Large language models (LLMs) are increasingly used in agentic systems that reason over multiple steps, call tools, and interact with clinical knowledge sources. This survey examines whether LLM-based medical agents approximate expert diagnostic reasoning or reproduce its outputs without its causal and epistemic foundations. We organize agent design along five axes: reasoning paradigms, agent architectures, tool use, degree of autonomy, and embodiment. We then analyze learning and adaptation, evaluation, safety, and deployment as cross-cutting concerns that interact with these axes. Drawing on research in machine learning, biomedical informatics, and clinical AI, we examine how design choices shape diagnostic reliability and failure modes, and identify requirements for clinical and medical agents.

Recent advances have reframed Artificial Intelligence (AI) around world models, internal simulators that predict environmental dynamics and support general intelligence. This review surveys Large Language Models (LLMs) and Vision–Language Models (VLMs) not merely as text or vision processors, but as predictive systems capable of modeling state, causality, and dynamics. We organize the literature with a dual taxonomy: a theoretical taxonomy covering pixel-space generative simulators, latent-space predictive architectures, neuro-symbolic and code-coupled models, autoregressive foundation models, and hybrid mixtures of world models; and an application taxonomy covering autonomous driving, embodied robotics, interactive simulation and gaming, and social and agentic simulation. Synthesizing recent works, we conclude that although autoregressive transformers exhibit emergent topological and causal reasoning, robust physical world modeling requires moving beyond next-token prediction toward hybrid architectures that integrate generative diffusion models with symbolic grounding.

Spatial reasoning, the ability to represent and reason about spatial relations, geometry, topology, and embodied environments, is increasingly vital for evaluating and deploying large language models (LLMs) and vision–language models (VLMs). While multimodal foundation models show strong performance on spatial benchmarks, navigation, and embodied planning, much of this stems from shallow pattern matching rather than grounded spatial understanding. This review offers a comprehensive, framework-driven analysis of spatial reasoning in LLMs and VLMs, covering tasks, representations, datasets, benchmarks, architectures, training, and failure modes. We also link spatial reasoning to embodied AI, emphasizing challenges in generalization, compositionality, and grounding. Our survey unifies diverse fields, NLP, vision, robotics, and cognitive science, to guide the development of truly spatially grounded foundation models.