Abstract:With the widespread application of large language models， the evaluation of large language models has become crucial. In addition to the performance of large language models in downstream tasks， some potential risks should also be evaluated， such as the possibility that large language models may violate human values and be induced by malicious input to trigger security issues. This paper analyzes the commonalities and differences between traditional software， deep learning systems， and large model systems. It summarizes the existing work from the dimensions of functional evaluation， performance evaluation， alignment evaluation， and security evaluation of large language models， and introduces the evaluation criteria for large models. Finally， based on existing research and potential opportunities and challenges， the direction and development prospects of large language models evaluation technology are discussed.

Abstract:As ChatGPT and other foundation-model-based products demonstrate powerful general performance， both academia and industry are actively exploring how to adapt these models to specific industries and application scenarios， a process known as the customization of domain-specific foundation models. However， the existing general-purpose foundation models may not fully accommodate the patterns of domain-specific data or fail to capture the unique needs of the field. Therefore， this paper aims to discuss the methodology for customizing domain-specific foundation models， including the definition and types of foundation models， the description of their general architecture， the theoretical foundations behind the effectiveness of foundation models， and several feasible methods for constructing domain-specific foundation models. By presenting this content， we hope to provide guidance and reference for researchers and practitioners in the customization of domain-specific foundation models.

Abstract:The chain of thought （CoT） prompts enable large language models to process complex tasks according to specific reasoning steps， allowing them to demonstrate stronger capabilities in common sense reasoning， mathematical logic reasoning， and interpretability. However， the main drawback of the CoT approach lies in its reliance on massive language models， which typically have billions of parameters and face challenges in large-scale deployment. To address this issue， this paper proposes a large model knowledge distillation method based on the CoT， aiming to fully leverage the thinking and reasoning capabilities of large language models. Through knowledge distillation techniques， the main goal is to guide smaller models in solving complex tasks.This study adopts a large model as the teacher model and a small model as the student model， fine-tuning the student model by acquiring reasoning data from the teacher model. Through a series of carefully designed methods， such as changing data generation methods， clustering-based sampling of question-answer examples， heuristic correction of examples， and adaptive generation of answers， this study makes the generation process of the teacher model more efficient， resulting in higher-quality and larger quantities of reasoning data. This enables better fine-tuning of the student model， allowing it to acquire strong reasoning capabilities and achieve efficient knowledge distillation. The framework of this study aims to establish an effective knowledge transfer mechanism， allowing the deep thinking of large models to effectively guide smaller models， providing more intelligent and efficient solutions for solving complex tasks. Through this approach， we hope to overcome the challenges of deploying large models and promote the application and advancement of language models in the real world.

Abstract:Addressing the decision-making conundrum faced by commanders in response to major sudden incidents， this paper proposes a coordination framework for collaborative disposal of multi-intelligent agents based on large language models. The framework optimizes collective decision-making efficiency and action planning through strategies such as agent role generation， multi-level Monte-Carlo tree and interactive prompt learning. It introduces hierarchical mechanisms and workflow management concepts， enhancing collaboration efficiency through the reward function shared among agents. A transparent and implicit communication model ensures node status consistency. Experimental results demonstrate that the framework performs well under various scenarios， significantly improving reaction speed and response efficiency compared to traditional task allocation methods.

Abstract:Fine-tuned large language models （LLMs） perform exceptionally well in various tasks， but centralized training poses user privacy leakage risks. Federated learning （FL） mitigates data sharing issues through local training， yet the large parameter size of LLMs challenges resource-constrained devices and communication bandwidth， making deployment in edge networks difficult. Considering split learning （SL）， federated split learning can effectively address these issues. Given the more pronounced influence of deep-layer model weights and the discovery that training certain layers yields slightly lower accuracy compared to training the entire model， we opt to split the model based on Transformer layers. Additionally， utilizing low-rank adaption （LoRA） can further reduce resource overhead and enhance security. Therefore， at each device， we only perform LoRA and training on the final few layers. These adapted layers are then uploaded to the server for aggregation. From the perspective of cost reduction and ensuring model performance， we propose a fine-tuning method for the pre-trained model RoBERTa based on federated split learning and LoRA. By jointly optimizing the computational frequency of edge devices and the rank of model fine-tuning， we maximize the rank to improve model accuracy under resource constraints. Simulation results indicate that only training the last three layers of the LLMs can improve model accuracy within a certain range （1—32） by increasing the rank. Additionally， increasing the per-round delay and the energy threshold of devices can further enhance model accuracy.

Abstract:Intelligent command is an important direction for the new command and control theories， and large language models are important support for the realization of intelligent command capabilities such as intelligent interaction， task planning， and auxiliary decision-making. Combining theory and practice， we outline the military capability requirements of the large model and design a large language model application framework for intelligent command. Then， the system architecture， information process， and collaborative architecture of the “Aiwu large model+” system are proposed and the key technologies for engineering implementation are proposed. Empirical cases of intelligent command are used to verify the multimodal interaction and military language understanding of the system. Collaboration and command control of manned/unmanned platforms can be expanded， which provides reference for research and implementation of the major national defense and military special projects and the intelligent command in the future.

Abstract:The existing imputation methods are roughly divided into statistical methods and deep learning methods. The statistical methods can only capture the linear time relationship， which makes it impossible to accurately capture the relationship of non-linear time series data. The deep learning imputation methods usually donot consider the correlation between different time series. To solve these problems， a new model jointing the tensor completion and the recurrent neural network is proposed. Firstly， the multivariate time series are modeled as a tensor， and the correlation of different time series is captured by low rank tensor completion. Secondly， a time based dynamic weight is proposed to fuse the tensor completion results with the prediction results of the recurrent neural network to avoid the accumulation of prediction error caused by continuous missing. The proposed method is evaluated on several real time series datasets， and the results show that the proposed model outperforms the existing models in term of imputation accuracy， which is helpful for improving classification and regression accuracy.

Abstract:Blind face restoration is to recover high quality face from unknown degradation， and the ill-posed problem often results in local texture missing or mismatched facial components for restored images， therefore a degraded blind face restoration algorithm based on feature fusion and embedding optimization is proposed. By extracting face prior features from degraded inputs， using multi-headed cross-attention for feature interaction fusion and global context modeling， embedding facial priors into the latent space of pre-trained generative networks， and carrying out optimization based on loss functions， local textures lost or damaged due to degradation are repaired to achieve a balance between realism and fidelity. Numerical experiments are conducted on three real degraded datasets， which outperform existing methods in terms of objective metrics and subjective quality， and the final ablation experiments validate the effectiveness of the degraded blind face restoration algorithm.

Abstract:Intuitionistic fuzzy similarity relations cause the similarity degree between objects in the intuitionistic fuzzy set too concentrated or the dissimilarity degree too high， leading to nreasonable classification results， and when constructing intuitionistic fuzzy similarity relation， the similarity degree and dissimilarity degree between objects are vulnerable to unimportant attributes. Therefore， a three-way decision model based on intuitionistic fuzzy similarity relation is proposed according to the intuitionistic fuzzy sets and the possibility theory. Firstly， the definitions of possibility measure and necessity measure are given. Combining with the Hausdorff measure， a distance formula is constructed and its properties are proved. The similarity degree and dissimilarity degree between objects in intuitionistic fuzzy sets are defined， and a new intuitionistic fuzzy similarity relationship is constructed.Then，the （λ₁，λ₂）-cut set under intuitionistic fuzzy similarity relation and the similar class under intuitionistic fuzzy （λ₁，λ₂）-cut set are defined， and the positive， negative and boundary fields of target set are further obtained. Finally， the rationality and effectiveness of the proposed model are verified through UCI data sets and examples.

Abstract:With the in-depth research on many-objective optimization problems， many-objective optimization problems with irregular Pareto frontiers pose challenges to existing methods due to their complex Pareto frontiers distribution. To address the above issues， a many-objective evolutionary algorithm based on the enhanced growing neural gas is proposed. This algorithm combines the learning characteristics of growing neural networks with the optimization characteristics of binary quality indicators to enhance the convergence pressure of the population at the irregular Pareto frontier. Firstly， an enhanced growing type of neural gas network is designed， which utilizes the topological information of the Pareto optimal frontier to guide the population to converge towards the Pareto optimal frontier direction. Then， a joint metric is proposed to comprehensively evaluate the convergence of individuals in conjunction with Pareto dominance information. Finally， an adaptive reference point based environment selection is proposed to enhance the diversity of the population in high-dimensional target space. To verify the performance of the proposed algorithm， 44 irregular many-objective optimization problems in the DTLZ and WFG benchmark problem sets are compared with five advanced many-objective evolutionary algorithms. Experimental results show that the overall performance of the proposed many-objective evolutionary algorithm based on enhanced growing neural gas is superior to the comparison algorithms.

Abstract:Under the criterion of maximum output signal-to-noise ratio （SNR）， the problem of difficult control of complex-valued coefficients in generalized eigenvalue （GEV） beamforming is encountered， and severe distortion of the output signal can be caused in complex acoustic environments. To address the issue of complex-valued coefficient estimation， a complex-valued coefficient estimation method based on minimum mean square error （MMSE） is proposed in this paper. By introducing a speech distortion weight factor （SDW）， the weight relationship between noise reduction and speech distortion is adjusted， thereby proposing a method for generalized eigenvalue robust beamforming based on SDW-MMSE. The power spectra of the target and noise signals are estimated using maximum likelihood method， and the main generalized eigenvectors are then determined. Furthermore， the complex-valued coefficients are estimated ， and the complex coefficients are combined with the principal generalized eigenvector to obtain the generalized eigenvalue robust beamforming filter vector based on SDW-MMSE. Through simulation experiments， it is demonstrated that the proposed beamforming method effectively eliminates coherent and incoherent noise， and exhibits robust performance with high output SNR and low speech distortion.

Abstract:With the rapid development of deep learning， the number of parameters and computational complexity of models have exploded， which pose challenges for deployment on mobile terminals. Model pruning has become the key to the implementation and application of deep learning models. At present， the pruning method based on regularization usually adopts L2 regularization combined with the importance standard based on the order of magnitude. It is an empirical method lacking theoretical basis， and its accuracy is difficult to guarantee. Inspired by the Proximal gradient method for solving sparse optimization problems， we propose a Prox-NAG optimization method that can directly generate sparse solutions on deep neural networks and a corresponding iterative pruning algorithm is designed. This method is based on L1 regularization and uses Nesterov momentum to solve the optimization problem. It overcomes the dependence of the original regularization pruning method on L2 regularization and order of magnitude standards， and is a natural extension of sparse optimization from traditional machine learning to deep learning. Pruning experiments are conducted on the ResNet series models on the CIFAR10 dataset， and the results show that the Prox-NAG pruning algorithm has improved its performance compared to the original pruning algorithm.

Abstract:Dialogue understanding aims to detect user intent given dialogue history. Due to the lack of domain knowledge， traditional dialogue understanding models fail to understand domain-specific entities. Knowledge-enhanced approaches are proposed to improve model performance with structured knowledge， where the knowledge is implicitly injected with knowledge embeddings. However， knowledge embeddings have to be updated with the update of the knowledge base， which brings extra costs. Besides， existing methods suffer from the knowledge noise and incorporate the context-irrelevant knowledge that changes the semantics of the utterance. To address the above issues， this paper proposes a multi-task learning dialogue understanding model with explicit knowledge injection（K-CAM）. K-CAM injects knowledge into the model using natural language knowledge without retraining the model for updated knowledge embeddings. A multi-task learning objective of joint intent detection， slot filling， and relevant knowledge recognition is further proposed to resist the knowledge noise problem. Extensive experimental results show that the proposed model K-CAM achieves a significant improvement of 4.87% and 2.09% in macro F₁ on the intent detection and slot filling tasks compared to other baselines.

Abstract:Trajectory reconstruction is a research field for trajectory restoration of low-sampling rate trajectory data. In recent years， in order to improve the accuracy of trajectory reconstruction， some work used deep learning models such as Seq2Seq to improve the efficiency and accuracy of trajectory recovery. However， most of the existing work ignores the long-distance dependencies between trajectory points， resulting in poor accuracy for trajectory reconstruction. Therefore， this paper proposes a trajectory recovery model， called ZTrajRec （Zero-based trajectory recovery） based on Transformer， which captures the long-distance dependency between trajectories through Transformer encoder， and uses the attention mechanism to take into account the similarity between current trajectory and historical trajectories to reconstruct the trajectory directly on the road network. Experimental results show that， on the real Beijing taxi dataset， ZTrajRec improves the recall rate by 3%—4%， compared to the results of the benchmark models. Finally， the result is visually analyzed to demonstrate its plausibility.

Abstract:In order to solve the problems of the insufficient feature information extraction and the difficulty in capturing local obvious feature information accurately in few-shot learning， a method combining class enhancement and multi-scale adaptation is proposed. Firstly， the class enhancement is performed on the image at the level of features， and rich semantic structures are encoded by associating each activation of the feature map with its neighborhood， thus making the extracted intra class features obvious and more conducive to the current classification task. Secondly， low-level representations of image features at different scales are extracted through multi-scale feature generation. Finally， the semantic correlation matrix on each scale is weighted and similarity elements are maximized to calculate the semantic similarity between the query image and each support set category image. After the fusion of multi-scale information， the target images are classified. In the 5-way 1-shot and 5-way 5-shot settings， the mean average precision （mAP） of this method on the miniImageNet dataset is 56.83% and 75.76% respectively， and it achieves 79.33% and 93.92%， 66.33% and 85.78% on the commonly used fine grained image dataset Standard Cars and CUB-200-2011 classification benchmarks， respectively， which are superior to the best results of the existing methods.

Abstract:To identify relevant extreme speech， a gender opposition speech recognition method of fusing multi-features and emoji sentiment lexicon is proposed. Firstly， BERT（Bidirectional encoder representation from transformer） is used to extract the character features of the input texts， and Word2Vec is used to extract the Wubi， Zhengma and Pinyin features of the input texts. Then， these features are fused and fed into the Bi-GRU（Bi-directional gated recurrent unit） network to obtain the deeper semantic information. Finally， the sentiment polarities are calculated with the full-connected layer and SoftMax function combining the emoji sentiment lexicon to determine whether the input texts are related gender opposition. Compared with the method without adding multi-features and emoji sentiment lexicon， the experiments on the self-collected Chinese gender opposition dataset show that the proposed model is improved on the F₁ value by 5.19%. In addition， the generalization of the proposed method is verified by experiments on the public Chinese sentiment analysis dataset Weibo_senti_100k.

Abstract:Structural brain network （SC） and functional brain network （FC） can reflect the changes in brain structure information caused by epilepsy from different perspectives. Currently， the fusion of two types of brain network information for auxiliary diagnosis of epilepsy has become one of the important studies in the field. However， common fusion models only fuse the information of the two types of brain networks at a single granularity， ignoring the multi-grained attribute of brain networks. This paper proposes an epilepsy identification method based on multi-modal multi-grained fusion network （MMFN）， which integrates the features of the multi-modal brain network from global and local granularities to take full advantage of multi-modal brain network information. Specifically， at the local granularity， two modules （i.e.， edge features fusion module and node features fusion module） are designed to reconstruct the feature maps of edge layer and node layer of two types of brain network， so that these two modes can learn features interactively. At the global granularity， a multimodal decomposition bilinear pooling module is designed to learn the joint representation of the two types of brain networks. Compared to current methods， experimental results show that the proposed method can improve the accuracy of epilepsy recognition significantly and assist doctors in the diagnosis of epilepsy.

Abstract:Mental arithmetic is a skill commonly used in daily life. It involves various cognitive processing processes that cause changes in brain activity， so research on its electroencephalogram （EEG） can help improve the level of research on cognitive tasks. Amplitude permutation entropy for global graph （APEGG） is proposed to apply to the study of EEG mental arithmetic， to make up that the traditional permutation entropy for graph （PEG） can not fully reflect changes of the neighboring nodes around brain network nodes， and overcome the problem of insensitive EEG signal amplitude. At first， the EEG brain network is constructed using the phase locking value （PLV）， the synchronization and correlation between multi-lead EEG signals are analyzed， and then the amplitude permutation entropy for global graph of the brain network at different frequency bands is calculated. Finally， support vector machine （SVM） is used for classification. EEG in public data sets is used for simulation， and the mental state of different frequency bands and resting state entropy scatterplot are analyzed， showing a larger difference. The classification results show better results compared with other algorithms.

Abstract:In the application environment of human activity recognition， it is still challenging to extract sufficiently reliable features from the original sensor data. The hybrid signal decomposition technology of discrete wavelet transform （DWT） and variational mode decomposition （VMD） is used to extract the salient feature vectors from the original sensor signals to identify various human activities. Using a variety of machine learning classification algorithms， such as K-nearest neighbor， random forest， LightGBM and XGBoost， the effectiveness of the proposed algorithm is tested on UCI-HAR and SCUT-NAA data sets. Experimental results show that by using the hybrid signal decomposition technology， the recognition accuracy of all classification algorithms has been improved， with the maximum classification accuracy of 98.91% for UCI-HAR dataset， which has improved by 1.79% compared to not joining the decomposition algorithm. The maximum classification accuracy of SCUT-NAA dataset reaches 95.52%， which has improved by 3.2%. In human activity recognition， through the use of DWT-VMD hybrid signal decomposition technique， more effective features can be extracted from the original signal and the recognition accuracy can be further improved， showing the certain practical value of the technique.

Abstract:This paper proposes a novel indoor positioning algorithm based on time sequences fusion in Pauta criterion-correlation coefficient convolutional neural networks （P-C-CNN）. The P-C-CNN approach integrates data points from different nodes and various time sequences， leveraging the interconnectedness of temporal and spatial data to enhance the accuracy and reliability of indoor positioning. Firstly， this method utilizes the Pauta criterion-correlation coefficient （P-C） algorithm to remove outliers in angle of arrival （AOA）-received signal strength （RSS） data， improving the quality of the training data. Secondly， the algorithm randomly selects data at intervals， reducing the training time of the model and effectively simulating the uncertainty of data selection in the online positioning phase， thus reducing overfitting of the model to the training data. Furthermore， the traditional single-frame information training method is unable to stably extract information features due to the mixture of noise. The proposed algorithm randomly selects multiple frames of fixed length from the continuously collected AOA-RSS data within time sequences fusion， and then employs convolutional neural networks （CNN） for feature extraction. This approach can avoid the issue of large error fluctuations commonly encountered in single-frame signal positioning. Finally， through extensive practical testing， this paper has validated the effectiveness of the proposed method. The experimental results demonstrate that in typical indoor environments， compared to fingerprint positioning algorithms that solely rely on RSS data or AOA information， the proposed algorithm achieves an improved classification accuracy from 91.6% to 96.4%， and the positioning accuracy is improved from 1.3 m to 0.3 m. Moreover， compared to the traditional model-based AOA-RSS joint positioning， this algorithm effectively addresses the influence of interference factors such as multipath effects observed in real-world measurements. The positioning accuracy is improved from 1.1 m to 0.3 m.

Abstract:Aiming at the problem of high computational overhead affecting the real-time localisation when paillier's algorithm is applied to indoor fingerprint privacy protection， this paper proposes a privacy-preserving algorithm for indoor fingerprinting positioning of mobile users to achieve trajectory anonymity and effectively improve the positioning performance. Since that the number of access points （APs） and reference points （RPs） involved in localization is the main factor affecting the time overhead of the encryption， the proposed algorithm divides the trajectory localization into continuous and discontinuous location localization. The number of APs and RPs involved in encryption is reduced by using the information of the before and after requests in continuous location localization， while the number of APs and RPs involved in encryption is reduced in discontinuous location localization. In continuous position localization， the number of APs and RPs involved in the encryption operation is reduced by using the information of before and after location requests； while in discontinuous positing localization， the coarse localization of users reduces the number of APs and RPs involved in the algorithm， thus improving the location efficiency. An optional scheme based on principal component analysis （PCA） is proposed to further improve the localization efficiency. Experimental results in a real-world environment show that the proposed algorithm can control the time required for a single positioning in both continuous and discontinuous positioning within 1 s. The positioning accuracy is improved by about 20% in continuous positioning， while the privacy protection has no effect on the positioning accuracy in discontinuous positioning. The overall performance of the localization algorithm is effectively improved.