Paper summary: A Survey of Agentic Reasoning for Large Language Models: Towards Recursively Self-Improving and Collective Agents

Problem

LLMs reason well in closed settings but struggle in open-ended, dynamic environments. Agentic reasoning reframes LLMs as agents that plan, act and learn through interaction, and the literature has grown quickly.

Method

The survey organises the field along three layers (Figure 1):

1. Foundational agentic reasoning: planning, tool use and search in stable environments.
2. Self-evolving agentic reasoning: improving through feedback, memory and adaptation.
3. Collective multi-agent reasoning: coordination, knowledge sharing and shared goals.

Across all three, it separates in-context reasoning (structured orchestration at test time) from post-training reasoning (RL and SFT).

Key results

The contribution is the map, plus a review of frameworks and benchmarks across real-world applications such as science, robotics, healthcare and autonomous research. These map closely onto the planning, memory and multi-agent posts in my own agentic AI phase this autumn.

Limitations

As a survey, it reflects the literature at the start of 2026.