Abstract. This paper presents a survey of the most significant machine learning methods and models as of the end of 2025 and proposes their systematic classification. Key trends from leading international conferences, including NeurIPS 2025, ICML 2025, and ICLR 2025, are analyzed. Special attention is given to State Space Models, Offline and Counterfactual Reinforcement Learning, Federated and Continual Learning, compact and efficient models, and multimodal systems. An analysis of industrial reports indicates the dominance of efficiency-first and privacy-by-design principles. In resource-constrained environments, priority is given to architectures capable of processing long sequences without transmitting raw data. The scientific novelty of this study lies in the development of a new classification model based on performance, energy consumption, security, and privacy metrics, as well as in the experimental validation of the advantages of hybrid approaches. Such models provide a balance between accuracy, speed, and privacy in decentralized and long-term systems. The practical significance of the research lies in offering concrete recommendations for model selection and integration when designing intelligent systems of various scales and application domains.

Keywords: machine learning, State Space Models, Mamba, Federated Learning, Continual Learning, small language models, hybrid architectures.