Agentic Large Language Models, a survey

TL;DR

A survey paper that organizes research on using LLMs as agents along three axes—reasoning, acting, and interacting—and presents a future research agenda.

Problem

Research on agentic LLMs has surged, but systematic classification and an integrated perspective are lacking. Researchers need help grasping the overall picture and setting future directions.

Approach

We classify the literature based on the core capabilities of agentic LLMs: reasoning, acting, and interacting. We analyze key studies within each category and comprehensively discuss cross-category synergies, applications, and risks.

Results & Contribution

We show that research on reasoning, acting, and interacting creates synergies. We identify practical applications in medical diagnosis, logistics, and financial market analysis, as well as the potential to accelerate scientific research itself. We also note that the inference-time behavior of agents can generate new training states, offering a solution to the problem of LLMs running out of training data. Open issues such as safety, liability, and security are clearly outlined.