Agentic Reasoning for Large Language Models: A Survey

Reasoning is a fundamental cognitive process that underpins inference, problem-solving, and decision-making in intelligent systems. While Large Language Models (LLMs) have demonstrated impressive reasoning abilities in controlled environments, such as those found in mathematics and coding benchmarks, they often struggle when faced with open-ended and dynamic real-world scenarios. The concept of "agentic reasoning" represents a significant paradigm shift in this area, seeking to transform LLMs from passive text generators into active, autonomous agents capable of planning, executing actions, and learning from their interactions with an environment. This comprehensive survey provides a structured roadmap for understanding agentic reasoning, organizing its various facets along three key dimensions. Firstly, it explores foundational agentic reasoning, which encompasses core single-agent functionalities like planning, tool utilization, and search mechanisms, all operating within relatively stable environments. Secondly, the survey delves into self-evolving agentic reasoning, examining how these agents can refine their capabilities through feedback loops, memory mechanisms, and adaptive strategies in environments that are constantly changing. Finally, it extends to collective multi-agent reasoning, where intelligence is distributed across multiple agents that coordinate their roles, share knowledge, and collaborate towards shared objectives. Across these different layers of agentic reasoning, the paper analyzes the inherent system constraints and various optimization settings. It draws a crucial distinction between in-context reasoning, which focuses on scaling test-time interaction through structured orchestration and adaptive workflow designs, and post-training reasoning, which involves optimizing agent behaviors through techniques like reinforcement learning and supervised fine-tuning. Furthermore, the survey meticulously reviews and contextualizes existing agentic reasoning frameworks within a broad spectrum of real-world applications and benchmarks. These applications span critical domains such as scientific discovery, robotics, healthcare, autonomous research, and mathematics, illustrating how diverse reasoning mechanisms are instantiated and evaluated across these varied fields. The authors synthesize these methods into a unified roadmap, bridging the theoretical aspects of thought with the practicalities of action, thereby offering actionable guidance for the development of agentic systems that can adapt to different environmental dynamics, optimization settings, and interaction paradigms. The paper concludes by outlining significant open challenges and proposing future directions for research in agentic reasoning. These include areas such as personalization, enabling long-horizon interactions, developing more sophisticated world modeling capabilities, achieving scalable multi-agent training, and establishing robust governance frameworks for the safe and effective deployment of these advanced AI systems in real-world contexts.