Agentic Reasoning for Large Language Models

TL;DR

Agentic Reasoning reframes large language models as autonomous agents that plan, act, and learn through continual interaction, bridging thought and action in dynamic environments. The survey organizes this modality across three layers—foundational, self-evolving, and collective—and two optimization modes: in-context reasoning during inference and post-training capability optimization. It surveys frameworks and benchmarks across mathematics, science, robotics, healthcare, and autonomous web exploration, culminating in a unified roadmap and identified challenges like personalization, long-horizon interaction, world modeling, scalable multi-agent training, and governance. By linking reasoning to action and collaboration, the work illuminates how adaptive, memory-enabled, multi-agent systems can achieve robust, real-world performance. Key contributions include a formal latent-space model, a comprehensive taxonomy, and a synthesis of representative applications and benchmarks that anchor future research in agentic reasoning.

Abstract

Reasoning is a fundamental cognitive process underlying inference, problem-solving, and decision-making. While large language models (LLMs) demonstrate strong reasoning capabilities in closed-world settings, they struggle in open-ended and dynamic environments. Agentic reasoning marks a paradigm shift by reframing LLMs as autonomous agents that plan, act, and learn through continual interaction. In this survey, we organize agentic reasoning along three complementary dimensions. First, we characterize environmental dynamics through three layers: foundational agentic reasoning, which establishes core single-agent capabilities including planning, tool use, and search in stable environments; self-evolving agentic reasoning, which studies how agents refine these capabilities through feedback, memory, and adaptation; and collective multi-agent reasoning, which extends intelligence to collaborative settings involving coordination, knowledge sharing, and shared goals. Across these layers, we distinguish in-context reasoning, which scales test-time interaction through structured orchestration, from post-training reasoning, which optimizes behaviors via reinforcement learning and supervised fine-tuning. We further review representative agentic reasoning frameworks across real-world applications and benchmarks, including science, robotics, healthcare, autonomous research, and mathematics. This survey synthesizes agentic reasoning methods into a unified roadmap bridging thought and action, and outlines open challenges and future directions, including personalization, long-horizon interaction, world modeling, scalable multi-agent training, and governance for real-world deployment.

Figures (12)

• Figure 1: An overview of agentic reasoning.
• Figure 2: Overview of Planning Reasoning in LLM agents, categorized into in-context planning and post-training planning.
• Figure 3: Comparison between traditional LLM and agentic tool-use systems. While traditional models operate in a closed world with fixed reasoning, agentic tool-use systems enable dynamic selection, orchestration, and integration of external tools, allowing agents to extend reasoning, improve precision, and dynamically adapt across domains.
• Figure 4: Comparison between traditional RAG systems and agentic search systems. Traditional RAG relies on static retrieval over a vector database, while agentic search introduces autonomous decision-making for when, what, and how to retrieve, enabling dynamic search, in-context retrieval, critique-and-adapt loops, and tool use.
• Figure 5: Illustration of three forms of agentic feedback mechanisms.Inference-time reflection enables real-time self-critique and revision during reasoning; offline adaptation consolidates feedback into model parameters for long-term improvement; and outcome-based feedback relies on validator signals (success or failure) to refine behavior through retry. Together, they represent a continuum from adaptive reflection to stable learning and efficient validation.