The Ascent of Machine Cognition: A Systematic Review of Large Reasoning Models

The evolution of Large Language Models (LLMs) has catalyzed significant advancements in artificial intelligence; however, their initial proficiency in fluent text generation often masked an underlying fragility in tasks demanding complex, multi-step reasoning. This review charts the systematic development of Large Reasoning Models (LRMs)—a functional classification characterized not by novel architectures, but by the augmentation of existing LLMs with sophisticated techniques designed to elicit, structure, and ground their latent cognitive abilities. We synthesize this burgeoning field across four principal vectors of innovation. The first is the elicitation of latent reasoning , a paradigm shift initiated by Chain-of-Thought (CoT) prompting, which revealed that structured interaction could unlock inherent problem-solving capabilities. The second vector involves the development of advanced reasoning architectures , transitioning from linear processes to deliberate, nonlinear frameworks such as Tree of Thoughts (ToT) and Graph of Thoughts (GoT), enabling exploration, backtracking, and synthesis. The third addresses the inherent limitations of LLMs through tool augmentation and program-aided reasoning , grounding models in external realities by offloading precise computation and information retrieval to external systems, and extending capabilities into formal domains and autonomous agency. The fourth vector confronts the critical frontier of self-correction and critique , a nascent area focused on imbuing models with the capacity for introspection and iterative refinement. By deconstructing these advancements and synthesizing them in comparative tables, we present a cohesive narrative of a field moving from stochastic pattern completion toward more robust, verifiable, and adaptable machine reasoning, and conclude with a critical evaluation of the profound open challenges that will define the future trajectory of artificial intelligence.