Fudan NLP Group - The Rise and Potential of Large Language Model Based Agents: A Survey
Here are the contents translated into English:
Concept and Social Applications of Agent
- The author notes that large language models have established a common sense consensus and ability to continuously learn and understand, but agents like humans have different cognition and abilities.
Extrinsic Social Behavior
- From an extrinsic social behavior perspective, agents can manifest as:
- Individual behavior: the basis for agent operation and development, including perception, action, and internalization.
- Group behavior: the behavior produced by two or more agents interacting voluntarily, including positive behaviors (such as cooperation), negative behaviors (such as conflict), and neutral behaviors (such as observation).
Intrinsic Personality
- From an intrinsic personality perspective, agents can manifest as:
- Cognitive ability: involving the process of acquiring and understanding knowledge.
- Emotional ability: primarily involving subjective feelings and emotional states, such as joy, anger, and sadness, as well as the ability to show empathy and compassion.
- Personality analysis: used to understand and analyze the personality characteristics of large language models, using mature evaluation methods such as the Big Five Personality or MBTI test.
Environment
- The author believes that a simulated society should have openness, persistence, scenario-ness, and organizationality.
- Openness: allowing agents to freely enter and exit the simulated society.
- Persistence: ensuring that the society has a continuous trajectory over time.
- Scenario-ness: emphasizing the existence and operation of subjects in specific environments.
- Organizationality: ensuring that the simulated society has rules and limitations similar to those in the physical world.
Significance
- Agents can be used to explore the boundaries of collective intelligence, such as simulating social networks to observe phenomena in dissemination studies or simulating policies’ impact on society to assist decision-making, etc.
- However, this type of simulation also poses certain risks, including:
- Harmful social phenomena
- Stereotypes and prejudices
- Privacy security issues
- Over-reliance and addiction
Prospective Open-Ended Questions
- How should the research on intelligent agents and large language models be mutually promoted to develop together?
- What are the implications of big models showing great potential in language understanding, decision-making, and generalization for building agents, and vice versa?
- What challenges and concerns will intelligent agents bring, such as illegal use, job displacement risks, and impacts on human well-being?
- As the number of agents increases, what opportunities and challenges will arise, including decreased system efficiency and information misinterpretation issues?
In summary, this paper reviews the concept of agent based on large language models and its social applications, discussing their significance, risks, and future development directions.
Translation
本文主要讨论了基于大语言模型的代理-agent的概念及其社会层面的应用。作者指出，大语言模型已经为我们建立一个社会底层的通识共识和不断学习理解的能力，但代理-agent却如人类个体一样，有不同的认知和能力。
外在社会行为
从外在的社会行为来看，代理-agent可以表现为：
- 个体行为：构成了代理自身运作和发展的基础，包括感知、行动以及内化行为。
- 群体行为：是指两个以上代理自发交互时产生的行为，包括积极行为（如协作为代表）、消极行为（如冲突为代表）及中立行为（如从众旁观）。
内在人格
从内在人格来看，代理-agent可以表现为：
- 认知能力：涉及获取和理解知识的过程。
- 情感能力：主要涉及主观感受和情绪状态，如喜怒哀乐，以及表现出同情和共情的能力。
- 性格分析：用于理解和分析大语言模型的性格特征，使用成熟的评测方式如大五人格或MBTI测试。
环境
作者认为一个模拟社会应当具有开放性、持久性、情景性和组织性。
- 开放性：允许代理自主地进出模拟社会。
- 持久性：是指社会具有随着时间的推移而发展的连贯轨迹。
- 情景性：强调主体在特定环境中的存在和运作。
- 组织性：确保模拟社会拥有类似于物理世界的规则与限制。
意义
代理社会可以用来探索群体智能的能力边界，例如通过模拟社交网络来观察传播学的现象，或是模拟政策对社会的影响来辅助决策等。然而，这种模拟社会也可能存在一定的风险，如有害的社会现象、刻板印象和偏见、隐私安全问题以及过度依赖与成瘾性。
前瞻开放性的问题
作者讨论了几项前瞻开放性的问题，例如：
- 智能代理与大语言模型的研究应该如何互相促进共同发展。
- 大模型在语言理解决策制定以及泛化能力等方面展现出的强大潜力是构建代理的关键角色，而代理的进展也为大模型提出了更高的要求。
- 智能代理会带来哪些挑战与隐忧，例如非法滥用、失业风险以及对人类福祉造成影响等。
- 随着代理数量的不断提升会带来哪些机遇和挑战，包括模拟系统效率的降低和信息失真误解的问题。
总之，这篇论文综述了基于大语言模型的代理-agent的概念及其社会层面的应用，探讨了其意义、风险以及未来发展方向。
Reference:
https://arxiv.org/pdf/2309.07864