Agentic Large Language Models, a survey (2503.23037v2)

Abstract: There is great interest in agentic LLMs, LLMs that act as agents. We review the growing body of work in this area and provide a research agenda. Agentic LLMs are LLMs that (1) reason, (2) act, and (3) interact. We organize the literature according to these three categories. The research in the first category focuses on reasoning, reflection, and retrieval, aiming to improve decision making; the second category focuses on action models, robots, and tools, aiming for agents that act as useful assistants; the third category focuses on multi-agent systems, aiming for collaborative task solving and simulating interaction to study emergent social behavior. We find that works mutually benefit from results in other categories: retrieval enables tool use, reflection improves multi-agent collaboration, and reasoning benefits all categories. We discuss applications of agentic LLMs and provide an agenda for further research. Important applications are in medical diagnosis, logistics and financial market analysis. Meanwhile, self-reflective agents playing roles and interacting with one another augment the process of scientific research itself. Further, agentic LLMs may provide a solution for the problem of LLMs running out of training data: inference-time behavior generates new training states, such that LLMs can keep learning without needing ever larger datasets. We note that there is risk associated with LLM assistants taking action in the real world, while agentic LLMs are also likely to benefit society.

Agentic LLMs: An Overview

The survey "Agentic LLMs, a Survey" explores an innovative facet of artificial intelligence research focused on LLMs that possess agent-like capabilities. This paper provides a structured overview of recent advancements in enhancing LLMs with agentic features such as reasoning, acting, and interacting. The authors categorize the landscape into three major areas: reasoning, action, and interaction, and propose a research agenda addressing how these elements connect and complement each other to bolster LLM society and applications in medicine, logistics, financial analysis, and scientific research, among others.

Description of Agentic LLMs

Agentic LLMs represent a shift from purely descriptive LLMs towards systems that can perform autonomous actions in real-world contexts. The defining characteristics of agentic LLMs include:

1. Reasoning: Agentic LLMs exhibit advanced reasoning abilities, which involve decision-making, retrieval, and reflection processes. They engage in multi-step reasoning, self-consistent reflection to improve performance, and rely on both implicit and explicit decision frameworks like the "tree of thoughts" method.
2. Acting: LLMs are enhanced with the capability to perform actions, integrate with tools, and interface with robotic systems. Through vision-language-action models, these LLMs enable practical applications such as automated diagnostics, market analysis, and task planning.
3. Interacting: This involves multi-agent systems where LLMs engage in social interactions, role-based task solving, and simulations of large-scale social dynamics. They demonstrate collective behaviors, social coordination, and emergent norms through enhanced communication and collaboration mechanisms.

Applications and Research Implications

Agentic LLMs have significant potential across various fields. They propose solutions for handling the scaling issues of LLM datasets by allowing models to continue learning dynamically through interaction data generated during inference. In medicine, logistics, and finance, agentic LLMs can serve as assistants making autonomous and informed decisions, thereby transforming these domains.

The survey underscores several implications for AI developments:

• Continued Learning: Through interaction, agentic LLMs discover new training states, mitigating the challenge of exhaustible training datasets. This leads to autonomous learning opportunities, a critical factor for advancing AI capabilities.
• Risk and Assurance: Despite their benefits, agentic LLMs acting in real-world scenarios pose potential risks, demanding careful analysis of security, liability, and ethical considerations in AI deployment.
• Theory of Mind and Social Dynamics: Interaction-based LLMs show potential for creating complex social simulations to paper social behaviors and emergent phenomena, thus contributing to broader social science research.
• Scientific Research Support: Agentic LLMs aid in hypothesis generation, experimentation, and data analysis by simulating human researchers and enhancing the scientific workflow.

Future Directions

Agentic LLMs provide numerous avenues for further research:

• Human-AI Interaction: Investigating adaptive techniques for improving conversation quality and collaboration between humans and LLMs.
• Open-ended Learning: Expanding frameworks for adaptive and interactive learning environments to facilitate continuous agent learning.
• Norms and Ethics in Multi-agent Societies: Further analysis is required for understanding the development of norms and behaviors in societies of agentic LLMs and ensuring alignment with human ethics.
• Security and Robustness: Developing mechanisms to mitigate risks associated with autonomous actions by agentic LLMs, especially in sensitive fields.

Conclusion

The exploration of agentic LLMs opens up transformative possibilities in how AI models interact, reason, and act within various domains. With their unique capabilities, these models not only promise to improve automated decision-making but also to serve as an integral part of intelligent systems that augment human efforts in complex tasks. The paper highlights the need for continued research into refining these models, ensuring stability, and addressing ethical concerns as we advance towards more sophisticated AI systems.