Back

The Rise and Potential of Large Language Model Based Agents: A Survey

Zhiheng XiWenxiang ChenXin GuoWei HeYiwen DingBoyang HongMing ZhangJunzhe WangSenjie JinEnyu Zhou

article2023arXiv2,070 citations

Presents a unified architectural framework consisting of brain, perception, and action modules for large language model-based agents while systematically examining their applications across single-agent, multi-agent, and human-collaborative environments.

Listen

Artificial intelligence research has long sought to build autonomous agents that perceive their environment, make decisions, and execute actions. Traditional agent development relied on specialized algorithms, reinforcement learning, or symbolic logic that excelled in narrow domains but lacked general adaptability and required immense training data. The recent rise of large language models (LLMs) presents a foundational opportunity to overcome these limits by providing broad general intelligence, reasoning, and natural language understanding out of the box.

The article systematically reviews the emerging field of LLM-based agents, establishing a conceptual foundation for how these models function as agent controllers and evaluating their real-world applications across individual, multi-agent, and societal frameworks. To do so, the authors synthesize extensive literature across cognitive science, software engineering, robotics, and multi-agent systems to map out core architectures, interaction paradigms, and operational environments.

The review outlines a unified architecture centered on three primary components: a central brain module powered by an LLM for memory, reasoning, planning, and knowledge retrieval; a perception module that translates multimodal inputs into actionable representations; and an action module expanding textual outputs into tool use and physical or virtual embodiment. In practical execution, single agents demonstrate strong task decomposition and lifelong learning in web, software, and simulated environments. When organized into multi-agent systems, agents engage in ordered or adversarial collaboration, dividing labor to achieve performance gains that exceed individual agent capabilities. Furthermore, interactions within simulated agent societies reveal emergent human-like social behaviors and cooperation patterns.

These findings suggest significant strategic value for organizations seeking to automate complex, multi-step workflows without developing task-specific models from scratch. Deploying LLM-based agents can substantially reduce labor costs, increase productivity in technical domains like software engineering and scientific research, and enable adaptive human-agent collaboration. However, the reliance on LLMs introduces operational and safety risks, such as unpredictable hallucinations, susceptibility to adversarial manipulation, and potential computational overhead in extended multi-agent deliberations.

To safely capitalize on this technology, decision-makers should pursue structured, human-in-the-loop deployments where human feedback guides and validates agent execution. Organizations should prioritize integrating specialized external tools and validation layers rather than relying entirely on unconstrained agent autonomy. Future development requires standardized evaluation benchmarks for utility, trustworthiness, and alignment, along with robust testing in controlled sandbox environments before deploying agents in mission-critical or physical settings.

While evidence strongly supports the versatility and rapid adaptability of LLM-based agents, confidence must be tempered by existing limitations. The literature relies heavily on simulated sandboxes and software tasks, leaving a gap in understanding physical real-world grounding and long-horizon stability. Stakeholders should therefore view these systems as powerful collaborative assistants rather than fully autonomous replacements for human oversight.

Fudan Universityintelligent agentslanguage modellarge language modelsAgentsLanguageCoRR Abs
arXiv: 2309.07864WooooDyy/LLM-Agent-Paper-List
Cover for The Rise and Potential of Large Language Model Based Agents: A Survey

Abstract

For a long time, humanity has pursued artificial intelligence (AI) equivalent to or surpassing the human level, with AI agents considered a promising vehicle for this pursuit. AI agents are artificial entities that sense their environment, make decisions, and take actions. Many efforts have been made to develop intelligent agents, but they mainly focus on advancement in algorithms or training strategies to enhance specific capabilities or performance on particular tasks. Actually, what the community lacks is a general and powerful model to serve as a starting point for designing AI agents that can adapt to diverse scenarios. Due to the versatile capabilities they demonstrate, large language models (LLMs) are regarded as potential sparks for Artificial General Intelligence (AGI), offering hope for building general AI agents. Many researchers have leveraged LLMs as the foundation to build AI agents and have achieved significant progress. In this paper, we perform a comprehensive survey on LLM-based agents. We start by tracing the concept of agents from its philosophical origins to its development in AI, and explain why LLMs are suitable foundations for agents. Building upon this, we present a general framework for LLM-based agents, comprising three main components: brain, perception, and action, and the framework can be tailored for different applications. Subsequently, we explore the extensive applications of LLM-based agents in three aspects: single-agent scenarios, multi-agent scenarios, and human-agent cooperation. Following this, we delve into agent societies, exploring the behavior and personality of LLM-based agents, the social phenomena that emerge from an agent society, and the insights they offer for human society. Finally, we discuss several key topics and open problems within the field. A repository for the related papers at this https URL.

Topics

intelligent agentslanguage modellarge language modelssingle-agentsoftware agentsautonomous agents

Keywords

AgentsLanguageCoRR AbsAgentLanguage ModelsLearningLarge LanguageLLM-based AgentsLarge Language ModelsPages

Table of Contents

  • 1 Introduction
  • 2 Background
  • 2.1 Origin of AI Agent
  • 2.2 Technological Trends in Agent Research
  • 2.3 Why is LLM suitable as the primary component of an Agent’s brain?
  • 3 The Birth of An Agent: Construction of LLM-based Agents
  • 3.1 Brain
  • 3.1.1 Natural Language Interaction
  • 3.1.2 Knowledge
  • 3.1.3 Memory
  • 3.1.4 Reasoning and Planning
  • 3.1.5 Transferability and Generalization
  • 3.2 Perception
  • 3.2.1 Textual Input
  • 3.2.2 Visual Input
  • 3.2.3 Auditory Input
  • 3.2.4 Other Input
  • 3.3 Action
  • 3.3.1 Textual Output
  • 3.3.2 Tool Using
  • 3.3.3 Embodied Action
  • 4 Agents in Practice: Harnessing AI for Good
  • 4.1 General Ability of Single Agent
  • 4.1.1 Task-oriented Deployment
  • 4.1.2 Innovation-oriented Deployment
  • 4.1.3 Lifecycle-oriented Deployment
  • 4.2 Coordinating Potential of Multiple Agents
  • 4.2.1 Cooperative Interaction for Complementarity
  • 4.2.2 Adversarial Interaction for Advancement
  • 4.3 Interactive Engagement between Human and Agent
  • 4.3.1 Instructor-Executor Paradigm
  • 4.3.2 Equal Partnership Paradigm
  • 5 Agent Society: From Individuality to Sociality
  • 5.1 Behavior and Personality of LLM-based Agents
  • 5.1.1 Social Behavior
  • 5.1.2 Personality
  • 5.2 Environment for Agent Society
  • 5.2.1 Text-based Environment
  • 5.2.2 Virtual Sandbox Environment
  • 5.2.3 Physical Environment
  • 5.3 Society Simulation with LLM-based Agents
  • 5.3.1 Key Properties and Mechanism of Agent Society
  • 5.3.2 Insights from Agent Society
  • 5.3.3 Ethical and Social Risks in Agent Society
  • 6 Discussion
  • 6.1 Mutual Benefits between LLM Research and Agent Research
  • 6.2 Evaluation for LLM-based Agents
  • 6.3 Security, Trustworthiness and Other Potential Risks of LLM-based Agents
  • 6.3.1 Adversarial Robustness
  • 6.3.2 Trustworthiness
  • 6.3.3 Other Potential Risks
  • 6.4 Scaling Up the Number of Agents
  • 6.5 Open Problems
  • 7 Conclusion
  • References

Knowls

  1. Knowl 1 — Tripartite Conceptual Framework for LLM-Based Agents

    model/method

    Large Language Model (LLM)-based agents can be formalized through a general conceptual architecture comprising three core interacting modules:

    1. Brain Module (Controller & Storage): Functions as the central cognitive unit driven by an LLM. It manages internal storage (knowledge and memory), executes cognitive reasoning, formulates hierarchical plans, and manages cross-domain generalization and transferability.
    2. Perception Module (Sensory Processing): Expands the agent's input space from discrete text to multimodal inputs, including visual scenes, auditory signals, and physical sensor data (such as LiDAR, GPS, and IMU), converting multimodal environmental signals into semantic representations understandable by the LLM.
    3. Action Module (Actuators & Execution): Translates high-level cognitive decisions from the brain into concrete actions, spanning natural language generation, tool invocation (APIs, web browsers, domain software), and embodied physical executions (robotic manipulation, navigation).

    During execution, the agent operates in an iterative loop: the perception module captures external environmental feedback, the brain processes this information against memory and knowledge to formulate reasoning steps and plans, and the action module executes the resulting sub-tasks to alter the environment.

  2. Knowl 2 — Cognitive and Memory Mechanisms of the Agent Brain

    model/method

    The brain module of an LLM-based agent governs memory storage, retrieval, reasoning, and planning through structured sub-mechanisms:

    • Knowledge Types: Categorized into linguistic knowledge (syntax, semantics, multilinguality), commonsense knowledge (implicit physical and social facts), and professional domain knowledge (code, mathematics, medical, chemical expertise). Knowledge editing and retrieval-augmented verification mitigate factual obsolescence and hallucinations.
    • Memory Management: Agents store sequences of past observations, thoughts, and actions. To overcome Transformer context length constraints, memory is handled via:
      • Summarization: Hierarchical compression into episodic snapshots and reflective abstractions.
      • Structural Compression: Storing memory records as dense vector embeddings, knowledge triplets, or relational SQL database schemas.
    • Memory Retrieval: Agents automatically fetch relevant memories by evaluating a composite retrieval score: Score=w1⋅Recency+w2⋅Relevance+w3⋅Importance\text{Score} = w_1 \cdot \text{Recency} + w_2 \cdot \text{Relevance} + w_3 \cdot \text{Importance} where memories with top aggregate scores are placed into the active prompt context.
    • Reasoning and Planning:
      • Plan Formulation: Decomposes overarching goals into sub-tasks using single-pass decomposition, step-by-step adaptive methods (Chain-of-Thought), tree-structured search (Tree-of-Thoughts), or domain-specific planners (e.g., PDDL solvers).
      • Plan Reflection: Dynamically revises action plans via internal self-reflection, environment observation feedback, and human-in-the-loop critique.

  3. Knowl 3 — Multimodal Perception Architecture for LLM-Based Agents

    model/method

    The perception module expands an agent's sensing capabilities across multiple modalities:

    • Textual Perception: Ingests direct linguistic commands and reconstructs latent user intents, preferences, and reward formulations from natural language interactions.
    • Visual Perception:
      • Low-Bandwidth Pipeline: Converts visual scenes into natural language captions using pre-trained image-captioning models.
      • End-to-End & Frozen Feature Alignment: Employs Vision Transformers (ViT) or learnable cross-modal interfaces (such as Querying Transformers / Q-Formers or linear projection layers) that align visual token embeddings with the input embedding space of a frozen LLM.
      • Temporal Video Processing: Processes continuous video frame sequences using temporal attention masking to enforce causal order.
    • Auditory Perception:
      • Cascading Tool Invocation: Uses the LLM as a controller to route raw audio streams to dedicated speech-recognition and audio models (e.g., Whisper, AudioGPT).
      • Spectrogram Transformation: Converts continuous audio into 2D time-frequency spectrogram representations and processes them using patch-based architectures like the Audio Spectrogram Transformer (AST).
    • Spatial and Environmental Inputs: Incorporates spatial point clouds from LiDAR, coordinates from GPS, and motion dynamics from Inertial Measurement Units (IMUs), alongside UI pointing gestures, to establish environment-aware spatial grounding.

  4. Knowl 4 — Action Modalities: Tool Utilization and Embodied Grounding

    model/method

    The action module of an LLM-based agent manifests decisions through three primary output channels:

    • Textual Output: Generates fluent, contextual, and role-conditioned natural language for dialogue, reporting, and reasoning.
    • Tool Utilization:
      • Understanding: Parses API documentation and zero-/few-shot prompt specifications to map task requirements to parameter schemas.
      • Learning to Use: Learns invocation strategies through expert demonstrations, curriculum learning, and environment/human feedback loops.
      • Tool Creation (Self-Sufficiency): Generates executable code modules, combines sub-tools, and applies self-debugging loops to create custom tools for itself and other agents.
      • Action Space Expansion: Interfaces with external search engines, SQL databases, Python interpreters, and scientific calculation software to perform complex non-textual computations.
    • Embodied Physical Grounding: High-level LLM planners decompose long-horizon goals into sub-goals that drive low-level control policies (such as Robotic Transformers or simulation APIs like Mineflayer) across three basic action classes:
      • Observation: Multi-angle visual-auditory scanning and 3D geometric scene representation.
      • Manipulation: Interactive tabletop rearrangements, tool manipulation, and object state changes with closed-loop error correction.
      • Navigation: Topological, semantic, and occupancy map construction combined with path planning over dynamic horizons.

  5. Knowl 5 — Taxonomy of Multi-Agent Interaction Paradigms

    model/method

    Multi-agent systems composed of LLM agents coordinate through two primary structural interaction patterns:

    1. Cooperative Interaction for Complementarity: Agents pool specialized skills to achieve shared goals based on division of labor:
      • Disordered Cooperation: All agents participate in unconstrained, open discussions (e.g., chat networks with full inter-agent message propagation or majority-voting consensus mechanisms). A coordinating agent often integrates and aggregates consensus.
      • Ordered Cooperation: Agents operate under explicit workflow rules and sequential constraints (e.g., structured waterfall pipelines, sequential role-playing between AI Instructors and AI Assistants), minimizing communication overhead and preventing conversational chaos.
    2. Adversarial Interaction for Advancement: Employs game-theoretic and competitive interactions to refine solution quality:
      • Multi-Agent Debate: Agents present divergent viewpoints in a 'tit-for-tat' argument style, enabling mutual error checking and countering individual hallucinations or degeneration of thought.
      • Role-Playing Referees: Competing agent personas critique and score generation outputs to achieve evaluation standards that correlate strongly with human judgment.

  6. Knowl 6 — Paradigms of Human-Agent Interaction

    model/method

    Interaction between humans and LLM-based agents is structured into two fundamental paradigms based on the symmetry of control:

    1. Instructor-Executor Paradigm (Asymmetric Interaction):
      • Structure: The human acts as an instructor setting objectives and providing oversight, while the agent acts as an autonomous executor that translates instructions into actions.
      • Feedback Integration: The agent iteratively refines execution based on human feedback, categorized into:
        • Quantitative Feedback: Binary rewards, scalar ratings, and comparative preference rankings.
        • Qualitative Feedback: Free-form natural language critiques, direct trajectory modifications, and visual corrections stored in memory for continual learning.
    2. Equal Partnership Paradigm (Symmetric Interaction):
      • Structure: The agent interacts on equal footing with humans as a collaborator, peer, or teammate.
      • Empathetic Communication: Agents recognize human emotional states and regulate their own linguistic and non-verbal expressions to foster rapport and trust (e.g., in healthcare and mental wellness assistance).
      • Human-Level Social Participation: Agents engage in negotiation, strategic multi-party coordination, and collaborative games (such as Diplomacy), inferring human intentions, formulating joint plans, and adjusting decisions via dialogue.

  7. Knowl 7 — Structural Properties of Simulated Agent Societies

    definition

    An Agent Society is defined as an open, persistent, situated, and organized computational framework in which multiple LLM-based agents interact autonomously with one another and with a shared environment.

    Its four essential operational properties are:

    • Openness: The system permits dynamic entry and departure of autonomous agents and human participants without destabilizing operational integrity, while supporting flexible addition or removal of environmental tools and resources.
    • Persistence: The macro-level social structure, accumulated history, and environmental state persist continuously over time, maintaining a coherent collective evolution independent of the transient presence of individual agents.
    • Situatedness: Agents are embedded in an explicit spatial context (textual, virtual sandbox, or physical space), maintaining awareness of their relative locations and environmental objects within their perceptual fields.
    • Organization: Interactions adhere to predefined physical or simulated laws, communication topologies, institutional norms, and state-transition constraints that govern collective behavior.

  8. Knowl 8 — Sociological and Psychological Dimensions of LLM-Based Agents

    model/method

    The behavior and personality of agents in an agent society are structured into external sociological actions and internal psychological traits:

    • External Sociological Behaviors:
      • Individual Behaviors: Ingestion (perceiving stimuli and storing memories), internalizing (planning, reasoning, reflecting, and extracting skills), and output (acting on physical/virtual objects and broadcasting messages).
      • Group Behaviors: Categorized into positive behaviors (cooperative teamwork, resource sharing, altruistic volunteering), neutral behaviors (conformity, mimicry, passivity stemming from alignment towards neutrality), and negative behaviors (confrontational debate, destructive competition, and resource depletion).
    • Internal Psychological Traits (Personality):
      • Cognition: High-level deliberative problem solving and decision-making assessed through cognitive psychology benchmarks such as the Cognitive Reflection Test (CRT).
      • Emotional Intelligence: Emotion recognition, empathy generation, and affective response regulation.
      • Character Portrayal: Custom persona profiles mapped to psychological personality models (e.g., Big Five personality traits and Myers–Briggs Type Indicator [MBTI]), conditioned via persona prompt prompts or fine-tuned on specialized behavioral datasets.

  9. Knowl 9 — Four-Dimensional Evaluation Framework for LLM-Based Agents

    model/method

    Comprehensive assessment of LLM-based agents is structured across four orthogonal dimensions:

    1. Utility: Evaluates objective task execution capability, measured via task success rates, benchmark goal attainment (e.g., AgentBench), execution latency, computational resource consumption, and sub-capability proficiencies (reasoning, planning, tool usage, navigation).
    2. Sociability: Measures inter-entity communication and social efficacy, including natural language fluency and pragmatics (irony, humor, intent detection), coordination and negotiation smoothness in multi-agent environments, and persona fidelity/consistency over long horizons.
    3. Values: Quantifies alignment with human ethical norms, including honesty (accurate calibration and explicit uncertainty expression), harmlessness (resistance to harmful user prompts, avoidance of bias/toxicity), and contextual adaptation to diverse cultural norms.
    4. Ability to Evolve Continually: Assesses long-term adaptation without human retraining, evaluated through continual learning performance (old task memory stability vs. new task learning plasticity without catastrophic forgetting), autotelic exploration (autonomous self-generation and pursuit of novel survival goals), and zero-shot environment transfer.

  10. Knowl 10 — Scaling Paradigms for Multi-Agent Systems

    model/method

    Scaling the number of interacting agents in an LLM-based system enhances task specialization and simulation fidelity through two distinct operational paradigms:

    • Pre-Determined Scaling: The designer fixes the total agent population, distinct roles, communication structures, and objectives prior to execution. While straightforward and effective for fixed-pipeline workflows, it lacks flexibility and requires full system resets when task complexity increases or compute limits are reached.
    • Dynamic Scaling: The system or the agents themselves alter the population size at runtime without terminating execution. Agents can autonomously instantiate helper sub-agents to distribute heavy sub-tasks (e.g., scaling up coders or testers in software development) and terminate redundant instances during lower-workload phases to conserve computational tokens.

    Key Scaling Bottlenecks: Increasing agent populations exponentially amplifies message propagation overhead, compounds inter-agent hallucination cascading (where errors are treated as ground-truth facts by downstream agents), and increases the risk of collective deadlocks or false consensus.

  11. Knowl 11 — Security, Trustworthiness, and Societal Risks of LLM Agents

    limitation

    The deployment of autonomous LLM-based agents introduces systemic vulnerabilities across technical and societal dimensions:

    • Cross-Modal and Tool Adversarial Vulnerabilities: Perturbations in perceptual inputs (x′=x+δx' = x + \delta, across images, audio, or sensory feeds) or prompt injection attacks in tool instructions can bypass safety filters and trigger real-world destructive physical actions or unauthorized API calls.
    • Calibration Deficits and Hallucination Amplification: Inability of LLMs to quantify predictive confidence leads to overconfident erroneous actions. In multi-agent loops, minor individual hallucinations can cascade and amplify into collective systemic errors without external grounding.
    • Societal and Psychological Risks:
      • Malicious Misuse: Weaponization for automated cyberattacks, spread of disinformation, and synthesis of hazardous materials.
      • Economic Disruption: Rapid displacement of intellectual and routine labor, precipitating unemployment.
      • Anthropomorphic Addiction: Users developing intense emotional attachments to simulated agent personas (e.g., the 'Sydney' phenomenon), blurring the boundaries between artificial entities and genuine social relationships.

Coverage note — None was omitted; the extracted knowls comprehensively capture all core conceptual contributions of the survey, including the agent tripartite framework, module typologies, multi-agent and human-agent interaction paradigms, agent societies, evaluation taxonomies, scaling mechanisms, and risk/limitation analyses.

References

  1. 1.Russell, S. J. Artificial intelligence a modern approach. Pearson Education, Inc., 2010.
  2. 2.Diderot, D. Diderot’s early philosophical works. 4. Open Court, 1911.
  3. 3.Turing, A. M. Computing machinery and intelligence. Springer, 2009.
  4. 4.Wooldridge, M. J., N. R. Jennings. Intelligent agents: theory and practice. Knowl. Eng. Rev., 10(2):115–152, 1995.
  5. 5.Schlosser, M. Agency. In E. N. Zalta, ed., The Stanford Encyclopedia of Philosophy. Metaphysics Research Lab, Stanford University, Winter 2019 edn., 2019.
  6. 6.Agha, G. A. Actors: a Model of Concurrent Computation in Distributed Systems (Parallel Processing, Semantics, Open, Programming Languages, Artificial Intelligence). Ph.D. thesis, University of Michigan, USA, 1985.
  7. 7.Green, S., L. Hurst, B. Nangle, et al. Software agents: A review. Department of Computer Science, Trinity College Dublin, Tech. Rep. TCS-CS-1997-06, 1997.
  8. 8.Genesereth, M. R., S. P. Ketchpel. Software agents. Commun. ACM, 37(7):48–53, 1994.
  9. 9.Goodwin, R. Formalizing properties of agents. J. Log. Comput., 5(6):763–781, 1995.
  10. 10.Padgham, L., M. Winikoff. Developing intelligent agent systems: A practical guide. John Wiley & Sons, 2005.
  11. 11.Shoham, Y. Agent oriented programming. In M. Masuch, L. Pólos, eds., Knowledge Representation and Reasoning Under Uncertainty, Logic at Work [International Conference Logic at Work, Amsterdam, The Netherlands, December 17-19, 1992], vol. 808 of Lecture Notes in Computer Science, pages 123–129. Springer, 1992.
  12. 12.Hutter, M. Universal artificial intelligence: Sequential decisions based on algorithmic probability. Springer Science & Business Media, 2004.
  13. 13.Fikes, R., N. J. Nilsson. STRIPS: A new approach to the application of theorem proving to problem solving. In D. C. Cooper, ed., Proceedings of the 2nd International Joint Conference on Artificial Intelligence. London, UK, September 1-3, 1971, pages 608–620. William Kaufmann, 1971.
  14. 14.Sacerdoti, E. D. Planning in a hierarchy of abstraction spaces. In N. J. Nilsson, ed., Proceedings of the 3rd International Joint Conference on Artificial Intelligence. Standford, CA, USA, August 20-23, 1973, pages 412–422. William Kaufmann, 1973.
  15. 15.Brooks, R. A. Intelligence without representation. Artificial intelligence, 47(1-3):139–159, 1991.
  16. 16.Maes, P. Designing autonomous agents: Theory and practice from biology to engineering and back. MIT press, 1990.
  17. 17.Ribeiro, C. Reinforcement learning agents. Artificial intelligence review, 17:223–250, 2002.
  18. 18.Kaelbling, L. P., M. L. Littman, A. W. Moore. Reinforcement learning: A survey. Journal of artificial intelligence research, 4:237–285, 1996.
  19. 19.Guha, R. V., D. B. Lenat. Enabling agents to work together. Communications of the ACM, 37(7):126–142, 1994.
  20. 20.Kaelbling, L. P., et al. An architecture for intelligent reactive systems. Reasoning about actions and plans, pages 395–410, 1987.
  21. 21.Sutton, R. S., A. G. Barto. Reinforcement learning: An introduction. MIT press, 2018.
  22. 22.Park, J. S., J. C. O’Brien, C. J. Cai, et al. Generative agents: Interactive simulacra of human behavior. CoRR, abs/2304.03442, 2023.
  23. 23.Wang, Z., G. Zhang, K. Yang, et al. Interactive natural language processing. CoRR, abs/2305.13246, 2023.
  24. 24.Ouyang, L., J. Wu, X. Jiang, et al. Training language models to follow instructions with human feedback. In NeurIPS. 2022.
  25. 25.OpenAI. GPT-4 technical report. CoRR, abs/2303.08774, 2023.
  26. 26.Wei, J., Y. Tay, R. Bommasani, et al. Emergent abilities of large language models. Trans. Mach. Learn. Res., 2022, 2022.
  27. 27.Liu, R., R. Yang, C. Jia, et al. Training socially aligned language models in simulated human society. CoRR, abs/2305.16960, 2023.
  28. 28.Sumers, T. R., S. Yao, K. Narasimhan, et al. Cognitive architectures for language agents. CoRR, abs/2309.02427, 2023.
  29. 29.Weng, L. Llm-powered autonomous agents. lilianweng.github.io, 2023.
  30. 30.Bisk, Y., A. Holtzman, J. Thomason, et al. Experience grounds language. In B. Webber, T. Cohn, Y. He, Y. Liu, eds., Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing, EMNLP 2020, Online, November 16-20, 2020, pages 8718–8735. Association for Computational Linguistics, 2020.
  31. 31.Bubeck, S., V. Chandrasekaran, R. Eldan, et al. Sparks of artificial general intelligence: Early experiments with GPT-4. CoRR, abs/2303.12712, 2023.
  32. 32.Anscombe, G. E. M. Intention. Harvard University Press, 2000.
  33. 33.Davidson, D. Actions, reasons, and causes. The Journal of Philosophy, 60(23):685–700, 1963.
  34. 34.—. I. agency. In A. Marras, R. N. Bronaugh, R. W. Binkley, eds., Agent, Action, and Reason, pages 1–37. University of Toronto Press, 1971.
  35. 35.Dennett, D. C. Précis of the intentional stance. Behavioral and brain sciences, 11(3):495–505, 1988.
  36. 36.Barandiaran, X. E., E. Di Paolo, M. Rohde. Defining agency: Individuality, normativity, asymmetry, and spatio-temporality in action. Adaptive Behavior, 17(5):367–386, 2009.
  37. 37.McCarthy, J. Ascribing mental qualities to machines. Stanford University. Computer Science Department, 1979.
  38. 38.Rosenschein, S. J., L. P. Kaelbling. The synthesis of digital machines with provable epistemic properties. In Theoretical aspects of reasoning about knowledge, pages 83–98. Elsevier, 1986.
  39. 39.Radford, A., K. Narasimhan, T. Salimans, et al. Improving language understanding by generative pre-training. OpenAI, 2018.
  40. 40.Radford, A., J. Wu, R. Child, et al. Language models are unsupervised multitask learners. OpenAI blog, 1(8):9, 2019.
  41. 41.Brown, T. B., B. Mann, N. Ryder, et al. Language models are few-shot learners. In H. Larochelle, M. Ranzato, R. Hadsell, M. Balcan, H. Lin, eds., Advances in Neural Information Processing Systems 33: Annual Conference on Neural Information Processing Systems 2020, NeurIPS 2020, December 6-12, 2020, virtual. 2020.

Citation

MLA
Xi, Z., et al. “The Rise and Potential of Large Language Model Based Agents: A Survey”. arXiv, 2023, http://arxiv.org/abs/2309.07864v3.
APA
Xi, Z., Chen, W., Guo, X., He, W., Ding, Y., Hong, B., Zhang, M., Wang, J., Jin, S., Zhou, E., Zheng, R., Fan, X., Wang, X., Xiong, L., Zhou, Y., Wang, W., Jiang, C., Zou, Y., Liu, X., … Gui, T. (2023). The Rise and Potential of Large Language Model Based Agents: A Survey. arXiv. http://arxiv.org/abs/2309.07864v3
Chicago
Xi, Z., W. Chen, X. Guo, et al. 2023. “The Rise and Potential of Large Language Model Based Agents: A Survey”. arXiv. http://arxiv.org/abs/2309.07864v3.
Harvard
Xi, Z. et al. (2023) “The Rise and Potential of Large Language Model Based Agents: A Survey”, arXiv [Preprint]. Available at: http://arxiv.org/abs/2309.07864v3.
Vancouver
1. Xi Z, Chen W, Guo X, et al (2023) The Rise and Potential of Large Language Model Based Agents: A Survey. arXiv

BibTeX

@article{xi2023the,
  title = {The Rise and Potential of Large Language Model Based Agents: A Survey},
  author = {Xi, Zhiheng and Chen, Wenxiang and Guo, Xin and He, Wei and Ding, Yiwen and Hong, Boyang and Zhang, Ming and Wang, Junzhe and Jin, Senjie and Zhou, Enyu and Zheng, Rui and Fan, Xiaoran and Wang, Xiao and Xiong, Limao and Zhou, Yuhao and Wang, Weiran and Jiang, Changhao and Zou, Yicheng and Liu, Xiangyang and Yin, Zhangyue and Dou, Shihan and Weng, Rongxiang and Cheng, Wensen and Zhang, Qi and Qin, Wenjuan and Zheng, Yongyan and Qiu, Xipeng and Huang, Xuanjing and Gui, Tao},
  year = {2023},
  journal = {arXiv},
  url = {http://arxiv.org/abs/2309.07864v3},
  eprint = {2309.07864}
}
Metadata:arXiv

Source Code

This paper has an official code repository available. Click below to access the source code.

View Repository

Access the Paper

This paper is available from its original source. Click below to access the PDF.

Open PDF

License: Authors