The pursuit of artificial intelligence (AI) has long been a goal for humanity, with AI agents seen as a promising avenue for achieving this. These AI agents are artificial entities that can sense their environment, make decisions, and take actions. While significant efforts have been made since the mid-20th century to develop intelligent AI agents, these endeavors have primarily focused on advancing algorithms or training strategies to improve specific capabilities or performance on particular tasks. However, what the AI community lacks is a sufficiently general and powerful model that can serve as a foundation for designing AI agents capable of adapting to diverse scenarios. This is where large language models (LLMs) come into play. LLMs are versatile and demonstrate remarkable capabilities, making them potential sparks for Artificial General Intelligence (AGI) and offering hope for building general AI agents. Researchers have leveraged LLMs as the basis for building AI agents and have achieved significant progress in this area. To provide a comprehensive understanding of LLM-based agents, this survey starts by tracing the concept of agents from its philosophical origins to its development in AI. It then explains why LLMs are suitable foundations for AI agents. Building upon this foundation, the survey presents a conceptual framework for LLM-based agents consisting of three main components: brain, perception, and action. This framework can be tailored to suit different applications and provides a starting point for designing adaptable AI agents. The survey further explores the extensive applications of LLM-based agents in various scenarios including single-agent scenarios, multi-agent scenarios, and human-agent cooperation. By delving into agent societies, it examines the behavior and personality of LLM-based agents and explores the social phenomena that emerge when these agents form societies. The insights gained from studying agent societies also offer valuable perspectives on human society. Finally, the survey discusses key topics and open problems within the field of LLM-based agent research. By addressing these challenges, researchers aim to further enhance the capabilities and potential of AI agents based on LLMs. In summary, this survey provides a detailed overview of the rise and potential of large language model-based agents. It highlights the significance of LLMs as foundations for AI agents and explores their applications in various scenarios. By examining agent societies it offers insights into both AI and human society while identifying key research topics and open problems that need to be addressed in order to advance the field of LLM-based agent research .
- - The pursuit of artificial intelligence (AI) has long been a goal for humanity
- - AI agents are artificial entities that can sense their environment, make decisions, and take actions
- - Efforts have primarily focused on advancing algorithms or training strategies to improve specific capabilities or performance on particular tasks
- - Large language models (LLMs) are versatile and demonstrate remarkable capabilities, making them potential sparks for Artificial General Intelligence (AGI)
- - LLMs serve as a foundation for designing AI agents capable of adapting to diverse scenarios
- - A conceptual framework for LLM-based agents consists of three main components: brain, perception, and action
- - LLM-based agents have extensive applications in single-agent scenarios, multi-agent scenarios, and human-agent cooperation
- - Agent societies emerge when LLM-based agents form societies and offer valuable perspectives on human society
- - Key topics and open problems within the field of LLM-based agent research need to be addressed to enhance the capabilities and potential of AI agents based on LLMs.
1. People have been trying to create artificial intelligence (AI) for a long time.
2. AI agents are artificial beings that can sense things, make decisions, and do things.
3. Scientists have been working on improving the abilities of AI agents by making better computer programs or training methods.
4. Large language models (LLMs) are very smart and can do many things, which could help us create even smarter AI in the future.
5. LLMs are important for designing AI agents that can adapt to different situations.
Definitions- Artificial Intelligence (AI): Creating machines that can think and act like humans.
- Agents: Artificial beings that can sense their surroundings, make choices, and take actions.
- Algorithms: Sets of instructions or rules that computers follow to solve problems or perform tasks.
- Performance: How well something does a task or job.
- Language Models: Computer programs that understand and generate human language.
- Artificial General Intelligence (AGI): Machines that have the same level of intelligence as humans and can do any task a human can do.
The Pursuit of Artificial Intelligence: Exploring the Potential of Large Language Model-Based Agents
Humans have long sought to create intelligent artificial agents, with AI research beginning in the mid-20th century. Since then, significant efforts have been made to develop algorithms and training strategies that can improve specific capabilities or performance on particular tasks. However, what has been lacking is a sufficiently general and powerful model that can serve as a foundation for designing AI agents capable of adapting to diverse scenarios. This is where large language models (LLMs) come into play. LLMs are versatile and demonstrate remarkable capabilities, making them potential sparks for Artificial General Intelligence (AGI) and offering hope for building general AI agents.
In this article, we will explore the rise and potential of large language model-based agents by tracing their concept from its philosophical origins to its development in AI. We will discuss why LLMs are suitable foundations for AI agents before presenting a conceptual framework for LLM-based agents consisting of three main components: brain, perception, and action. We will also examine the extensive applications of LLM-based agents in various scenarios including single-agent scenarios, multi-agent scenarios, and human-agent cooperation while exploring agent societies to gain valuable insights into both AI and human society. Finally, we will identify key research topics and open problems within the field of LLM-based agent research that need to be addressed in order to advance this field further.
Philosophical Origins
The concept of an agent has its roots in philosophy; it was first introduced by John Locke who defined an "agent" as an entity that acts upon its environment according to certain principles or laws which govern its behavior. Over time this definition evolved into what we now know as an autonomous agent – an entity capable of sensing its environment through sensors or other means such as vision or hearing; processing information using some form of reasoning; making decisions based on this information; taking actions accordingly; learning from past experiences; adapting itself over time; communicating with other entities if necessary; achieving goals set by itself or others; etc., all without any external control or guidance from humans.
Development In AI
AI researchers began exploring autonomous agents since the mid 20th century when they started developing algorithms designed specifically for problem solving tasks such as game playing or robotics navigation tasks among others. These algorithms were designed using techniques such as heuristics search methods like A*, genetic algorithms (GAs), reinforcement learning (RL), deep learning (DL), etc., which enabled them to solve complex problems autonomously without any direct intervention from humans but still lacked sufficient generality needed for creating truly intelligent systems capable of adapting themselves dynamically across different environments/scenarios without requiring manual tuning each time they encountered something new .
Large Language Models As Foundations For Agents
LLMs are becoming increasingly popular due their versatility which makes them suitable foundations for building adaptable AI agents capable tackling diverse real world challenges without needing manual tuning every time they encounter something new . They offer unprecedented levels power compared traditional machine learning models due their ability capture intricate relationships between words phrases sentences allowing them generate highly accurate predictions even when given limited data . Furthermore recent advancements natural language processing NLP have enabled researchers leverage these models create sophisticated conversational bots virtual assistants chatbots etc .
Conceptual Framework For LLM Based Agents
To provide comprehensive understanding how these models used build effective intelligent agents survey presents conceptual framework consisting three main components brain perception action Brain component responsible decision making process usually implemented recurrent neural networks RNNs Perception component responsible gathering sensory input environment usually implemented convolutional neural networks CNNs Action component responsible carrying out instructions generated brain usually implemented reinforcement learning RL Tailor fit different applications provides starting point designing adaptable intelligent systems
Applications Of LLM Based Agents
Researchers leveraging power large language models create sophisticated adaptive systems wide range applications Single agent scenarios involve one single autonomous system performing task Multi agent scenarios involve multiple interacting autonomous systems cooperating achieve common goal Human Agent Cooperation involves combination both human intelligence artificial intelligence working together accomplish task Examples include self driving cars automated customer service virtual personal assistants medical diagnosis robots etc
Agent Societies And Emergent Social Phenomena
Studying societies composed solely artificial entities allows researchers gain valuable insights behavior personality llm based agents well social phenomena emerge when these form societies These phenomena similar those found human societies example emergence leadership roles division labor competition cooperation trust negotiation conflict resolution etc Understanding dynamics interactions between individuals groups helps better understand nature complex social structures
< h 3 >Key Research Topics Open Problems
Despite significant progress made area there several important topics open problems remain address order continue advancing field llm based agent research These include improving interpretability robustness scalability safety privacy security efficiency transferability explainability multi modality integration multimodal interaction dynamic adaptation knowledge representation hybrid architectures distributed architectures among others Addressing these challenges help enhance capabilities potential llm based systems enable creation more advanced general purpose artificial intelligences