A Survey of the Recent Foundation Models Based on Their Architectural and Learning Strategies

The foundation models have transformed the decision-making process by utilizing large-scale architectures and advanced learning algorithms that exhibit high generalization capabilities across a variety of tasks. This survey is a detailed overview of the recent progress in foundation models, both in terms of architectural advances, such as Transformers, sparse attention mechanisms, and Mixture-of-Experts, as well as advancements in learning paradigms, such as self-supervised learning, reinforcement learning from human feedback, and few-shot adaptation. We characterize and examine recent multimodal systems till December 2025 such as DeepSeek, GPT, and CLIP, and identify their effectiveness and trade-offs. Prominent large-scale models are used to illustrate emerging opportunities in multilingual intelligence, while underscoring their key ethical considerations. The survey also discusses the main problems, such as scalability, bias, robustness, and generalization, and outlines the future research directions, including neuro-symbolic systems and quantum computing. Integrating technical innovations with the emphasis on social impact, this survey can be used as a guide for researchers and professionals in developing sustainable and secure AI systems for future.