Agentic Reasoning for Large Language Models: A Survey

Large Language Models (LLMs) have demonstrated remarkable capabilities in various tasks, particularly within well-defined and closed-world settings, such as those found in standard benchmarks for mathematics and code. However, their performance often diminishes significantly when faced with open-ended and dynamic environments. This limitation has spurred the development of 'agentic reasoning,' a new paradigm that aims to bridge the gap between an LLM's inherent reasoning abilities and its capacity for autonomous action and learning. The recently published survey on OpenReview, titled "Agentic Reasoning for Large Language Models: A Survey," delves into this transformative approach. It reframes LLMs not merely as predictive text generators but as autonomous agents that can plan, execute actions, and learn from continuous interaction with their surroundings. This shift is crucial for developing AI systems that can operate effectively in complex, real-world scenarios. The survey offers a systematic roadmap for understanding agentic reasoning, organizing its methodologies along three complementary dimensions. Firstly, it characterizes environmental dynamics, distinguishing between foundational single-agent capabilities (like planning, tool use, and search in stable environments), self-evolving agentic reasoning (where agents refine capabilities through feedback and adaptation), and collective multi-agent reasoning (involving collaboration and knowledge sharing among multiple agents). Secondly, the paper examines optimization settings, differentiating between in-context reasoning, which scales test-time interaction through structured orchestration, and post-training reasoning, which optimizes behaviors via reinforcement learning and supervised fine-tuning. Finally, it reviews agentic reasoning frameworks within various real-world applications and benchmarks, including science, robotics, healthcare, and autonomous research, illustrating how different reasoning mechanisms are implemented and evaluated across these diverse domains. By synthesizing these methods, the survey provides actionable guidance for the development of agentic systems. It also identifies several open challenges and future directions for the field, such as personalization, long-horizon interaction, world modeling, scalable multi-agent training, and the establishment of robust governance frameworks for real-world deployment. This comprehensive overview underscores the potential of agentic reasoning to enhance LLM capabilities, moving towards more intelligent, adaptive, and autonomous AI.