LLM based agents: a survey
Here are the contents of the document in English:
The Significance of Simulated Societies: The 25-person town from Stanford University is a vivid example that shows how agent societies can be used to explore the boundaries of collective intelligence and accelerate social science research.
Risks and Challenges: The author discusses the risks present in agents based on large language models, such as harmful social phenomena, stereotypes and biases, privacy security issues, and addiction-like behavior.
Open-Ended Questions: The paper explores open-ended questions, including how research on intelligent agents and large language models can mutually promote each other’s development, the potential of large models in understanding and generalizing language, and whether intelligent agents can truly be implemented and avoid harming the real world.
Opportunities and Challenges: As the number of agents increases, it will bring more opportunities and challenges, such as improving the credibility and authenticity of simulated societies by increasing the number of individual agents, but also making communication and message transmission problems more complex.
The Debate on AGI: The paper discusses the debate about whether agents are a path to Artificial General Intelligence (AGI), including whether self-supervised language modeling can truly show intelligence, and how to coordinate individual agents to overcome groupthink and cognitive biases.
Translation
这篇论文探讨了基于大语言模型的代理在模拟社会中的应用和潜在风险。作者提出了几个重要观点：
- 模拟社会的意义：斯坦福大学的25人小镇是一个生动的例子，展示了代理社会可以用来探索群体智能的能力边界以及加速社会科学研究。
- 风险和挑战：作者讨论了存在于基于大语言模型的代理中的风险，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等等。
- 前瞻开放性的问题：论文探讨了一些开放性的问题，例如智能代理与大语言模型的研究如何互相促进共同发展，大模型在语言理解决策制定和泛化能力方面的潜力，以及智能代理能否真正落地并避免对真实世界带来的危害。
- 机遇和挑战：随着代理数量的不断提升，将带来更多的机会和挑战，如模拟社会中提升代理的个体数量可以显著提升可信度与真实性，但也会导致通信和消息传播问题变得复杂。
- 关于AGI的争论：论文提到了关于代理是否是通向AGI道路的争论，包括是否自回归语言建模可以显现真正的智能，以及如何协调单个代理以克服团体迷思和个体认知偏差。
总之，这篇论文探讨了基于大语言模型的代理在模拟社会中的应用和潜在风险，并提出了许多开放性的问题供进一步研究和讨论。
Reference:
https://arxiv.org/pdf/2309.07864