Engineering and intelligent systems increasingly require decision-making under heterogeneous evidence. These sources include multi-source data, predictive models, competing objectives, feasibility constraints, and uncertainty. Evolutionary algorithms (EAs) are widely used in such settings, yet the literature is typically organised by algorithmic lineage, which obscures how and where information is fused within the evolutionary process. This survey reframes EAs as adaptive information fusion architectures. We introduce a fusion-centric taxonomy spanning data-level integration of heterogeneous observations, model/feature-level integration through surrogate and learning components, objective-level integration through multi- and many-objective formulations, constraint handling as feasibility-signal integration, and decision-level integration through ensembles, distributed (island) evolution, and multi-run aggregation. Building on this perspective, we synthesise major EA families according to their dominant fusion mechanisms and review fusion-driven applications across structural and mechanical design, energy and smart grids, robotics and control, communications, healthcare engineering, and neural architecture search. To complement this qualitative synthesis, we propose lightweight quantitative indicators—integration depth, integration diversity, and decision outcome entropy—to characterise algorithm–fusion–domain alignment and to identify recurring success and failure modes, including surrogate bias, over-aggregation, and fusion over-complexity. We conclude with practical design guidelines and discuss emerging directions, including federated evolutionary fusion and reliability considerations in learning- and large language model (LLM)-assisted fusion.

Generative Adversarial Networks (GANs) often exhibit unstable training dynamics and strong dependence on manually designed architectures, limiting scalability and robustness. This paper proposes HEAGAN, an efficient hybrid evolutionary neural architecture search framework for GANs. HEAGAN searches the generator architecture only, while the discriminator follows a fixed AutoGAN-style backbone throughout training and evaluation. HEAGAN couples a guided hybrid evolutionary search with a partial weight-sharing supernet to automate generator architecture discovery for unsupervised image synthesis. The search integrates genetic algorithms and particle swarm optimisation, while constrained weight sharing with dynamic node activation reduces weight coupling and mitigates multi-model forgetting during supernet optimisation. To improve training stability, an adaptive robust loss is incorporated, and a multi-objective formulation based on NSGA-II jointly optimises generation quality and distributional fidelity. Experiments on CIFAR-10, STL-10, CIFAR-100, and ImageNet32 show that HEAGAN consistently identifies competitive generator architectures with improved stability and strong transferability across datasets. On CIFAR-10, HEAGAN achieves an Inception Score of 8.99 +/- 0.11 and a Fr & eacute;chet Inception Distance of 8.20. Ablation studies further validate the individual contributions of the proposed components.

Traditional Neural Architecture Search (NAS) methods often struggle with hyperparameter tuning and are prone to local optima, which limits the diversity of the population and hampers the search for optimal architectures. This paper introduces DiffEvo-NAS, a novel approach that leverages a diffusion evolution algorithm, which interprets the evolution process as a denoising operation and the reverse evolution as diffusion. By incorporating directional denoising (similar to directional selection) and adding small noise (analogous to mutations), the algorithm continuously refines individuals while maintaining diversity, enabling effective exploration of the search space and avoiding local optima. The effectiveness of DiffEvo-NAS is demonstrated through experiments on the CIFAR-10 and CIFAR-100 datasets, where it achieves a test error rate of 2.55% and 16.15%, respectively. On ImageNet, it achieves competitive performance with a top-1 error rate of 24.61% and top-5 error rate of 7.44%. These results confirm that DiffEvo-NAS offers significant improvements in both exploration and exploitation, outperforming other evolutionary NAS methods in terms of population diversity and architecture performance. As a stepping stone in the field, DiffEvo-NAS lays the groundwork for further innovations in NAS, particularly in the integration of diffusion models and evolutionary computation.

Each LoRA checkpoint compactly stores task-specific updates in low-rank weight matrices, offering an efficient way to adapt large language models to new tasks and domains. In principle, these weights already encode what the adapter does and how well it performs. In this paper, we ask whether this information can be read directly from the weights, without running the base model or accessing training data. A key obstacle is that a single LoRA update can be factorized in infinitely many ways. Without resolving this ambiguity, models trained on the factors may fit the particular factorization rather than the underlying update. To this end, we propose , which maps each LoRA update to a provably canonical form via QR decomposition followed by SVD, so that all equivalent factorizations share the same representation. The resulting components are then tokenized and processed by a Transformer to produce a weight-space embedding. Across language and vision LoRA collections, W2T achieves strong results on attribute classification, performance prediction, and adapter retrieval, demonstrating that LoRA weights reliably indicate model behavior once factorization ambiguity is removed. Code is available at https://github.com/xiaolonghan2000/Weight2Token.

Designing effective neural architectures remains a central challenge in deep learning, and Neural Architecture Search (NAS) has become a popular tool for automating this process. However, many existing NAS approaches depend on hand-crafted architecture descriptors or shallow performance predictors, which fail to capture the structural complexity of candidate networks and often lead to unreliable search guidance. We introduce Graph Embedding Comparator with Isomorphic Multi-Comparison (GEC-IMC), an evolutionary NAS framework that learns architecture representations directly from their graph structure. A graph convolutional network encodes architectures into embeddings, while a contrastive learning strategy ensures that architectures with similar accuracy are mapped closer in the embedding space. On top of these embeddings, a comparator estimates the relative performance between two architectures, enabling more precise pairwise assessments during search. To further increase robustness, GEC-IMC incorporates an isomorphic multi-comparison mechanism, which evaluates multiple structurally equivalent variants of each architecture and aggregates their pairwise outcomes into a global score. This ranking score provides consistent feedback for evolutionary selection. Experiments on standard NAS benchmarks demonstrate that GEC-IMC achieves state-of-the-art performance with improved robustness over existing predictors. Ablation studies confirm the complementary roles of embedding learning and multi-comparison in enhancing search efficiency.

This paper addresses two persistent challenges in Evolutionary Neural Architecture Search (ENAS): high computational cost and premature convergence. We propose a novel method—One-Step Diffusion Distillation Evolutionary Algorithm with Score Implicit Matching for NAS (OSD2E-SIM-NAS)—which significantly reduces the complexity of the search process while maintaining performance. The key innovation is the distillation of the iterative diffusion-based evolutionary process into a single inference step via Score Implicit Matching (SIM), reducing computational complexity from O(T) to O(1). SIM leverages implicit score matching to enable dynamic variance modulation, naturally balancing exploration and exploitation—an essential characteristic inherited from diffusion models. To further mitigate premature convergence and improve search robustness, we introduce a diversity-preserving mechanism that combines simulated annealing, stochastic teacher sampling, and periodic injection of new individuals. We validate the portability and effectiveness of OSD2E-SIM-NAS across both image classification and generative tasks. The method achieves test error rates of 2.55% on CIFAR-10, 16.15% on CIFAR-100, and 24.61% on ImageNet. In generative tasks on CIFAR-100, it attains an Inception Score (IS) of 9.01± 0.07, a Fréchet Inception Distance (FID) of 15.06, and a remarkably low search cost of 0.39 GPU-days, demonstrating superior efficiency, diversity, and adaptability across tasks.

Automatic data classification is commonly performed using supervised learning algorithms that build models to classify new instances. However, labelled data are often scarce and expensive to obtain. Semi-supervised learning (SSL) offers a viable alternative by exploiting a few labelled and many unlabelled instances. Among various SSL approaches, graph-based methods are particularly appealing as they rely on neighbourhood structure rather than explicit parametric distributional assumptions. Nevertheless, their performance heavily depends on graph construction, which requires defining a suitable similarity (or distance) function. Since no single similarity function performs optimally across all datasets, this paper introduces two evolutionary approaches – GEAS and ST-GPAS – to automatically design similarity functions for graph-based SSL. Extensive experiments on 50 public datasets show that the proposed methods yield similarity functions leading to superior transductive learning performance compared with traditional distance-based measures and representative metric-learning algorithms. The improvements, measured in Micro-F1 and Macro-F1, are statistically significant under nonparametric tests. GEAS and ST-GPAS achieve closely matched performance, with small but consistent advantages depending on the supervision regime.

Neural network weights are typically viewed as the end product of training, while most deep learning research focuses on data, features, and architectures. However, recent advances show that the set of all possible weight values (weight space) itself contains rich structure: pretrained models form organized distributions, exhibit symmetries, and can be embedded, compared, or even generated. Understanding such structures has tremendous impact on how neural networks are analyzed and compared, and on how knowledge is transferred across models, beyond individual training instances. This emerging research direction, which we refer to as Weight Space Learning (WSL), treats neural weights as a meaningful domain for analysis and modeling. This survey provides the first unified taxonomy of WSL. We categorize existing methods into three core dimensions: Weight Space Understanding (WSU), which studies the geometry and symmetries of weights; Weight Space Representation (WSR), which learns embeddings over model weights; and Weight Space Generation (WSG), which synthesizes new weights through hypernetworks or generative models. We further show how these developments enable practical applications, including model retrieval, continual and federated learning, neural architecture search, and data-free reconstruction. By consolidating fragmented progress under a coherent framework, this survey highlights weight space as a learnable, structured domain with growing impact across model analysis, transferring, and weight generation. We release an accompanying resource at https://github.com/Zehong-Wang/Awesome-Weight-Space-Learning.

Accurate segmentation of low-contrast images plays a crucial role in computer-aided diagnosis and treatment, particularly for early lesion detection and clinical decision support. To address the limitations of existing approaches in boundary localisation and multi-scale context modelling, we propose a lightweight and efficient hybrid segmentation framework, referred to as SwinFuseNet. The proposed architecture combines the strengths of detection-based and Transformer-based models. Specifically, the Global Pyramid Attention Backbone Network integrates a shifted window-based Transformer mechanism to enhance the global representation of blurred lesions in low-contrast images. In the feature aggregation stage, two dedicated modules-Dynamic Zoom Fusion and Spatial Interaction Fusion-are introduced to adaptively integrate information from multiple layers, effectively refining local boundary representations and fine-grained structural features. Additionally, a lightweight attention subnetwork is employed to highlight salient regions while suppressing background noise, thereby improving overall segmentation precision. Experiments conducted on four publicly available low-contrast image segmentation datasets (ISIC 2018, PH 2, LUNA16, Kvasir-SEG and Brisc2025) demonstrate that the proposed method significantly outperforms existing models, including variants of U-Net and a recent detection-based segmentation framework. On the ISIC 2018 dataset, the proposed network achieves a Dice coefficient of 0.9873 and an Intersection-over-Union score of 0.9566, representing improvements of 4.48% and 6.53% respectively over the current state-of-the-art, showing a remarkable 78% and 60% improvement, respectively, toward perfection from the best alternative algorithm. These results confirm the effectiveness and practical relevance of the proposed method in the domain of low-contrast medical image segmentation.

Generative Adversarial Network (GAN) have become a dominant paradigm in Zero-Shot Learning (ZSL) for synthesizing features of unseen classes. However, the efficacy of these models relies heavily on their network architectures, the manual design of which is labor-intensive. While Neural Architecture Search (NAS) automates this process, existing approaches typically adopt a “divide-and-conquer” strategy that decouples the optimisation of the generator and discriminator to ensure stability. This separation neglects the intrinsic co-adaptive nature of adversarial training, often leading to mismatched architectures and sub-optimal performance. To address these limitations, this paper proposes a novel framework named Multi-Population Co-Evolutionary Generative Adversarial Network Architecture Search (MC-GANS). MC-GANS reformulates the search as a co-evolutionary task that evolves the generator and discriminator as a symbiotic system, integrating three key strategies: (1) a Wasserstein distance-based adaptive mechanism to dynamically balance adversarial training stability; (2) a diversity-guided multi-population strategy to prevent premature convergence; and (3) a channel attention mechanism is integrated into the search space to enhance feature selectivity. MC-GANS outperformed the state-of-the-art generative ZSL methods on two datasets and achieves competitive results on the other dataset. The maximum performance improvement in harmonic mean reaches up to 2.3%. Ablation studies and theoretical analysis further confirm the effectiveness of each component and the model’s ability to capture complex data distributions. Codes and models are available at https://github.com/Wang-Zhaoming/MC-GANS.

Foreign object detection is a critical task in power transmission line inspection, where large-scale power trans mission infrastructure networks must be monitored reliably using UAV-based or mobile visual systems operating under strict energy constraints. Existing deep learning-based detectors achieve high accuracy but incur substantial computational and power costs, limiting their suitability for long-duration or large-area inspection. Spiking neural networks (SNNs) offer an energy-efficient alternative; however, their application to complex object detection tasks has been constrained by architectural limitations that prevent the incorporation of advanced feature fusion mod ules commonly used in modern detectors. To address this limitation, SpikeC3-YOLO is proposed. The framework introduces the SpikeC3 module, which supports C3-level feature fusion while preserving full spike compatibility. A unified pre-convolution spiking strategy is employed to preserve event-driven computation while supporting complex feature interactions. Comprehensive experiments on a real transmission line inspection dataset show that the proposed method achieves 57.30% mAP@0.5:0.95, comparable to the ANN-based YOLOv8 (55.80%), while reducing theoretical energy consumption by 48.3% relative to the baseline Spike-YOLO. Furthermore, an ultra-efficient variant maintains competitive detection performance while consuming only 13.8% of the baseline energy. These results demonstrate that sophisticated detection architectures can be realised within spiking neural networks without sacrificing energy efficiency, enabling practical, energy-aware visual inspection solutions for large-scale power transmission lines.

Detecting small objects reliably is particularly difficult in modern neural architectures, where scale imbalance, background clutter, and high object density frequently degrade feature quality and prediction accuracy. To address these challenges, we propose YOLO-Super Resolution and Attention (YOLO-SRA), a multi-scale neural architecture enhanced with attention and super-resolution. The architecture introduces High-Resolution Feature Enhancement (HRFE) to better represent small objects without incurring high computational cost, a Grouped Multi-Scale Split Attention (GMSA) mechanism to efficiently extract features from densely distributed objects, and Weighted Fine-Grained Cross-Scale Fusion (WFCF) network for adaptive multi-scale feature integration with Unmanned Aerial Vehicle (UAV)-specific adjustments. The Spatial-Attentive Non-Maximum Suppression (SA-NMS) strategy is further employed to reduce missed detections in overlapping regions. Extensive experiments on the VisDrone dataset demonstrate that YOLO-SRA outperforms the baseline YOLOv11, achieving 12.5% and 11.3% increase in mAP 50 and mAP 50 : 95 , respectively, while reducing the parameter count by 6.0M. These results highlight the effectiveness of the proposed approach as a methodological contribution to neural systems for small object detection in aerial image.

Crowd counting aims to estimate the number of individuals in images, and the use of multimodal data has been shown to significantly enhance counting accuracy. However, such approaches are highly sensitive to the loss or corruption of data from any single modality, leading to severe performance degradation. To address this limitation, a new problem setting-Modality-Reconfigurable Crowd Counting-is introduced, in which a model is required to maintain robust performance even when one of the input modalities (e.g., RGB or thermal) is perturbed or entirely unavailable. Modality reconfigurability is achieved through effective cross-modal information transfer, enabled by a Feature Patches Generator that leverages Margin Ranking Loss across multiple network layers to align and transfer discriminative features between modalities. Additionally, a Negative Knowledge Transfer Prevention module is incorporated to suppress misleading or detrimental cross-modal signals. State-of-the-art performance is demonstrated on RGB-T crowd counting benchmarks, with consistent accuracy maintained under both complete and degraded modality conditions.

Neural architecture search (NAS) has become an essential approach for automating neural network design and improving performance across a range of tasks. Nevertheless, applying NAS to non-Euclidean data, such as graphs, remains challenging due to the high computational cost of evaluating candidate architectures. This paper introduces GraphCETF, a novel evolutionary graph neural architecture search (GNAS) framework that integrates a training-free evaluation strategy with surrogate-assisted methods to accelerate the search process while maintaining high performance. By exploiting training-free evaluation techniques, GraphCETF markedly reduces the computational cost of assessing neural architectures, measured in GPU hours and runtime, and enables more effective exploration of the search space in terms of faster convergence during evolutionary search. We evaluate GraphCETF on six benchmark graph datasets, demonstrating its ability to discover high-quality architectures with substantially lower computational overhead compared with conventional approaches. Experimental results show that GraphCETF achieves competitive performance while significantly reducing search costs, making it a promising solution for efficient GNAS.

Simultaneous Localization and Mapping (SLAM) systems typically rely on a prior map constructed during an initial deployment. In real-world environments, however, structural and semantic changes gradually invalidate this map, leading to degraded localization accuracy and, in severe cases, localization failure. This limitation poses a major obstacle to the long-term deployment of mobile robots in dynamic environments. This paper proposes a lifelong mapping framework with multi-view projection fusion (LLMF) that enables efficient map maintenance while preserving a consistent global coordinate system. The framework introduces two key design components. First, a multi-view point cloud projection alignment strategy based on Bird's-Eye View (BEV) and frontal view (FV) projections is employed to align point cloud maps acquired at different times without re-labeling previously defined operational points. Second, an image-based change detection and map update mechanism is developed, transforming computationally expensive 3D point cloud comparisons into efficient 2D image processing operations. The proposed framework is evaluated through qualitative experiments on the open-source MulRan dataset and quantitative long-term experiments conducted over more than nine months in a real farm environment. Experimental results demonstrate that LLMF maintains localization accuracy while significantly reducing the computational cost of change detection, lowering processing time from several hours to a few minutes. These results indicate that the proposed framework provides a practical and scalable engineering solution for long-term robot operation in changing environments.

Fine-tuning large-scale Transformer-based models is computationally expensive due to the enormous parameter space. Low-Rank Adaptation (LoRA) substantially reduces the number of trainable parameters while maintaining performance; however, identifying the optimal LoRA configuration - such as rank r, scaling factor [Formula: see text], and insertion positions - remains challenging. To address this issue, we propose a zero-shot proxy metric, termed Gradient Projection Score (GPS), which enables rapid evaluation of candidate configurations using only a few forward and backward passes. Building upon this metric, we further introduce EvoLoRA, a zero-shot evolutionary architecture search method that jointly optimizes three objectives: performance proxy, evaluation stability, and trainable parameter size. EvoLoRA automatically discovers effective LoRA configurations across different models and datasets. Experimental results demonstrate that GPS is strongly correlated with final model performance; moreover, on tasks such as image classification and object detection, EvoLoRA markedly reduces search and training costs while generally outperforming other fine-tuning methods and manually designed LoRA configurations.

The Pareto set (PS) of a continuous multi-objective optimization problem exhibit a distribution along a low-dimensional manifold structure. This regularity property significantly contributes to generating high-quality offspring in large-scale multi-objective evolutionary algorithms (LSMOEAs). However, conventional regularity model-based algorithms face several challenges when dealing with large-scale multi-objective optimization problems (LSMOPs), including high computational costs for modeling, difficulty in capturing the true PS structure, and neglecting individual directional information. To address these challenges, we propose a dual-information offspring reproduction strategy that considers both the distribution information of the population and the directional information of the outstanding individuals. Specifically, this strategy comprises a sampling approach based on an augmented regularity model specifically designed for LSMOPs. Leveraging this model, we explore and exploit the decision space to sample a promising set of solutions. Additionally, the strategy also involves a search method based on competitive learning among individuals. By assigning a positive evolutionary direction to losing solutions, we update the losing solutions to generate high-quality offspring. We continuously refine the proposed regularity model to approximate the true PS more closely. In extensive experiments on large-scale multi-objective benchmark functions, we compare our algorithm with eight state-of-the-art algorithms. The results demonstrate that our approach excels in handling LSMOPs.

The detection of drone aerial images has become a research hotspot due to the extensive application of drones in various fields. However, the presence of a large number of small objects and complex scenes in the images poses severe challenges to the detection task. As an emerging technology, dynamic neural networks, with their input adaptive adjustment mechanism, provide important theoretical and technical support for solving the problem of detecting small objects in complex scenes of drone aerial images. In this work, we propose Dynamic You Only Look Once Object Detector (Dynamic-YOLO). Firstly, we improve You only look once (YOLO) v9. By leveraging the Convolutional Block Attention Module (CBAM) attention mechanism and Space-to-Depth Convolution (SPD-Conv), we optimise its backbone network to enhance the model’s feature extraction ability. Then, we adjust the structure of its neck network, delete the detection head P5 for large objects, and add the detection head P2 for tiny objects. After that, we combine a dynamic router with the improved YOLOv9 to form a dynamic detector, achieving adaptive processing of images. The experimental results on the Vision Meets Drone (VisDrone) benchmark dataset have demonstrated the excellent performance and flexibility of Dynamic-YOLO. When the Floating Point Operations (FLOPs) increase to 101.5 G, the highest mean Average Precision at Intersection over Union 0.5:0.95 (mAP@0.5:0.95) reaches 32.3% and mAP@0.5 hits 51%, representing a 4.0% and 10% improvement over YOLOv9-M. In addition, by dynamically adjusting the parameters of the router, the computational resource consumption and detection accuracy of the model can be flexibly adjusted, effectively achieving an optimised balance between detection performance and resource costs.

Designing effective architectures for Generative Adversarial Networks (GANs) remains a challenging task due to training instability and the complexity of balancing generator and discriminator performance. While Neural Architecture Search (NAS) has shown promise in automating architecture design, existing NAS-GAN approaches often suffer from limited design flexibility and high computational demands. This paper introduces a gradient-based NAS framework, termed Differentiable Architecture Search for GANs with Rollback Mechanism (RASGAN), aimed at addressing these limitations. RASGAN incorporates a hyperparameter rollback to indirectly optimise evaluation metrics such as the Inception Score (IS) and Fr & eacute;chet Inception Distance (FID), leading to higher-quality generative models. Moreover, the search space integrates lightweight convolutional operations to reduce computational and storage overhead without compromising performance. On unconditional image generation tasks, the proposed method achieves competitive results: on CIFAR-10, RASGAN attains IS = 8.98 and FID = 10.31; on STL-10, IS = 10.55 and FID = 22.37. Compared to existing NAS-GAN methods, the architectures discovered by RASGAN are not only more effective but also significantly more efficient, exhibiting reduced parameter size while maintaining strong generative performance.

Evolutionary neural architecture search (ENAS) treats neural network design as an optimisation problem and addresses it via evolutionary computation. Despite being flexible and enabling automated design, ENAS typically suffers from high computational costs due to the need to train a network at each fitness evaluation. Surrogate-assisted ENAS methods mitigate the severity of this challenge by replacing the computationally expensive fitness function with an approximate computationally cheap fitness function for some fitness evaluations of the run. Currently, a major research challenge in the field is the smooth integration of such surrogate models (and, often, data collection mechanisms) within ENAS frameworks. This paper puts forth a simple yet effective way to address this challenge. During the initial stage of the optimisation, the proposed algorithm, score predictor-assisted ENAS (SPNAS), evolves a small population of candidate architectures using ground truth fitness, i.e., the testing error rate of the network following its training. The data collected in this stage are then used to train a multi-layer perceptron network that builds an alternative fitness function. Unlike algorithms in previous studies, this novel alternative fitness does not approximate the error rate but is designed to preserve its order relation over populations of candidate architectures. Thus, this approach naturally allows for a computationally cheap population ranking. Most of the evolution is then carried out with the surrogate (i.e., alternative) fitness on a large population without retraining the surrogate model or calculating the ground truth fitness. Experiments conducted on the EvoXBench platform show that on its seven search spaces, SPNAS achieves excellent results in terms of error rate despite the modest use of ground-truth fitness calls.