Elamur Research
☰
home / papers / 2023 / A Survey on Large Language Model based Autonomous Agents

A Survey on Large Language Model based Autonomous Agents

Written for the Research library (claude-opus-5-5) from the paper text. Original — arxiv.org/pdf/2308.11432

Summary

Classic autonomous agents were usually trained with simple policies inside small, closed environments, so they struggled to act in a humanlike way in open-ended settings. Large language models (LLMs) carry broad world knowledge from web-scale training, and this led to a fast-growing body of work that uses them as the core controller of an agent. This survey by Wang et al. reviews that work from three angles: how such agents are built, where they are used, and how they are evaluated. For building agents, the authors propose a single architecture made of four modules (profiling, memory, planning and action). They also sort methods for giving agents task skills by whether the underlying LLM is fine-tuned. The survey then covers uses in the social sciences, natural sciences and engineering. It compares subjective and objective ways of evaluating agents. It ends with open problems. Because most systems in this area were built in isolation, the survey's common taxonomy is a useful reference for anyone designing or comparing LLM-driven agents.

Key points

  • Four-module framework: The profiling module sets the agent's role. The memory and planning modules let the agent use past experience and plan what to do next. The action module turns those decisions into outputs. Profiling shapes memory and planning, and all three feed into action.
  • Ways to create agent profiles: profiles can be written by hand (as in Generative Agent, MetaGPT and ChatDev), generated by an LLM, or aligned with real datasets. One example of the last approach gives GPT-3 personas based on demographic data from ANES respondents.
  • Memory and planning: Memory is described by its structure (unified or hybrid), its format (natural language, databases, embeddings or lists) and its operations (reading, writing and reflection). Planning is split into approaches that use feedback and those that don't. Feedback can come from the environment, a human or a model.
  • Scope: A growth chart counts papers published between January 2021 and August 2023. It groups them into categories such as tool, embodied, game, web, simulation and assistant agents (for example AutoGPT, Voyager and ToolBench).
  • Open challenges: These include alignment that still lets simulations reproduce negative human behaviour, fragile prompt frameworks, hallucination, LLMs knowing more than the humans they are meant to simulate, and slow inference when each action needs many LLM calls.

Our summary of the paper, in our own words — not its text. The work belongs to its authors; read it at the links above.

Fields: AI & machine learning

Cited by

443 papers of our library cite it (of the 140673 whose bibliographies we read); 18 of them have a page here, 425 not yet.

Cite

The paper itself, not this page. BibTeX from the library's record of the work. DOI 10.1007/s11704-024-40231-1 · arXiv 2308.11432

@misc{wang2023survey,
  title = {{A Survey on Large Language Model based Autonomous Agents}},
  author = {Wang, Lei and Ma, Chen and Feng, Xueyang and Zhang, Zeyu and Yang, Hao and Zhang, Jingsen and Chen, Zhiyuan and Tang, Jiakai and Chen, Xu and Lin, Yankai and Zhao, Wayne Xin and Wei, Zhewei and Wen, Ji-Rong},
  year = {2023},
  doi = {10.1007/s11704-024-40231-1},
  eprint = {2308.11432},
  archivePrefix = {arXiv},
  url = {https://doi.org/10.1007/s11704-024-40231-1}
}

References

11 references, 3 of them with a card of ours. ↗ leads to the work, «find» searches for one we have no id for yet.

  1. 1. Mnih V, Kavukcuoglu K, Silver D, Rusu A A, Veness J, Bellemare M G, Graves A, Riedmiller M, Fidjeland A K, Ostrovski G, others . Human-level control through deep reinforcement learning. nature, 2015, 518(7540): 529–533 2. Lillicrap T P, Hunt J J, Pritzel A, Heess N, Erez T, Tassa Y, Silver D, Wie… find
  2. Political Analysis, 2023, 31(3): 337–351 Fischer K A. Reflective linguistic programming (rlp): A stepping stone in socially-aware agi (socialagi). arXiv preprint arXiv:2305.12647, 2023 Rana K, Haviland J, Garg S, Abou-Chakra J, Reid I, Suenderhauf N. Sayplan: Grounding large language models using 3d… find our card
  3. A generative agent with fast and slow thinking for complex interactive tasks. Advances in Neural Information Processing Systems, 2024, 36 Evans J S B, Stanovich K E. Dual-process theories of higher cognition: Advancing the debate. Perspectives on psychological science, 2013, 8(3): 223–241 Wang X, We… find our card
  4. Intelligent Computing, 2024, 3: 0063 Karpas E, Abend O, Belinkov Y, Lenz B, Lieber O, Ratner N, Shoham Y, Bata H, Levine Y, Leyton-Brown K, others . Mrkl systems: A modular, neuro-symbolic architecture that combines large language models, external knowledge sources and discrete reasoning. arXiv prep… find our card
  5. Science, 2023, 380(6650): 1108–1109 Zhou X, Li G, Liu Z. Llm as dba. arXiv preprint arXiv:2308.05481, 2023 He Z, Wu H, Zhang X, Yao X, Zheng S, Zheng H, Yu B. Chateda: A large language model powered autonomous agent for eda. In: 2023 ACM/IEEE 5th Workshop on Machine Learning for CAD (MLCAD). 2023, 1… find
  6. Computational Linguistics, 2024, 1–58 Li C, Wang J, Zhu K, Zhang Y, Hou W, Lian J, Xie X. Emotionprompt: Leveraging psychology for large language models enhancement via emotional stimulus. arXiv e-prints, 2023, arXiv–2307 Zhuo T Y, Li Z, Huang Y, Shiri F, Wang W, Haffari G, Li Y F. On robustness of… find
  7. ACM Computing Surveys, 2023, 55(12): 1–38 find
  8. Jingsen Zhang is currently studying for a Ph.D. degree at Renmin University of China, Beijing, China. His research interests include recommender system. Lei Wang is a Ph.D. candidate at Renmin University of China, Beijing. His research focuses on recommender systems and agent-based large language mo… find
  9. Yankai Lin received his B.E. and Zhewei Wei received his Ph.D. of Ph.D. degrees from Tsinghua Uni-Computer Science and Engineer-versity in 2014 and 2019. find
  10. Af-ing from Hong Kong University of ter that, he worked as a senior re-Science and Technology. He did searcher in Tencent WeChat, and postdoctoral research in Aarhus joined Renmin University of China University from 2012 to 2014, and in 2022 as a tenure-track assistant professor. His main joined Ren… find
  11. Wayne Xin Zhao received his Ph.D. in Computer Science from Peking Ji-Rong Wen is a full professor, the University in 2014. His research executive dean of Gaoling School interests include data mining, natu-of Artificial Intelligence, and the ral language processing and infor-dean of School of Informa… find