paitesanshi / llm-agent-survey

A Survey on LLM-based Autonomous Agents

Autonomous agents are designed to achieve specific objectives through self-guided instructions. With the emergence and growth of large language models (LLMs), there is a growing trend in utilizing LLMs as fundamental controllers for these autonomous agents. While previous studies in this field have achieved remarkable successes, they remain independent proposals with little effort devoted to a systematic analysis. To bridge this gap, we conduct a comprehensive survey study, focusing on the construction, application, and evaluation of LLM-based autonomous agents. In particular, we first explore the essential components of an AI agent, including a profile module, a memory module, a planning module, and an action module. We further investigate the application of LLM-based autonomous agents in the domains of natural sciences, social sciences, and engineering. Subsequently, we delve into a discussion of the evaluation strategies employed in this field, encompassing both subjective and objective methods. Our survey aims to serve as a resource for researchers and practitioners, providing insights, related references, and continuous updates on this exciting and rapidly evolving field.

📍 This is the first released and published survey paper in the field of LLM-based autonomous agents.

Paper link: A Survey on Large Language Model based Autonomous Agents

Update Records

• 🔥 [25/3/2024] Our survey paper has been accepted by Frontiers of Computer Science, which is the first published survey paper in the field of LLM-based agents.

• 🔥 [9/28/2023] We have compiled and summarized papers related to LLM-based Agents that have been accepted by Neurips 2023 in the repository LLM-Agent-Paper-Digest. This repository will continue to be updated with accepted agent-related papers in the future.

• 🔥 [9/8/2023] The second version of our survey has been released on arXiv.

  Updated contents

  • 📚 Additional References

    • We have added 31 new works until 9/1/2023 to make the survey more comprehensive and up-to-date.

  • 📊 New Figures

    • Figure 3: This is a new figure illustrating the differences and similarities between various planning approaches. This helps in gaining a clearer understanding of the comparisons between different planning methods.
    • Figure 4: This is a new figure that describes the evolutionary path of model capability acquisition from the "Machine Learning era" to the "Large Language Model era" and then to the "Agent era." Specifically, a new concept, "mechanism engineering," has been introduced, which, along with "parameter learning" and "prompt engineering," forms part of this evolutionary path.

  • 🔍 Optimized Classification System

    • We have slightly modified the classification system in our survey to make it more logical and organized.

• 🔥 [8/23/2023] The first version of our survey has been released on arXiv.
  

Table of Content

• 🤖 Construction of LLM-based Autonomous Agent
• 📍 Applications of LLM-based Autonomous Agent
• 📊 Evaluation on LLM-based Autonomous Agent
• 🌐 More Comprehensive Summarization
• 👨‍👨‍👧‍👦 Maintainers
• 📚 Citation
• 💪 How to Contribute
• 🫡 Acknowledgement
• 📧 Contact Us

🤖 Construction of LLM-based Autonomous Agent

Model	Profile	Memory		Planning	Action	CA	Paper	Code
		Operation	Structure
WebGPT	-	-	-	-	w/ tools	w/ fine-tuning	Paper	-
SayCan	-	-	-	w/o feedback	w/o tools	w/o fine-tuning	Paper	Code
MRKL	-	-	-	w/o feedback	w/ tools	-	Paper	-
Inner Monologue	-	-	-	w/ feedback	w/o tools	w/o fine-tuning	Paper	Code
Social Simulacra	GPT-Generated	-	-	-	w/o tools	-	Paper	-
ReAct	-	-	-	w/ feedback	w/ tools	w/ fine-tuning	Paper	Code
LLM Planner	-	-	-	w/ feedback	w/o tools	Environment feedback	Paper	Code
MALLM	-	Read/Write	Hybrid	-	w/o tools	-	Paper	-
aiflows	-	Read/Write/ Reflection	Hybrid	w/ feedback	w/ tools	-	Paper	Code
DEPS	-	-	-	w/ feedback	w/o tools	w/o fine-tuning	Paper	Code
Toolformer	-	-	-	w/o feedback	w/ tools	w/ fine-tuning	Paper	Code
Reflexion	-	Read/Write/ Reflection	Hybrid	w/ feedback	w/o tools	w/o fine-tuning	Paper	Code
CAMEL	Handcrafting & GPT-Generated	-	-	w/ feedback	w/o tools	-	Paper	Code
API-Bank	-	-	-	w/ feedback	w/ tools	w/o fine-tuning	Paper	-
Chameleon	-	-	-	w/o feedback	w/ tools	-	Paper	Code
ViperGPT	-	-	-	-	w/ tools	-	Paper	Code
HuggingGPT	-	-	Unified	w/o feedback	w/ tools	-	Paper	Code
Generative Agents	Handcrafting	Read/Write/ Reflection	Hybrid	w/ feedback	w/o tools	-	Paper	Code
LLM+P	-	-	-	w/o feedback	w/o tools	-	Paper	-
ChemCrow	-	-	-	w/ feedback	w/ tools	-	Paper	Code
OpenAGI	-	-	-	w/ feedback	w/ tools	w/ fine-tuning	Paper	Code
AutoGPT	-	Read/Write	Hybrid	w/ feedback	w/ tools	w/o fine-tuning	-	Code
SCM	-	Read/Write	Hybrid	-	w/o tools	-	Paper	Code
Socially Alignment	-	Read/Write	Hybrid	-	w/o tools	Example	Paper	Code
GITM	-	Read/Write/ Reflection	Hybrid	w/ feedback	w/o tools	w/ fine-tuning	Paper	Code
Voyager	-	Read/Write/ Reflection	Hybrid	w/ feedback	w/o tools	w/o fine-tuning	Paper	Code
Introspective Tips	-	-	-	w/ feedback	w/o tools	w/o fine-tuning	Paper	-
RET-LLM	-	Read/Write	Hybrid	-	w/o tools	w/ fine-tuning	Paper	-
ChatDB	-	Read/Write	Hybrid	w/ feedback	w/ tools	-	Paper	-
S3	Dataset alignment	Read/Write/ Reflection	Hybrid	-	w/o tools	w/ fine-tuning	Paper	-
ChatDev	Handcrafting	Read/Write/ Reflection	Hybrid	w/ feedback	w/o tools	w/o fine-tuning	Paper	Code
ToolLLM	-	-	-	w/ feedback	w/ tools	w/ fine-tuning	Paper	Code
MemoryBank	-	Read/Write/ Reflection	Hybrid	-	w/o tools	-	Paper	Code
MetaGPT	Handcrafting	Read/Write/ Reflection	Hybrid	w/ feedback	w/ tools	-	Paper	Code
L2MAC	Handcrafting	Read/Write/ Reflection	Hybrid	w/ feedback	w/ tools	-	Paper	Code
LEO	-	-	-	w/ feedback	w/o tools	w/ fine-tuning	Paper	Code
JARVIS-1	-	Read/Write/ Reflection	Hybrid	w/ feedback	w/ tools	w/o fine-tuning	Paper	Code
CLOVA	-	Read/Write/ Reflection	Hybrid	w/ feedback	w/ tools	w/ fine-tuning	Paper	Code
LearnAct	-	-	-	w/ feedback	w/ tools	w/ fine-tuning	Paper	Code

• More papers can be found at More comprehensive Summarization.
• CA means the strategy of model capability acquisition.

📍 Applications of LLM-based Autonomous Agent

Title	Social Science	Natural Science	Engineering	Paper	Code
Drori et al.	-	Science Education	-	Paper	-
SayCan	-	-	Robotics & Embodied AI	Paper	Code
Inner monologue	-	-	Robotics & Embodied AI	Paper	Code
Language-Planners	-	-	Robotics & Embodied AI	Paper	Code
Social Simulacra	Social Simulation	-	-	Paper	-
TE	Psychology	-	-	Paper	Code
Out of One	Political Science and Economy	-	-	Paper	-
LIBRO	CS&SE	-	-	Paper	-
Blind Judgement	Jurisprudence	-	-	Paper	-
Horton	Political Science and Economy	-	-	Paper	-
DECKARD	-	-	Robotics & Embodied AI	Paper	Code
Planner-Actor-Reporter	-	-	Robotics & Embodied AI	Paper	-
DEPS	-	-	Robotics & Embodied AI	Paper	-
RCI	-	-	CS&SE	Paper	Code
Generative Agents	Social Simulation	-	-	Paper	Code
SCG	-	-	CS&SE	Paper	-
IGLU	-	-	Civil Engineering	Paper	-
IELLM	-	-	Industrial Automation	Paper	-
ChemCrow	-	Document and Data Management; Documentation, Data Managent; Science Education	-	Paper	-
Boiko et al.	-	Document and Data Management; Documentation, Data Managent; Science Education	-	Paper	-
GPT4IA	-	-	Industrial Automation	Paper	Code
Self-collaboration	-	-	CS&SE	Paper	-
E2WM	-	-	Robotics & Embodied AI	Paper	Code
Akata et al.	Psychology	-	-	Paper	-
Ziems et al.	Psychology; Political Science and Economy; Research Assistant	-	-	Paper	-
AgentVerse	Social Simulation	-	-	Paper	Code
SmolModels	-	-	CS&SE	-	Code
TidyBot	-	-	Robotics & Embodied AI	Paper	Code
PET	-	-	Robotics & Embodied AI	Paper	-
Voyager	-	-	Robotics & Embodied AI	Paper	Code
GITM	-	-	Robotics & Embodied AI	Paper	Code
NLSOM	-	Science Education	-	Paper	-
LLM4RL	-	-	Robotics & Embodied AI	Paper	-
GPT Engineer	-	-	CS&SE	-	Code
Grossman et al.	-	Experiment Assistant; Science Education	-	Paper	-
SQL-PALM	-	-	CS&SE	Paper	-
REMEMBER	-	-	Robotics & Embodied AI	Paper	-
DemoGPT	-	-	CS&SE	-	Code
Chatlaw	Jurisprudence	-	-	Paper	Code
RestGPT	-	-	CS&SE	Paper	Code
Dialogue shaping	-	-	Robotics & Embodied AI	Paper	-
TaPA	-	-	Robotics & Embodied AI	Paper	-
Ma et al.	Psychology	-	-	Paper	-
Math Agents	-	Science Education	-	Paper	-
SocialAI School	Social Simulation	-	-	Paper	-
Unified Agent	-	-	Robotics & Embodied AI	Paper	-
Wiliams et al.	Social Simulation	-	-	Paper	-
Li et al.	Social Simulation	-	-	Paper	-
S3	Social Simulation	-	-	Paper	-
Dialogue Shaping	-	-	Robotics & Embodied AI	Paper	-
RoCo	-	-	Robotics & Embodied AI	Paper	Code
Sayplan	-	-	Robotics & Embodied AI	Paper	Code
aiflows	-	-	CS & SE	Paper	Code
ToolLLM	-	-	CS&SE	Paper	Code
ChatDEV	-	-	CS&SE	Paper	-
Chao et al.	Social Simulation	-	-	Paper	-
AgentSims	Social Simulation	-	-	Paper	Code
ChatMOF	-	Document and Data Management; Science Education	-	Paper	-
MetaGPT	-	-	CS&SE	Paper	Code
L2MAC	-	-	CS&SE	Paper	Code
Codehelp	-	Science Education	CS&SE	Paper	-
AutoGen	-	Science Education	-	Paper	-
RAH	-	-	CS&SE	Paper	-
DB-GPT	-	-	CS&SE	Paper	Code
RecMind	-	-	CS&SE	Paper	-
ChatEDA	-	-	CS&SE	Paper	-
InteRecAgent	-	-	CS&SE	Paper	-
PentestGPT	-	-	CS&SE	Paper	-
Codehelp	-	-	CS&SE	Paper	-
ProAgent	-	-	Robotics & Embodied AI	Paper	-
MindAgent	-	-	Robotics & Embodied AI	Paper	-
LEO	-	-	Robotics & Embodied AI	Paper	-
JARVIS-1	-	-	Robotics & Embodied AI	Paper	-
CLOVA	-	-	CS&SE	Paper	-
AgentTrust	-	Social Simulation	✓	Paper	Code	-

• More papers can be found at More comprehensive Summarization.

📊 Evaluation on LLM-based Autonomous Agent

Model	Subjective	Objective	Benchmark	Paper	Code
WebShop	-	Environment Simulation; Multi-task Evaluation	✓	Paper	Code
Social Simulacra	Human Annotation	Social Evaluation	-	Paper	-
TE	-	Social Evaluation	-	Paper	Code
LIBRO	-	Software Testing	-	Paper	-
ReAct	-	Environment Simulation	✓	Paper	Code
Out of One, Many	Turing Test	Social Evaluation; Multi-task Evaluation	-	Paper	-
DEPS	-	Environment Simulation	✓	Paper	-
Jalil et al.	-	Software Testing	-	Paper	Code
Reflexion	-	Environment Simulation; Multi-task Evaluation	-	Paper	Code
IGLU	-	Environment Simulation	✓	Paper	-
Generative Agents	Human Annoation; Turing Test	-	-	Paper	Code
ToolBench	Human Annoation	Multi-task Evalution	✓	Paper	Code
GITM	-	Environment Simulation	✓	Paper	Code
Two-Failures	-	Multi-task Evalution	-	Paper	-
Voyager	-	Environment Simulation	✓	Paper	Code
SocKET	-	Social Evaluation; Multi-task Evaluation	✓	Paper	-
Mobile-Env	-	Environment Simulation; Multi-task Evaluation	✓	Paper	Code
Clembench	-	Environment Simulation; Multi-task Evaluation	✓	Paper	Code
Mind2Web	-	Environment Simulation; Multi-task Evaluation	✓	06/2023	Paper	Code
Dialop	-	Social Evaluation	✓	Paper	Code
Feldt et al.	-	Software Testing	-	Paper	-
CO-LLM	Human Annoation	Environment Simulation	-	Paper	Code
Tachikuma	Human Annoation	Environment Simulation	✓	Paper	-
WebArena	-	Environment Simulation	✓	Paper	Code
RocoBench	-	Environment Simulation; Social Evaluation; Multi-task Evaluation	✓	Paper	Code
AgentSims	-	Social Evaluation	-	Paper	Code
AgentBench	-	Multi-task Evaluation	✓	Paper	Code
BOLAA	-	Environment Simulation; Multi-task Evaluation; Software Testing	✓	Paper	Code
Gentopia	-	Isolated Reasoning; Multi-task Evaluation	✓	Paper	Code
EmotionBench	Human Annotation	-	✓	Paper	Code
PTB	-	Software Testing	✓	Paper	-
MintBench	-	Multi-task Evaluation	✓	Paper	Code
MindAgent	-	Environment Simulation; Multi-task Evaluation	✓	Paper	-
JARVIS-1	-	Environment Simulation	-	Paper	-

• More papers can be found at More comprehensive Summarization.

---

🌐 More Comprehensive Summarization

We are maintaining an interactive table that contains more comprehensive papers related to LLM-based Agents. This table includes details such as tags, authors, publication date, and more, allowing you to sort, filter, and find the papers of interest to you.

👨‍👨‍👧‍👦 Maintainers

• Lei Wang@Paitesanshi
• Chen Ma@Uily
• Xueyang Feng@XueyangFeng

📚 Citation

If you find this survey useful, please cite our paper:

@misc{wang2023survey,
      title={A Survey on Large Language Model based Autonomous Agents}, 
      author={Lei Wang and Chen Ma and Xueyang Feng and Zeyu Zhang and Hao Yang and Jingsen Zhang and Zhiyuan Chen and Jiakai Tang and Xu Chen and Yankai Lin and Wayne Xin Zhao and Zhewei Wei and Ji-Rong Wen},
      year={2023},
      eprint={2308.11432},
      archivePrefix={arXiv},
      primaryClass={cs.AI}
}

💪 How to Contribute

If you have a paper or are aware of relevant research that should be incorporated, please contribute via pull requests, issues, email, or other suitable methods.

🫡 Acknowledgement

We thank the following people for their valuable suggestions and contributions to this survey:

• Yifan Song@Yifan-Song793
• Qichen Zhao@Andrewzh112
• Ikko E. Ashimine@eltociear

📧 Contact Us

If you have any questions or suggestions, please contact us via:

• Email: [email protected], [email protected]

llm-agent-survey's Issues

多智能体

能否增加一个信息，是否为多智能体还是单智能体

I'd like to share recent work "Empowering Large Language Model Agents through Action Learning"

Hello,

Thanks for your comprehensive and inspiring paper list! I'd like to share our recent work titled "Empowering Large Language Model Agents through Action Learning," which may be of interest to the paper list readers. The paper may be added to the Planning Section.

Paper: https://arxiv.org/abs/2402.15809
Code: https://github.com/zhao-ht/LearnAct
This work proposes the LearnAct framework, which employs an iterative learning approach to dynamically create and refine learnable actions (skills). By evaluating and amending actions in response to errors observed during unsuccessful training episodes, LearnAct systematically increases the efficiency and adaptability of actions undertaken by Large Language Model (LLM) agents.
The experiment conducted within the contexts of Robotic Planning and Alfworld environments demonstrated that LearnAct can significantly enhance agent performance on given tasks.

I hope this contributes to the great paper list!

Equation 1

Thank you for providing this comprehensive and outstanding survey.

Is it possible that "argmax" should be used instead of "argmin" in Equation 1 on page 6?

AppAgent

web url：https://appagent-official.github.io/
Code：https://github.com/mnotgod96/AppAgent
paper：https://arxiv.org/abs/2312.13771

能否在交互式表格中加入是否需要finetune这列呀

就是这个表格
https://abyssinian-molybdenum-f76.notion.site/237e9f7515d543c0922c74f4c3012a77?v=0a309e53d6454afcbe7a5a7e169be0f9

Pointer to early LLM-based simulation agent

Hi,

Thanks for compiling this thorough and excellent review (article and website)!

Here's just a pointer to an early paper which set up a simulation agent and might be worth mentioning:

• "Natural-Language Multi-Agent Simulations of Argumentative Opinion Dynamics" https://www.jasss.org/25/1/2.html

Best,
Gregor

Not an issue, just wanted to say thank you for keeping this up to date

This aspect of our world is moving so fast its very difficult for us to keep track. Resources like this one you have put together are incredibly helpful

One reference on LLM Agents playing Trust Games

Congratulations on your recent solid survey paper and impressive paper list!

We have a related paper on LLM Agents playing Trust Games.

Can Large Language Model Agents Simulate Human Trust Behaviors?

• arxiv : https://arxiv.org/abs/2402.04559
• code : https://github.com/camel-ai/agent-trust
• project website : https://www.camel-ai.org/research/agent-trust
• We discover the trust behaviors of LLM agents under the framework of Trust Games, and the high behavioral alignment between LLM agents and humans regarding the trust behaviors, particularly for GPT-4, indicating the feasibility to simulate human trust behaviors with LLM agents.
• abstract: Large Language Model (LLM) agents have been increasingly adopted as simulation tools to model humans in applications such as social science. However, one fundamental question remains: can LLM agents really simulate human behaviors? In this paper, we focus on one of the most critical behaviors in human interactions, trust, and aim to investigate whether or not LLM agents can simulate human trust behaviors. We first find that LLM agents generally exhibit trust behaviors, referred to as agent trust, under the framework of Trust Games, which are widely recognized in behavioral economics. Then, we discover that LLM agents can have high behavioral alignment with humans regarding trust behaviors, particularly for GPT-4, indicating the feasibility to simulate human trust behaviors with LLM agents. In addition, we probe into the biases in agent trust and the differences in agent trust towards agents and humans. We also explore the intrinsic properties of agent trust under conditions including advanced reasoning strategies and external manipulations. We further offer important implications of our discoveries for various scenarios where trust is paramount. Our study provides new insights into the behaviors of LLM agents and the fundamental analogy between LLMs and humans.

Introducing a new paper on role-playing LLM agents (ACL 2024 Findings)

Hi, what a fantastic resource for developing intelligent LLM agents!

I wanted to highlight a recent paper presented at ACL 2024 Findings: TimeChara: Evaluating Point-in-Time Character Hallucination of Role-Playing Large Language Models.
This study focuses on assessing hallucinations in role-playing LLM agents when they simulate characters at specific moments in time.

We would greatly appreciate it if you could consider adding our paper to your survey.
Thanks!

Update Chart

Hi, I was wondering if you have an updated list of LLM papers? You have a really nice chart that goes until August 2023 but it would be great to have an updated version or at least a list of all LLM papers by date. Do you have this?

APi-Bank URLs are dead

The URLs for paper "API-Bank" are missing.

Currently they point to https://github.com/Paitesanshi/LLM-Agent-Survey/blob/main/url

unified agent framework

Hello, will the unified agent framework mentioned in the article be open source?

citation version

I have a quick question regarding your citations.
I've noticed that a significant proportion of your citations (over 90%) are the arXiv version of papers.
But many of them have already been published.
I'm curious about why.

Add Suspicion-Agent

Suspicion Agent: Playing Imperfect Information Games with Theory of Mind Aware GPT-4

Category: Imperfect Information Game, Psychology

Project Demo
Game Data Replay
Code

[165] reference

First, I really thank you for the contribution that you guys have made. I am still reading it. BTW, no offense, I notice some commas are missing in the [165] reference.:joy:

Introducing our NeurIPS 2023 paper

Hi!

This list is an invaluable resource in the area of building intelligent agents with LLMs.

I wanted to take a moment to bring your attention to a recent NeurIPS-23 paper from our lab: Leveraging Pre-trained Large Language Models to Construct and Utilize World Models for Model-based Task Planning. Instead of getting plans from LLMs directly, it allows the agent to use external planners to reliably search for plans (somewhat in a similar vein to tool-augmented LLMs).

We would be grateful if you would consider including our papers in your survey. We believe it would greatly benefit the readers interested in this burgeoning area of LLM-driven intelligent agents.

Best regards

Missing Related Work

Dear Authors,

Thank you for your efforts in proposing this survey paper.

We are the authors of “Chameleon,” a framework designed to seamlessly integrate LLM agents with various external tools (https://arxiv.org/abs/2304.09842, https://github.com/lupantech/chameleon-llm). Since its release in April 2023, our work has attracted significant attention from the AI research community.

We would be honored if you could include our work in the github repo (and a discussion in the next revision of your paper if it is possible). We believe that our framework complements the discussions in your work and could offer additional insights to the readers.

We thank you for considering our request and look forward to your positive response.