To address the limitations of the Snow Ablation Optimizer (SAO) algorithm in solving complex multimodal and composite-structure problems, including slow convergence and susceptibility to local optima, this paper proposes the Hybrid Snow Ablation Optimizer (HGSAO) integrated with a dynamic multi-layer velocity guidance mechanism that integrates a dynamic hierarchical cooperative search mechanism and an adaptive local enhancement strategy. The core contribution of this work lies in the systematic deconstruction and reorganization of the original algorithmic framework through three progressive innovations: First, a dynamic hierarchical mechanism based on individual performance metrics and evolutionary trends is developed, which adaptively partitions the population into three functional subgroups (elite, potential, and ordinary), establishing a structural foundation for subsequent differentiated strategy implementation. Second, a hybrid update model for velocity and position evolution, incorporating historical information and population-state dynamics, is constructed, enabling targeted guided search operations for distinct subgroups to coordinate global exploration and local exploitation effectively. Third, an adaptive local enhancement strategy with dual perception capabilities for iteration progression and convergence status is embedded, which conducts multi-strategy collaborative exploitation in the neighborhood of high-quality solutions during the late algorithmic stages to improve convergence accuracy and stability. A comprehensive experimental evaluation on the IEEE CEC 2022 benchmark suite demonstrates that HGSAO achieves superior optimization accuracy, convergence speed, and robustness across unimodal, multimodal, hybrid, and composite functions, with the Wilcoxon rank-sum test confirming its statistically significant superiority. Furthermore, feature selection experiments conducted across 12 public datasets confirm HGSAO’s effectiveness in high-dimensional classification, achieving higher classification accuracy with reduced feature dimensions, thereby demonstrating strong potential for engineering applications.
End-to-end learned image compression (LIC) methods have superseded traditional transform-based methods. However, most LIC methods still prioritize pixel-level reconstruction without sufficiently preserving semantic information, thereby limiting their performance. To address this limitation, we propose a generic semantic-aware image compression (SaIC) framework that leverages knowledge distillation to integrate semantic priors into the LIC decoder, thereby guiding it to reconstruct images with higher fidelity. Specifically, we first constructed a semantic teacher decoder that incorporates multiscale semantic representations derived from a pretrained network. These semantic representations were dynamically fused with the intermediate features of the teacher decoder through semantic feature fusion blocks (SFFBs), which were equipped with a gating mechanism. Subsequently, a student decoder was designed to mimic the teacher through reconstruction-level distillation. Thus, the student decoder inherited semantic awareness without the need for introducing additional computational or bitrate overhead during inference. Comprehensive experiments on four benchmark LIC models demonstrated that the proposed SaIC framework is model-agnostic and delivers robust performance gains. The framework consistently improved the average PSNR by 0.1–0.4 dB and significantly optimized the BD-rate compared to the original baselines.
Trajectory prediction aims to accurately generate future trajectories and has significant applications in fields such as autonomous driving and robotic navigation. Diffusion model-based trajectory prediction algorithms, as one of the most mainstream approaches, have made remarkable progress. However, most of them struggle to effectively capture both long-term and short-term dependencies in temporal trajectories, and diffusion models with fixed diffusion steps are inadequate in adapting to substantial variations in motion complexity across different trajectories. To address these issues, this paper proposes a stochastic trajectory prediction framework based on Diffusion step Scheduling with Time-frequency Analysis, termed DSTA. DSTA jointly incorporates a Multi-scale Time-frequency Enhancement Module (MTEM) and an Energy Entropy-guided Diffusion Scheduler (EEDS) to model temporal dependencies and motion complexity. Specifically, MTEM introduces the discrete wavelet transform and employs a decomposition-aggregation architecture to perform multi-scale time-frequency analysis of trajectories. EEDS computes the energy entropy of trajectory features and maps it to trajectory-specific diffusion steps, thereby allocating adaptive diffusion generation budgets for diverse trajectories. Furthermore, a Step-Scheduled Diffusion Module (SSDM) is designed to utilize the MTEM-extracted trajectory features as priors and generate future trajectories based on the EEDS-assigned diffusion steps. Extensive comparative, visualization, ablation, and efficiency experiments are conducted on four datasets (ETH-UCY, NBA, SDD, NFL), fully demonstrating the effectiveness and superiority of the proposed framework.
Influence spread remains one of the most extensively studied phenomena in social network analysis due to its relevance for understanding real-world problems. It is well known that an isolated node can hardly exert significant influence, while overly large initial seeds lead to trivially high spread. This paper analyzes which characteristics initial seeds should have — when constructed from a given actor — in order to produce diverse and well-differentiated influence spread outcomes across actors, thereby yielding meaningful centrality criteria. To this end, we conduct controlled experiments using the General Influence Spread Rank (GISR), a centrality measure that evaluates the influence spread capacity of a node together with selected neighbors under specific parameter settings. The experiments consider 27 networks with varying levels of connectivity under 96 different parameter configurations. The results, based on correlation analysis and growth curve evaluation, highlight the importance of both neighborhood depth and initial activation probability, depending on network topology and the underlying diffusion model.
High-entropy alloys (HEAs) exhibit exceptional mechanical strength, thermal stability, and corrosion resistance; however, their vast compositional design space and the limited availability of structured composition–property data make systematic design and screening challenging. To address this challenge, this study proposes a reflection-enhanced, LLM-assisted literature-to-design framework for first-pass HEA screening and surrogate-guided inverse design. The workflow combines a dual-LLM extraction–reflection module for literature mining and data validation with composition-based surrogate modeling, genetic algorithm-driven candidate generation, and experimental validation in an integrated decision-support pipeline. The dual-LLM module first extracts structured composition–property data from unstructured scientific publications, achieving a precision of 0.943 on a benchmark derived from 25 articles and 397 entries. Using the curated dataset, an XGBoost surrogate model was trained to predict Young’s modulus from elemental composition, yielding a 10-fold cross-validated R2 of 0.640, an RMSE of 15.977 GPa, and an MAE of 10.985 GPa. The trained surrogate model was then coupled with a genetic algorithm as a constraint-aware candidate-generation procedure to identify candidate HEA compositions within target Young’s modulus ranges. Experimental validation on eight fabricated alloys showed good agreement for samples in well-represented regions of the training space, whereas larger deviations were observed in orientation- and microstructure-sensitive cases, highlighting the applicability boundary of the present composition-only surrogate model. These results provide a practical and transferable workflow for first-pass HEA screening and inverse design while clarifying where additional microstructural descriptors are needed for more reliable property prediction.
