Agentic Reasoning for Large Language Models

The provided text outlines the paradigm of agentic reasoning, where large language models (LLMs) transition from passive text generators to autonomous agents that plan, act, and learn through environment interaction. This survey organizes the field into three layers: foundational capabilities like tool use and planning, self-evolving mechanisms that utilize feedback and memory to improve, and collective intelligence involving multi-agent collaboration. Researchers distinguish between in-context reasoning, which optimizes performance at inference time through structured workflows, and post-training reasoning, which embeds these skills into model weights via fine-tuning or reinforcement learning. The roadmap further explores real-world applications in robotics, healthcare, and science, while identifying benchmarks to measure agent performance. Ultimately, the sources provide a systematic framework for developing more adaptive and goal-oriented AI systems.