With the rapid development of large language models, their broad application prospects have attracted significant attention from both the academic and industrial communities.Before a large language model is applied to practice,its performance and potential risks need to be comprehensively evaluated.In recent years,the evaluation methods of large language models have been studied from multiple perspectives by researchers.In this paper，the evaluation metrics,methods and benchmarks of large language models in terms of performance,robustness,and alignment,are reviewed systematically and the advantages and disadvantages of various evaluation metrics and methods are analyzed.Finally,the future research directions and challenges of large language model evaluation are discussed.

In order to evaluate the comprehensive performance of brain-inspired intelligence,it is necessary to establish a standard brain-inspired intelligence evaluation framework.In this paper,the theoretical foundations of brain-inspired intelligence evaluation are elaborated,including brain science theory,brain-inspired computing theory,the synergistic development of brain science and brain-like computing,and the key issues in brain-inspired intelligence evaluation.Brain-inspired intelligence evaluation methods,including the brain-mechanism based assessment methods and the brain-inspired modeling based assessment methods,are discussed in detail.Corresponding brain-inspired evaluation metrics were proposed alongside the construction of a difficulty-tiered standardized dataset.In addition，the hardware-software integrated evaluation of the brain-inspired intelligent systems is further considered,an integrated application framework is proposed,and a systematic evaluation practice is carried out by focusing on the brain-inspired localization and navigation tasks.

Recently,artificial intelligence,particularly deep learning,has developed rapidly，which deeply empowers traditional industries and leads the realization of a new round of industrial technology revolution.However,the size of transistors on electronic chips is gradually approaching the physical limit,resulting in the inability of traditional electronic neural networks to meet the exponential increasing demand for computing power.Benefiting from the unique advantages of photon,optical computing technology merges optoelectronic technology with neural network models,which has the advantages of parallelization,high speed,low power consumption and multi-dimensional processing.The research progress of optical neural networks is reviewed,concentrating on the computational architecture of optical diffractive neural networks.The challenges facing the practical implementation of large-scale optical diffractive neural networks are analysed and summarized，and the future development trend is prospected.

Knowledge graph is an important form of knowledge representation,which can integrate and organize information effectively.It has been widely used in search engines,intelligent question answering and recommendation systems.Traditional knowledge graph construction relies on manual annotation and rule-based systems,which is huge in scale and uneven in quality,and is difficult to adapt to the dynamic changes of massive data.Recently,large models have shown superior performance in knowledge generation.However,there is still a lack of research on large language models to enhance knowledge graph error repairing.Therefore,a progressive error correction method for knowledge graphs,assisted by large language models,has been proposed.Using embedding models to evaluate the quality of knowledge triples and high-quality triples as prompts for learning content,knowledge correction by large language models is realized.Based on extensive experiments,the proposed method significantly enhances the reasoning ability of knowledge graphs.

With the increasing complexity of the internal structure of deep neural network (DNN) ,it is difficult for people to have an intuitive understanding of its internal operation mechanism,so the probability of model errors is greatly increased.Therefore,an effective DNN robustness test method is needed to solve the trust crisis of the model to ensure the reliability and security of the software system.The existing DNN robustness test methods mostly target the coverage of neurons for generating perturbation samples,without introducing more information about the internal model,resulting in a high degree of perturbation and a large amount of redundancy in the generated perturbation samples,which greatly limits the ability to improve model robustness.A new adversarial example generation method is proposed.Firstly,a decision tree is constructed by the last convolutional layer of the model.The judgment path in the decision tree is regarded as the decision path,and each filter in the path is analyzed to find out the impact factor.Finally,the perturbed samples were generated according to the decision path and impact factors.The test results show that the test samples generated are 78% less than the existing state-of-the-art fuzzing method DLFuzz in terms of perturbation degree on average,and the number of original samples perturbed by our method is 27.7% more on average.

With the development of UAV technology,effectively measuring the intelligence level of UAVs has become an important issue.Traditional field measurement methods are costly,inefficient,and sensitive to environmental factors.Therefore,based on Unreal Engine and the AirSim platform,a virtual simulation platform for measuring UAV perception and decision-making capabilities was proposed,using the performance of UAV flight tasks in virtual environments to assess their perception and decision-making capabilities.The platform simulates  UAV flight tasks in complex environments with high-precision scenes and dynamic factors such as weather and traffic,supporting the testing and evaluation of various UAV perception and decision-making algorithms.Compared to field measurements,by this method costs and time are significantly reduced while enhancing flexibility.Test results show that the platform has significant advantages in interactivity and simulation effectiveness,providing effective support for optimizing and developing UAV algorithms,with broad application prospects and practical value.

Digital watermarking achieves data traceability by embedding unique identifiers in the data,which not only improves the reliability of the output of the intelligent model,but also enhances the public’s trust in the AI system.However,the current digital watermarking technology,especially in the field of AI,lacks a scientific and unified evaluation process and specification.The application of image watermarking technology and audio watermarking technology in the field of artificial intelligence are focused on,combining the actual use scenarios of specific intelligent models,scientific evaluation indexes are designed and the standardisation of evaluation is achieved.Meanwhile,an evaluation platform is constructed to achieve test automation and evaluation integration,which provides strong support for improving the traceability and application level of watermarking technology in the application field of artificial intelligence models,and is of great significance for ensuring the safe and controllable development of artificial intelligence.

At present, artificial intelligence technology is booming, and a variety of artificial intelligence models and products are launched at home and abroad，and artificial intelligence has constantly influence on people’s lives.Data is one of the core factors in artificial intelligence，and the development of artificial intelligence technology benefits from the high-quality data.Therefore,the measurement and evaluation of data for artificial intelligence is an important precondition to achieve legality,safety and fairness of artificial intelligence.For the metrology and evaluation of data,the legitimacy,authenticity,diversity,balance,data privacy protection and ethics,data quantity are proposed as the fundamental contents of measurement and evaluation of data.