How readers use multiple sources of information to determine where to move their eyes during reading is a central question in eye movement control. In Chinese reading, both preview and contextual information have been shown to modulate saccade target selection. However, it remains unclear how contextual constraint and parafoveal preview jointly influence saccade targeting. The present study examined the relationship between contextual constraint and parafoveal preview in saccade target selection during Chinese reading. Participants' eye movements were recorded while they read sentences containing three-character target words embedded in either high- or low-constraint contexts. Using the boundary paradigm, we manipulated the preview validity of the target words: the preview was either identical to the target word (valid preview) or consisted of three pseudo-characters (invalid preview). The results showed a significant interaction between contextual constraint and preview validity on initial landing positions within the target word region and on saccade lengths entering the target word. Specifically, contextual constraint affected saccade targeting under valid preview conditions, whereas this effect was eliminated under invalid preview conditions. These findings indicate that saccade target selection during Chinese reading reflects a dynamically interactive process that integrates both preview (bottom-up) and contextual (top-down) information. The results provide empirical evidence that refines cognitive models of eye movement control in Chinese reading.

Influence maximization (IM) is a fundamental problem in complex network analysis, with a wide range of real-world applications. To date, existing approaches to influential node identification in IM have predominantly relied on standard graphs, failing to capture higher-order intrinsic interactions embedded in many real-world systems. Hypergraphs can be employed to better capture higher-order interactions. However, using hypergraphs may lead to an excessively large search space and increased complexity in modeling cascading dynamics, making it challenging to accurately identify influential nodes. Therefore, in this study, we propose a new hypergraph-modeled IM method, based on the Discrete Particle Swarm Optimization algorithm and the threshold model. In the proposed method, a particle (i.e., a candidate solution) represents the selection information of seed nodes, and the fitness function is designed to accurately and efficiently evaluate the influence of seed nodes via a two-layer local influence approximation. We also propose a degree-based initialization strategy to improve the quality of initial solutions and develop rules for updating particles' velocity and position, incorporated with a local search to drive particles toward better solutions. Experimental results demonstrate that the proposed method outperforms baseline methods on both synthetic and real-world hypergraphs. In addition, ablation studies validate the effectiveness of both the local search and the initialization strategies.

Defect recognition is an essential aspect of intelligent manufacturing, but it is a challenging task with noise and unpredictable uncertainties, where convolutional neural networks (CNNs) struggle to achieve good performance. The fuzzy neural network (FNN) emerges as a promising approach to handle uncertainties. However, conventional methods for designing FNNs are tedious and error-prone. A solution is to automatically search for efficient FNN, which can be achieved by neural architecture search (NAS). To achieve NAS for FNN, we propose an efficient classifier-assisted evolutionary multiobjective FNN framework for defect recognition. Considering the characteristics of FNN (e.g., difficult to train and prone to overfitting), we first construct the architecture search as a constrained multiobjective optimization problem, the network accuracy and the architecture size are two conflicting objectives, and the constraint is used to filter out low-quality architectures. Then, we design the search space to incorporate the fuzzy module and develop the corresponding architectural representation and evolutionary operators. Furthermore, the complex regression task of performance evaluation is transformed into a classification task, and a classifier is designed to simplify the performance evaluation process. Massive experiments on four defect recognition datasets (i.e., ELPV, CODEBRIM, MIXEDWM38, and WM-811K) show that the architectures can effectively handle inherent uncertainties from datasets. Our method achieves 94.77% accuracy on ELPV, 81.82% accuracy on CODEBRIM, 98.99% accuracy on MIXEDWM38, and 98.22% accuracy on WM-811K, respectively.

To address ambiguous discrimination of intermediate flavor styles and low overall accuracy in existing flue-cured tobacco flavor style classification methods, we propose an innovative spectral modeling approach integrating data-driven and prior knowledge-guided strategies. Specifically, a pre-trained tobacco spectral model and limited labeled spectral data are used as prior knowledge, enabling pseudo-label annotation via transfer learning and self-training. Consistency training and an improved prediction confidence metric are further introduced to enhance pseudo-label reliability. The proposed discrimination model captures multi-level, multi-particle internal vibration characteristics of organic molecules in tobacco. Experimental results demonstrate that the proposed method reaches an accuracy of 94.0% in identifying medium-flavor flue-cured tobacco within the constructed dataset. It delivers better classification performance compared with several mainstream algorithms adopted in this field. This approach alleviates the difficulty in distinguishing medium-flavor tobacco samples and improves the overall classification accuracy for the three typical flavor styles of flue-cured tobacco.

Federated learning (FL) enables collaborative training across distributed edge devices, but deploying lightweight models in dynamic edge environments remains challenging. Existing methods typically require retraining whenever device resource constraints change, resulting in excessive computational overhead. We propose DFMPQ, a dynamic federated mixed-precision quantization framework that enables retraining-free deployment at the edge. DFMPQ trains a weight-sharing mixed-precision supernet via FL, which jointly represents diverse bit-width configurations. After training, resource-aware quantized subnets can be derived on demand to satisfy heterogeneous and time-varying resource constraints without additional optimization. Optimizing such a supernet in federated settings is difficult due to optimization interference among heterogeneous bit-widths and the coupling of quantization noise with non-IID data. DFMPQ addresses these issues through semantic-aware training and aggregation mechanisms that stabilize supernet optimization. In addition, a sensitivity-guided greedy search strategy is adopted to efficiently identify suitable quantization configurations under given resource budgets. Extensive experiments on multiple datasets and network architectures demonstrate that DFMPQ achieves competitive accuracy with significantly reduced computational cost, enabling efficient deployment for dynamic edge computing environments.

Accurate energy consumption prediction in Laboratory Stirred Millss is critical for process optimization but remains challenging due to the limitations of existing mechanistic and empirical models in capturing nonlinear, multi-parameter interactions. This study proposes a Backpropagation (BP) neural network model to predict the specific energy consumption of a WTM-5 Laboratory Stirred Mills grinding Panzhihua vanadium-titanium magnetite. Based on experimental data covering variations in residence time, rotational speed, filling rate, and pulp concentration, the model effectively characterizes the nonlinear relationships between operational parameters and energy use. The BP neural network demonstrated significantly higher predictive accuracy compared to traditional methods. Furthermore, Spearman correlation analysis identified pulp concentration as the most critical factor influencing specific energy consumption. These findings establish the BP model as a robust tool for real-time prediction and parameter optimization in fine grinding processes, laying a foundation for advanced deep learning applications in mineral processing.

Data-driven models are increasingly used to monitor complex industrial systems under varying operating conditions; however, their practical adoption is often constrained by limited labeled data and a lack of interpretability from an end-user perspective, where low-rank representations and context-irrelevant diagnosis are preferred. This study proposes a generative modeling framework for aero-engine fault clustering that emphasizes user-oriented interpretability, bridging data-driven learning with engineering diagnostic practice. The model adds a low-rank discriminative regularization to a mixture-based conditional variational autoencoder, enabling context-free clustering of multiple fault types. First, the learned latent structure yields compact and separable representations for multiple faults. Second, the associated covariance patterns are mapped back to sensor correlations within a reference context, aligning with how maintenance engineers recognize fault symptoms. Experimental results on both N-CMAPSS and real-world aero-engine datasets verify that the proposed approach yields superior clustering performance and reveals interpretable degradation patterns, supporting reliable diagnosis and informed decision-making in complex system health management.

With the rapid advancement of large language models (LLMs) in both academia and industry, their growing size and complexity have introduced significant challenges in terms of computational cost and deployment efficiency. To address these issues, a wide range of inference optimization techniques—including but not limited to model compression—have been proposed to accelerate LLM inference while preserving model performance. This survey provides a comprehensive overview of LLM inference acceleration strategies, analyzing them from multiple perspectives, including foundational principles, algorithmic techniques, real-world applications, and open research challenges. We begin by introducing core concepts underlying inference optimization and propose a new taxonomy that categorizes existing approaches, including quantization, pruning, distillation, efficient architectures, compilation, and hardware-aware methods. Following the lifecycle of LLM development and deployment, we examine how these techniques interact with model training, fine-tuning, and serving. Furthermore, we highlight key applications of efficient LLMs and discuss emerging trends and unresolved issues in the field. By synthesizing recent advances, this survey aims to provide actionable insights and practical guidance for researchers and practitioners working with scalable and efficient LLM systems.

Neural architecture search (NAS) faces a significant challenge due to the high computational cost of architecture evaluation. Zero-shot (ZS) proxies have been proposed as an efficient alternative to reduce evaluation overhead; however, their limited rank correlation often leads to sub-optimal architecture selection. To alleviate this issue, recent studies attempt to combine multiple ZS proxies, yet most existing approaches rely on heuristic aggregation guided by domain knowledge and remain restricted to a single evaluation scenario, which limits both effectiveness and interpretability. In this paper, we propose a novel perspective that treats ZS proxies as architectural features and formulates their combination as an evolutionary symbolic learning problem. Specifically, we leverage a genetic programming based symbolic learning framework to automatically evolve interpretable proxy combinations, enabling flexible exploration of symbolic structures without manual design. The evolved symbolic models capture complex and non-linear relationships between ZS proxies and architectural performance while maintaining strong interpretability. Furthermore, we systematically model three evaluation scenarios, including pointwise, pairwise, and listwise settings, and design scenario-specific fitness functions to guide evolutionary search toward different ranking objectives. Extensive experiments across 19 tasks using 13 representative ZS proxies are conducted to evaluate the proposed framework. The results demonstrate that the evolved symbolic models consistently achieve improved rank correlation and provide valuable insights into proxy interaction patterns under different evaluation scenarios, highlighting the effectiveness and generalizability of the proposed approach for NAS.

In evolutionary neural architecture search (ENAS), the accuracy predictors (i.e., regression models) have been successfully applied to save computational costs for the evaluation of network architectures. However, the accuracy of these predictors is largely limited by the small amount of evaluated architectures that may be difficult to obtain. Such accuracy predictors with prediction bias often lead to an inaccurate ranking, misleading the selection of ENAS. To alleviate the above limitations, we design an efficient and novel listwise ranking predictor (LRP) for ENAS to directly predict the ranking of each architecture instead of the numerical accuracy value of each architecture. Specifically, the training data is constructed by the proposed random encoding-combination (REC) strategy, which can generate substantial training data using the small number of evaluated architectures (data level). These specially constructed training data are used to train LRP, which can convert the complex regression task into a ranking task to reduce ranking bias (model level). The proposed NAS method is compared with state-of-the-art NAS methods on widely-used benchmark datasets and practical application. Experimental results demonstrate that LRP can alleviate the ranking disorder problem and outperform others in terms of both effectiveness and efficiency.

A game-theoretic risk assessment framework is proposed for Industrial Control Systems (ICS), integrating Bayesian Attack Graphs (BAG) with attacker-defender utility modeling. The approach constructs a game based on enhanced vulnerability-based risk quantification, deriving Nash equilibrium strategies to estimate attack and defense probabilities. These probabilities are incorporated into BAG to enable dynamic inference of system attack likelihoods and associated risks based on observed evidence. Experimental results demonstrate improved assessment accuracy over traditional methods. The framework supports informed security decision-making in ICS by enhancing situational awareness.

The great success of graph neural networks (GNNs) in graph-structured data tasks benefits from the powerful structure learning abilities of their architectures. For complex datasets, too deep GNNs can suffer from over-smoothing problem, leading to degradation of prediction performance. Recently, a graph fusion architecture designed by domain experts mitigates the over-smoothing problem. However, this design paradigm is labor and computation-intensive. Drawing on the idea of neural architecture search, this paper proposes evolutionary graph fusion architecture search (EGFAS), which can automatically find graph fusion architecture with the capability to solve the over-smoothing problem. Specifically, we design a graph fusion architecture search space, which includes multiple aggregation functions and fusion operations. An efficient encoding method is used to represent candidate architectures. Furthermore, we propose a surrogate model to solve the expensive overhead of performance evaluation. Experiments on five benchmark datasets are carried out to confirm the superiority of the proposed algorithm. It is shown that the proposed method leads to better results comparing to the state-of-the-art algorithms in terms of search effectiveness and adaptability to different datasets.

In neural architecture search (NAS), the relativistic predictor has recently emerged as an attractive technique to solve ranking issue for performance evaluation by predicting the relativistic ranking of architecture pair rather than the absolute performance of an architecture. However, it suffers from a significant cold-start issue, requiring a large amount of evaluated architectures to train an effective predictor on new datasets. In this paper, we propose a transferable relativistic predictor (TRP). Specifically, we construct a proxy dataset using the transferable cheaper-to-obtain performance estimation to softly label the rank between architectural pairs. The soft label with a smooth and easy-to-optimize loss function facilitates the learning of expressive and generalizable representations on the proxy dataset. Furthermore, we construct Chebyshev interpolation for correlation curve to adaptively determine the number of evaluated architectures required on each dataset. Extensive experimental results in different search spaces show the superior performance of TRP compared with state-of-the-art predictors. TRP requires only 54 and 73 evaluated architectures for a warm start on the CIFAR-10 and CIFAR-100 under the DARTS search space.

Medical information extraction (IE) is an essential aspect of electronic health records (EHRs), but it is a challenging task that converts plain text into structured knowledge, where domain models struggle to achieve performance. Recently, large language models (LLMs), which have demonstrated remarkable capabilities in text understanding and generation, have emerged as a promising method for handling natural language texts. However, LLMs are too dependent on elaborate prompts, resulting in extensive expert knowledge and manual prompt templates needed. In this work, we propose a novel method for the automatic prompt design, called Particle Swarm Optimization-based Prompt using a Large language model (PSOPL). As an efficient method for medical information extraction from EHRs, PSOPL can allow particle swarm optimization (PSO) to automate design prompts by leveraging LLM’s ability to generate coherent text token-by-token. Specifically, starting with a small number of initial prompts, evolutionary operators in PSOPL guide the LLM to generate new candidate prompts iteratively, and the PSOPL evaluates population fitness to retain the optimal prompts. In this way, PSOPL can achieve prompt evolution without model training and reduce the human effort and requirement for domain knowledge. We conducted experiments for open-source LLMs (e.g., Alpaca-7B, GPT-J-6B) and closed-source LLM (e.g., GLM-4), on public medical datasets (e.g., CMeEE, CMeIE, CHIP-CDEE) covering information extraction tasks (e.g., named Entity recognition, relation extraction, event extraction) to verify the method’s generalizability. The experimental results demonstrate the potential of using PSO-based LLMs to design prompts automatically, allowing for the swift extraction of important information about patients in the EHRs.

Benefiting from the principles of information aggregation and multi-layer stacking, graph neural networks (GNNs) have achieved remarkable success in graph-structured data tasks. However, deeper GNNs often suffer from over-smoothing, which weakens node-level discrimination. To address this, skip connections and feature fusion techniques have been introduced, but the manual design of these strategies typically require significant computational resources and expert knowledge. Inspired by neural architecture search (NAS), this paper proposes EGFAS-BD, an evolutionary framework for graph fusion architecture search with a binary-to-decimal encoding-based performance predictor. This is an evolutionary algorithm (EA)-based framework specifically designed for graph fusion scenarios. We first construct an expanded search space integrating diverse aggregation and fusion operations to enhance architectural expressiveness. A compact binary-to-decimal encoding scheme is then used to improve information density, reducing storage overhead and simplifying predictor training. Additionally, we introduce an active filling strategy based on core-set selection, which leverages population-level distributions to enhance sampling efficiency and prediction accuracy. Experiments on eight benchmark datasets and three real-world medical graph datasets demonstrate that EGFAS-BD outperforms state-of-the-art NAS methods in both search efficiency and generalization performance. Ablation studies further verify that the proposed encoding and prediction strategies significantly accelerate convergence and reduce computational cost.

Biomedical image segmentation plays a critical role in clinical applications such as disease diagnosis and surgical planning. While deep learning, especially U-Net and its variants, have achieved impressive performance in this domain, their success heavily relies on manual architectural design, which limits scalability and generalizability. Neural architecture search (NAS) offers a promising solution by automating model design, but existing NAS-based methods still suffer from high computational costs and limited architectural diversity. In this work, we propose a novel NAS framework, termed ranking-aware predictorassisted architecture design (RPA2D), for efficient and accurate biomedical image segmentation. RPA2D introduces a lightweight and expressive search space comprising diverse convolution and pooling operations tailored for medical images. Furthermore, we design a contrastive learning-based ranking-aware performance predictor that enables accurate architecture ranking under limited supervision. Extensive experiments on public biomedical datasets demonstrate that RPA2D consistently outperforms both handcrafted and NAS-based baselines in segmentation accuracy while significantly reducing search overhead.

Deep learning has emerged in many practical applications, such as vascular segmentation, fault diagnosis, and human detection. More recently, convolutional neural networks (CNNs), representative techniques of deep learning, have been used to solve emotion recognition. However, the current design of CNNs for emotion recognition is highly dependent on domain knowledge and needs a large amount of trial and error. For this reason, an evolutionary CNN framework is developed to automatically find network architecture for EEG-based emotion recognition. Specifically, we firstly design a search space based on three advanced network basic units. Based on this, a flexible variable-length encoding is proposed and the corresponding reproduction operators (i.e., crossover and mutation) are designed. To reduce search overhead, this paper proposes an acceleration strategy based on the similarity metric for population memory. A series of experimental results show that the architecture by ECNN-ER method achieves higher accuracy (96.47%) compared to the state-of-the-art results (i.e., DARTS-PV) on the DEAP dataset, as well as competitive results (accuracy = 95.78%) on the DREAMER dataset.

Deep learning has emerged in many practical applications, such as vascular segmentation, emotion recognition, and target detection. Moreover, convolutional neural networks (CNNs), representative techniques of deep learning, have been used to solve face recognition. However, the current design of CNNs for face recognition is highly dependent on domain knowledge and needs a large amount of trial and error. In this paper, we propose a new method based on the brain storm optimization algorithm to tackle the network structure and training parameters selection problem of CNN for face detection. In our method, an efficient mixed-length encoding strategy is designed to represent the CNN network information and the model training parameters. Specifically, convolutional layers, pooling layers, and fully connected layers can be designed automatically, and then an optimal set of selection results can be obtained. A series of experimental results show that the architecture by our method achieves higher accuracy (98.7%) compared to the state-of-the-art results on the FDDB dataset, as well as competitive results (accuracy = 94.3%, 92.5%, and 84.9%) on the WIDER FACE dataset.

In multiobjective evolutionary neural architecture search (NAS), existing predictor-based methods commonly suffer from the rank disorder issue that a candidate high-performance architecture may have a poor ranking compared with the worse architecture in terms of the trained predictor. To alleviate the above issue, we aim to train a Pareto-wise end-to-end ranking classifier to simplify the architecture search process by transforming the complex multiobjective NAS task into a simple classification task. To this end, a classifier-based Pareto evolution approach is proposed, where an online classifier is trained to directly predict the dominance relationship between the candidate and reference architectures. Besides, an adaptive clustering method is designed to select reference architectures for the classifier, and an $\alpha $ -domination-assisted approach is developed to address the imbalance issue of positive and negative samples. The proposed approach is compared with a number of state-of-the-art NAS methods on widely used test datasets, and computation results show that the proposed approach is able to alleviate the rank disorder issue and outperforms other methods. Especially, the proposed method is able to find a set of promising network architectures with different model sizes ranging from 2M to 5M under diverse objectives and constraints.

Defect recognition is an important task in intelligent manufacturing. Due to the subjectivity of human annotation, the collected defect data usually contains a lot of noise and unpredictable uncertainties, which have a great negative influence on defect recognition. It is a significant challenge to discover an effective defect recognition model with satisfactory uncertainty processing ability. A natural way is to automatically search for an efficient deep model, which can be realized by neural architecture search (NAS). To achieve this, we propose an efficient fuzzy NAS framework for defect recognition, where the searched architecture can effectively handle uncertain information from the given datasets. Specifically, we first design a fuzzy search space and the related encoding strategy for fuzzy NAS. Then, we propose a comparator-based evolutionary search approach, where an online end-to-end comparator is learned to directly determine the selection of candidate architectures from the evolutionary population. The comparator works in an end-to-end way and it transforms the complex ranking problem of evaluating architectures into a simple classification task, which overcomes the rank disorder issue suffered from traditional performance predictors. A series of experimental results demonstrate that the architecture with fewer #Params (1.22 M) search by fuzzy neural architecture search framework for defect recognition method achieves higher accuracy (92.26%) compared to the state-of-the-art results (i.e., DARTS-PV) on the ELPV dataset, as well as competitive results (accuracy = 76.4%, #Params = 1.04 M) on the CODEBRIM dataset. Experimental results show the effectiveness and efficiency of our proposed method in handling uncertain problems.