FOUNDATION MODELS BEYOND LANGUAGE: A COMPREHENSIVE STUDY OF MULTIMODAL, AGENTIC, AND RETRIEVAL-AUGMENTED ARCHITECTURES FOR REAL-WORLD DECISION MAKING

Abstract

Foundation models have transformed artificial intelligence by demonstrating remarkable adaptability across a range of tasks. However, language-only approaches remain insufficient for real-world decision-making, where context requires perception across modalities, adaptive agency, and dynamic grounding in external knowledge. This study critically examines the evolution of foundation models beyond language by analyzing three emerging paradigms: multimodal models, agentic architectures, and retrieval-augmented systems. A systematic and analytical review was undertaken to explore how each paradigm addresses the limitations of traditional language-based models. Multimodal models expand the perceptual capacity of artificial intelligence through the integration of text, vision, and structured data. Agentic models move beyond passive output generation to autonomous reasoning and planning, supported by memory augmentation and external tool use. Retrieval-augmented models reduce hallucination and increase reliability by linking parametric knowledge with external databases. Comparative synthesis reveals that each paradigm contributes unique strengths but also introduces challenges related to scalability, safety, interpretability, and evaluation. The study highlights the theoretical alignment of these paradigms with cognitive functions such as perception, reasoning, memory, and action. It further emphasizes the need for ethical and regulatory safeguards to address bias, transparency, and human–artificial intelligence collaboration. The findings suggest that the future of foundation models lies in the integration of multimodality, agency, and retrieval, paving the way for unified decision-making systems that are both technically advanced and socially responsible.