenecks are identified, and successful collaboration models are examined. This comprehensive survey provides the first systematic examination combining academic research insights with industrial development analysis. It thus offers valuable guidance for researchers, engineers, and policymakers working toward more capable, affordable, and socially integrated humanoid robots.

Addressing Artificial Intelligence (AI) faculty deficiencies in higher education, this paper develops the AI+ Dual-qualified Teacher Capability Ecology Model (AI-DTCEM) based on Capability Ecology Theory. The model is developed after a thorough analysis of the current state of new engineering talent cultivation in universities and the innovative practical abilities required in the AI+ environment. This paper proposes an implementation framework characterized by “three-dimensional collaboration, four-tier progression, and five-element drive.” Additionally, it uses the collaborative education project involving Hangzhou Normal University, Zhejiang University, and Hangzhou Ruishu Technology Co., Ltd. as a backdrop to introduce a deep collaborative education model, showcasing the theoretical and practical achievements of this project. Using NetLogo as the simulation platform, this paper designs a 96-month system dynamics experiment to compare and analyze the outcomes of four scenarios: the baseline experiment, the AI-enhanced experiment, the policy-driven experiment, and the comprehensive optimization experiment. This study reveals the following findings: (1) Policy-driven initiatives are crucial for the successful construction of dual-qualified teacher teams, with the policy-driven scenario achieving the highest overall skill level (9.332). (2) The application of AI technology significantly enhances teacher skill development, resulting in AI skill improvements ranging from 116.6% to 163.4%. (3) The comprehensive optimization scenario (utilizing a collaborative mechanism) achieves systemic advantages, realizing a 100% dual-qualified teacher ratio. However, this comes with diminishing marginal returns on investment. This research provides a theoretical foundation, quantitative analysis, and practical pathways for developing dual-qualified teacher teams in the AI+ era.

Hyperuricemia has seen a continuous increase in incidence and a trend towards younger patients in recent years, posing a serious threat to human health and highlighting the urgency of using technological means for disease risk prediction. Existing risk prediction models for hyperuricemia typically include two major categories of indicators: routine blood tests and biochemical tests. The potential of using routine blood tests alone for prediction has not yet been explored. Therefore, this paper proposes a hyperuricemia risk prediction model that integrates Particle Swarm Optimization (PSO) with machine learning, which can accurately assess the risk of hyperuricemia by relying solely on routine blood data. In addition, an interpretability method based on Explainable Artificial Intelligence(XAI) is introduced to help medical staff and patients understand how the model makes decisions. This paper uses Cohen’s d value to compare the differences in indicators between hyperuricemia and non-hyperuricemia patients and identifies risk factors through multivariate logistic regression. Subsequently, a risk prediction model is constructed based on the parameter optimization of five machine learning models using the PSO algorithm. The accuracy and sensitivity of the proposed particle swarm fusion Stacking model reach 97.8% and 97.6%, marking an improvement in accuracy of over 11% compared to the state-of-the-art models. Finally, a sensitivity analysis of factors affecting the prediction results is conducted using the XAI method. This paper has also developed a Health Portrait Platform that integrates the proposed risk prediction models, enabling real-time online health risk assessment. Since only routine blood test data are used, the new model has better feasibility and scalability, providing a valuable reference for assessing the risk of hyperuricemia occurrence.

The intermittent nature of solar energy poses significant challenges to the integration of photovoltaic (PV) power generation into the electrical grid. Consequently, the precise forecasting of PV power output becomes essential for efficient real-time power system dispatch. To meet this demand, this paper proposes a deep learning model, the CA-Transformer, specifically designed for PV power output prediction. To overcome the shortcomings of traditional correlation coefficient methods in dealing with nonlinear relationships, this study utilizes the Copula function. This approach allows for a more flexible and accurate determination of correlations within time series data, enabling the selection of features that exhibit a high degree of correlation with PV power output. Given the unique data characteristics of PV power output, the proposed model employs a 1D-CNN model to identify local patterns and trends within the time series data. Simultaneously, it implements a cosine similarity attention mechanism to detect long-range dependencies within the time series. It then leverages a parallel structure of a 1D-CNN and a cosine similarity attention mechanism to capture patterns across varying time scales and integrate them. In order to show the effectiveness of the model proposed in this study, its prediction results were compared with those of other models (LSTM and Transformer). The experimental results demonstrate that our model outperforms in terms of PV power output prediction, thereby offering a robust tool for the intelligent management of PV power generation.