Agentic Reasoning for Large Language Models

Abstract

Agentic reasoning marks a paradigm shift by reframing LLMs as autonomous agents that plan, act, and learn through continual interaction. The paper characterizes environmental dynamics through three layers: foundational agentic reasoning (establishing core single-agent capabilities including planning, tool use, and search in stable environments), self-evolving agentic reasoning (studying how agents refine capabilities through feedback, memory, and adaptation), and collective multi-agent reasoning (extending intelligence to collaborative settings). The survey distinguishes between in-context reasoning (test-time interaction) and post-training reasoning (reinforcement learning optimization).

Key Contributions

• Structured framework organizing agentic reasoning across foundational, self-evolving, and multi-agent layers
• Distinction between in-context reasoning (test-time) and post-training reasoning (RL optimization)
• Comprehensive review spanning science, robotics, healthcare, autonomous research, and mathematics
• Identification of open challenges: personalization, long-horizon interaction, governance

Methodology

Three-dimensional organizational scheme examining environmental dynamics, reasoning scales, and real-world applications across multiple domains.

Results

The survey establishes agentic reasoning as a paradigm shift enabling LLMs to operate effectively in dynamic settings through systematic planning and adaptation mechanisms. Covers applications across science, robotics, healthcare, and mathematics.

Limitations

• Focus primarily on text-based reasoning
• Multi-agent governance and safety remain open problems
• Long-horizon interaction capabilities still developing

Source: Agentic Reasoning for Large Language Models by Tianxin Wei et al.