Source-linked AI summary
A Survey on Large Language Model based Autonomous Agents
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
TL;DR
Earlier autonomous agents were often trained in restricted environments, limiting their fit with open-domain human-like decision processes. This survey unifies research on LLM-based agents across construction, applications, and evaluation, while identifying challenges and future directions.
Problem
Prior agents commonly relied on heuristic policies and isolated environments, diverging from human learning and limiting human-level decisions in unconstrained settings.
Method
The survey proposes a unified framework and taxonomies covering LLM-agent construction, applications, evaluation strategies, capability acquisition, and future directions.
Results
The survey systematically reviews LLM-based autonomous-agent research across construction, applications, and evaluation, organizing major techniques and their development histories.
Takeaways & Limitations
The survey provides newcomers with comprehensive background knowledge and identifies challenges intended to guide future research on LLM-based autonomous agents.
Takeaways & Limitations
LLM-based human simulations may be distorted because models possess substantially more web knowledge than average individuals.
Abstract
from arXiv · showhide
Autonomous agents have long been a prominent research focus in both academic and industry communities. Previous research in this field often focuses on training agents with limited knowledge within isolated environments, which diverges significantly from human learning processes, and thus makes the agents hard to achieve human-like decisions. Recently, through the acquisition of vast amounts of web knowledge, large language models (LLMs) have demonstrated remarkable potential in achieving human-level intelligence. This has sparked an upsurge in studies investigating LLM-based autonomous agents. In this paper, we present a comprehensive survey of these studies, delivering a systematic review of the field of LLM-based autonomous agents from a holistic perspective. More specifically, we first discuss the construction of LLM-based autonomous agents, for which we propose a unified framework that encompasses a majority of the previous work. Then, we present a comprehensive overview of the diverse applications of LLM-based autonomous agents in the fields of social science, natural science, and engineering. Finally, we delve into the evaluation strategies commonly used for LLM-based autonomous agents. Based on the previous studies, we also present several challenges and future directions in this field. To keep track of this field and continuously update our survey, we maintain a repository of relevant references at https://github.com/Paitesanshi/LLM-Agent-Survey.
1 Introduction
The paper surveys LLM-based autonomous agents as a response to limitations of earlier agents learned in isolated environments, organizing the field around construction, applications, and evaluation. It also establishes taxonomies, identifies challenges, and discusses future directions.
- Motivation: Earlier autonomous agents relied on simple heuristic policies and isolated environments, making them unlike human learning and far from replicating human intelligence.The authors contrast these limitations with human learning from a wider variety of environments.
- Motivation: LLMs have motivated autonomous-agent research by providing broad world knowledge, human-like intelligence potential, natural-language interaction, flexibility, and explainability.LLM-based agents can perform informed actions without training on specific domain data.
- Survey scope: The paper proposes a unified perspective on agent construction and systematically reviews the rapidly developing field to support comprehensive understanding and future research.The survey frames architecture as hardware fundamentals and capabilities as software resources.
- Survey scope: The survey organizes research on LLM-based autonomous agents around construction, application, and evaluation.For construction, it examines agent architecture and methods for enhancing capabilities across tasks.
- Contributions: The survey establishes comprehensive taxonomies, identifies challenges, and discusses potential future directions for LLM-based autonomous agents.These aims are intended to provide newcomers with an understanding of the field.
2 LLM-based Autonomous Agent Construction
LLM-based autonomous agent construction centers on designing architectures that exploit LLM capabilities and enabling agents to acquire task-specific abilities. The proposed unified framework organizes architecture design around profiling, memory, planning, and action modules.
- Design motivation: Constructing autonomous agents requires more than question answering because agents must fulfill roles, perceive and learn from environments, and evolve like humans.Rational agent architectures are therefore needed to help LLMs maximize their capabilities.
- Unified architecture: The framework comprises profiling, memory, planning, and action modules that respectively define roles, support recall and future planning, and translate decisions into outputs.Together, these modules connect role specification, environmental interaction, and action generation.
- Profiling module: Agent profiles can be handcrafted, generated by LLMs, or aligned with real-world datasets to shape roles, attributes, personalities, and relationships.The appropriate profile information depends on the application scenario, such as emphasizing psychology for studies of human cognition.
- Memory module: The memory module stores environmental information and recorded experiences to support future actions, self-evolution, and more consistent behavior.The section organizes memory design around structures, formats, and operations.
- Memory module: Memory structures draw on human memory by modeling short-term memory, long-term memory, or both through unified and hybrid designs.Unified memory relies on in-context prompts, whereas hybrid memory buffers recent perceptions and consolidates important information over time.
3 LLM-based Autonomous Agent Application
LLM-based autonomous agents are applied across social science, natural science, and engineering, leveraging language comprehension, reasoning, common-sense understanding, and tool use. Representative applications include psychological simulation and mental-health support, social and political modeling, scientific documentation and experimentation, education, and software engineering.
- Social science: In psychology, LLM-based agents reproduce results aligned with human-participant studies, while larger models tend to deliver more “hyper-accuracy distortion.”They are also used for simulation experiments and mental-health support.
- Social science: 120 Reddit posts showed conversation agents can help users cope with anxiety, social isolation, and depression on demand, but may sometimes produce harmful content.The study systematically analyzed their effectiveness for mental well-being support.
- Social science: LLM-based agents simulate social networks and virtual towns to study harmful-information propagation, behavioral effects, community regulation, and daily human life.Examples include Social Simulacra, Generative Agents, AgentSims, and related social-network simulations.
- Natural science: In natural science, agents support documentation, data management, question answering, experiment planning, experimental design, execution, and scientific education through internet and database tools.ChemCrow additionally recommends experimental procedures while emphasizing potential safety risks.
- Computer Science & Software Engineering: In software engineering, LLM-based agents automate coding, testing, debugging, and documentation, with multi-agent frameworks collaborating through natural-language conversations across the software development life cycle.ChatDev demonstrates efficient and cost-effective generation of executable software systems.
4 LLM-based Autonomous Agent Evaluation
LLM-based autonomous agents are evaluated through subjective human judgments and objective quantitative methods. Subjective approaches capture human-centered qualities, while objective evaluation organizes metrics, protocols, and benchmarks, and the two may be combined for comprehensive assessment.
- Subjective Evaluation: Subjective evaluation uses human judgments when datasets or quantitative metrics are unavailable or difficult to design, including intelligence and user-friendliness.It is especially suitable for assessing capabilities that are difficult to quantify directly.
- Subjective Evaluation: Human annotation directly scores or ranks agent outputs, whereas the Turing test asks evaluators to distinguish agent-generated outputs from human outputs.The Turing test treats indistinguishable results as evidence of human-like performance on the evaluated task.
- Subjective Evaluation: Subjective evaluation reflects human criteria but faces high costs, inefficiency, and population bias, motivating LLMs as intermediaries for assessment.Examples include GPT-based assessment in ChemCrow and structured multi-agent critique in ChatEval.
- Objective Evaluation: Objective evaluation uses computable quantitative metrics and comprises three aspects: evaluation metrics, protocols, and benchmarks.Representative task-success metrics include success rate, reward/score, coverage, and accuracy/error rate; protocols include real-world simulation, social evaluation, multi-task evaluation, and software testing.
- Objective Evaluation: Benchmarks evaluate agents across interactive, task-oriented, end-to-end, tool-use, and multi-agent settings, including ALFWorld, IGLU, Minecraft, WebArena, GentBench, and RocoBench.Objective evaluation provides quantitative insights but cannot yet perfectly measure every agent capability, so combining subjective and objective strategies may be more comprehensive.
5 Related Surveys
Existing surveys cover LLM background, technologies, applications, deployment challenges, alignment, evaluation, and downstream-task capabilities, but none specifically focus on the rapidly emerging field of LLM-based agents. This study addresses that gap by compiling 100 relevant works on LLM-based agents.
- Survey coverage: Prior surveys examine LLM backgrounds, mainstream technologies, downstream applications, deployment challenges, alignment, evaluation, and capabilities or limitations.These works span training, application, and evaluation of large models.
- Research gap: Before this paper, no work had specifically focused on the rapidly emerging and promising field of LLM-based agents.The paper positions its survey as addressing this gap in prior literature.
- Contribution: The study compiles 100 relevant works on LLM-based agents.This compilation supports the paper’s focused survey of the emerging field.
6 Challenges
LLM-based autonomous agents remain at an early stage, facing challenges in role-playing, human alignment, prompt robustness, hallucination, believable simulation, and action efficiency. Addressing these issues requires improved training, prompting, architectures, behavioral constraints, and inference efficiency.
- Role-playing: Agents must accurately play specialized roles, but existing LLMs may inadequately model uncommon roles and human cognitive psychology, resulting in limited self-awareness.Proposed approaches include fine-tuning on real-human data and designing tailored prompts or architectures, though preserving performance on common roles remains challenging.
- Human alignment: Agent-based simulation requires generalized human alignment so agents can reflect diverse human values across purposes and applications.Existing powerful LLMs, including ChatGPT and GPT-4, are described as mostly aligned with unified human values; prompting strategies may enable realignment.
- Prompt robustness: Memory and planning modules increase prompt complexity, while minor prompt alterations can produce substantially different outcomes and undermine robust agent behavior.Supplementary modules are embedded to support rational behavior, consistent operation, and effective communication.
- Hallucination: Hallucination can cause autonomous agents to produce false information confidently, leading to incorrect or misleading code, security risks, and ethical issues.The problem has been observed during code generation when agents receive simplistic instructions.
- Human simulation: Believable human simulations must constrain LLM use of knowledge that real-world users would not have access to beforehand.Without appropriate strategies, agents may make decisions using extensive knowledge unavailable to simulated users.
- Action efficiency: Slow autoregressive LLM inference greatly affects agent-action efficiency because agents may query models multiple times for memory retrieval, planning, and other actions.The need for repeated queries compounds the impact of inference speed on overall action efficiency.
7 Conclusion
The survey systematically reviews LLM-based autonomous-agent research through taxonomies of agent construction, applications, and evaluation, while also proposing challenges in the field.
- 7 Conclusion: The survey organizes research on LLM-based autonomous agents around construction, application, and evaluation.For each aspect, it provides a detailed taxonomy connecting existing studies and summarizing major techniques and their development histories.
- 7 Conclusion: The survey also proposes several challenges for LLM-based autonomous agents.