# A Survey on Large Language Model based Autonomous Agents

**Authors:** Lei Wang, Chen Ma, Xueyang Feng, Zeyu Zhang, Hao Yang, Jingsen Zhang, Zhiyuan Chen, Jiakai Tang, Xu Chen, Yankai Lin, Wayne Xin Zhao, Zhewei Wei, Ji-Rong Wen  
**arXiv:** [2308.11432](https://arxiv.org/abs/2308.11432)  
**Date:** 2023-08-22  
**Summary page:** https://www.summarizepaper.com/en/paper/2308.11432v1/

## Key points

- Autonomous agents incorporating large language models (LLMs) have transformed the field with human-level intelligence.
- LLM-based agents show potential in social science, natural science, and engineering applications.
- Diverse agent architectures are developed for different tasks, focusing on utilizing LLM capabilities.
- Memory systems of LLM-based agents feature hybrid structures with short-term and long-term functions for flexible planning.
- Memory formats in these agents use natural languages for reasoning and programming, enhancing storage and access efficiency.
- Surveys provide insights into LLM development and applications in downstream tasks; human alignment research addresses biases.
- Reasoning abilities are crucial for intelligence, influencing decision-making and problem-solving capabilities.

## Summary

The field of autonomous agents has undergone a significant transformation in recent years with the incorporation of large language models (LLMs) to achieve human-level intelligence. These LLM-based agents have shown immense potential in various applications across social science, natural science, and engineering domains. Researchers have developed diverse agent architectures tailored to different tasks and environments, with a focus on harnessing the full capabilities of LLMs. One key aspect that sets these LLM-based agents apart is their memory systems. Hybrid memory structures are commonly employed, distinguishing between short-term and long-term functions. The short-term component buffers recent perceptions while long-term memory consolidates important information over time. This dual-layered approach allows for flexible planning based on stable knowledge and insights. Memory formats also play a crucial role in how information is stored and accessed within these agents. Natural languages are often used for task reasoning and programming, enabling semantic-rich storage and access. Embeddings enhance memory reading efficiency, providing a comprehensive understanding of stored information. Additionally, related surveys have emerged to provide detailed insights into the development of LLMs and their applications in various downstream tasks. Human alignment with LLMs is an active area of research to address biases and ensure ethical deployment. Reasoning abilities are also highlighted as crucial for intelligence, influencing decision-making processes and problem-solving capabilities. Overall, the field of autonomous agents based on LLMs is rapidly evolving with ongoing challenges and future directions being explored by researchers. The integration of advanced memory systems and diverse memory formats showcases the potential for these agents to achieve human-like decision-making processes across a wide range of applications.