A Survey of Frontiers in LLM Reasoning: Inference Scaling, Learning to Reason, and Agentic Systems

Zixuan Ke Zixuan Ke1 Fangkai Jiao Fangkai Jiao3,4 Yifei Ming Yifei Ming1 Xuan-Phi Nguyen Xuan-Phi Nguyen1 Austin Xu Austin Xu1 Do Xuan Long Do Xuan Long2,4 Minzhi Li Minzhi Li2,4 Chengwei Qin Chengwei Qin3 Peifeng Wang Peifeng Wang1 Silvio Savarese Silvio Savarese1 Caiming Xiong Caiming Xiong1 Shafiq Joty Shafiq Joty1,3
1Salesforce AI Research 2National University of Singapore 3Nanyang Technological University 4A*STAR, Singapore

[Survey Paper]

Abstract

Reasoning is a fundamental cognitive process that enables logical inference, problem-solving, and decision-making. With the rapid advancement of large language models (LLMs), reasoning has emerged as a key capability that distinguishes advanced AI systems from conventional models that empower chatbots. In this survey, we categorize existing methods along two orthogonal dimensions: (1) Regimes, which define the stage at which reasoning is achieved (either at inference time or through dedicated training); and (2) Architectures, which determine the components involved in the reasoning process, distinguishing between standalone LLMs and agentic compound systems that incorporate external tools, and multiagent collaborations. Within each dimension, we analyze two key perspectives: (1) Input level, which focuses on techniques that construct high-quality prompts that the LLM condition on; and (2) Output level, which methods that refine multiple sampled candidates to enhance reasoning quality. This categorization provides a systematic understanding of the evolving landscape of LLM reasoning, highlighting emerging trends such as the shift from inference-scaling to learning-to-reason (e.g., DeepSeek-R1), and the transition to agentic workflows (e.g., OpenAI Deep Research, Manus Agent). Additionally, we cover a broad spectrum of learning algorithms, from supervised fine-tuning to reinforcement learning such as PPO and GRPO, and the training of reasoners and verifiers. We also examine key designs of agentic workflows, from established patterns like generator-evaluator and LLM debate to recent innovations. Finally, we identify emerging trends, such as domain-specific reasoning systems, and open challenges, such as evaluation and data quality. This survey aims to provide AI researchers and practitioners with a comprehensive foundation for advancing reasoning in LLMs, paving the way for more sophisticated and reliable AI systems

Overview

As shown in Figure 1, this area has rapidly gained research attention, often referred to as LLM reasoning or reasoning language model. Reasoning requires LLMs to go beyond directly producing an answer from a question; instead, they must generate the thinking process (implicitly or explicitly) in the form of 'question → reasoning steps → answer'. Building on this, the ability of LLMs to reason effectively depends on two factors: how and at what stage reasoning is achieved, and what components are involved in the reasoning process. Accordingly, in this survey, we categorize existing research into two orthogonal dimensions (Figure 2): (1) Regime, refers to whether reasoning is achieved through inference-time strategies (aka. inference-time scaling) or through direct learning and adaptation (learning to reason); and (2) Architecture, refers to whether reasoning happens within a single, standalone LLM or within an interactive, agentic system. Figure 3 outlines the structure of this survey.

Growth Trend Figure
Figure 1: Growth trend in LLM reasoning. We show the cumulative number (in thousands) of papers published from 2022 to February 2025, based on Semantic Scholar keyword search. Research on regimes and architectures has accelerated notably since the introduction of Chain-of-Thought (CoT) in 2022.
Proposed Catagorization
Figure 2: The proposed categorization over regimes, architectures, and unified perspectives in this survey.
Proposed Taxonomy
Figure 3: Taxonomy of LLM reasoning research organized in this survey by regimes (inference scaling, learning to reason) and architectures (standalone LLM, single-agent, multi-agent). Each leaf node includes examples from the literature that focus on the corresponding category.

BibTeX

@article{ke2025reasoning_survey,
  title={A Survey of Frontiers in LLM Reasoning: Inference Scaling, Learning to Reason, and Agentic Systems},
  author={Zixuan Ke and Fangkai Jiao and Yifei Ming and Xuan-Phi Nguyen and Austin Xu and Do Xuan Long and Minzhi Li and Chengwei Qin and PeiFeng Wang and silvio savarese and Caiming Xiong and Shafiq Joty},
  journal={arXiv},
  year={2025},
}