Zhong-Zhi Li, Duzhen Zhang, Ming-Liang Zhang, Jiaxin Zhang, Zengyan Liu, Yuxuan Yao, Haotian Xu, Junhao Zheng et al.
This survey paper comprehensively examines the development of reasoning LLMs aimed at transitioning from intuitive System 1 to logical System 2 thinking.
Existing LLMs excel at fast decision-making (System 1) but fail to perform step-by-step logical reasoning (System 2) for complex problems, limiting performance in high-difficulty tasks such as mathematics, coding, and scientific reasoning.
The paper first overviews the progress of foundational LLMs and early System 2 techniques, then systematically organizes methods for building reasoning LLMs (e.g., reinforcement learning, process-based reward models, self-correction). It also analyzes the evolution of various reasoning models and core methodologies, providing performance comparisons on major benchmarks.
This survey provides a systematic taxonomy of the reasoning LLM field and clarifies technical differences among state-of-the-art models like o1/o3 and R1. It also suggests future research directions (e.g., multimodal reasoning, continual learning, efficiency improvements) and offers a real-time GitHub repository to track the latest developments.