Skip to main content
QUICK REVIEW

[论文解读] Exploring the Intersection of Large Language Models and Agent-Based Modeling via Prompt Engineering

Edward Junprung|arXiv (Cornell University)|Aug 14, 2023
Topic Modeling被引用 9
一句话总结

论文展示了通过提示工程实现的基于大语言模型的代理人仿真(两人谈判与六人谋杀悬疑),分析可行性并讨论诸如上下文窗口瓶颈等局限性。

ABSTRACT

The final frontier for simulation is the accurate representation of complex, real-world social systems. While agent-based modeling (ABM) seeks to study the behavior and interactions of agents within a larger system, it is unable to faithfully capture the full complexity of human-driven behavior. Large language models (LLMs), like ChatGPT, have emerged as a potential solution to this bottleneck by enabling researchers to explore human-driven interactions in previously unimaginable ways. Our research investigates simulations of human interactions using LLMs. Through prompt engineering, inspired by Park et al. (2023), we present two simulations of believable proxies of human behavior: a two-agent negotiation and a six-agent murder mystery game.

研究动机与目标

  • 将ABM与LLMs结合以建模超越传统代理规则的人类驱动交互的动机。
  • 通过塑造代理角色与提示来介绍LLM驱动的仿真以研究结果。
  • 将LLM驱动的仿真分为一对一、一对多和多对多三类。
  • 强调诸如4,096标记上下文窗口等瓶颈并讨论潜在改进。

提出的方法

  • 使用 OpenAI GPT-3.5-turbo 构建具有明确人设的对话式LLM代理。
  • 实现自主的往返式提问,代理在彼此的累计对话历史中作出响应。
  • 采用轮换交互机制以维持对话连贯性。
  • 利用记忆流在群体仿真中跨代理保持上下文。
  • 分析人设设计如何影响涌现策略与结果。
  • 报告最终提示标记计数并讨论可扩展性约束。
Figure 1: Autonomous dialog with back-and-forth prompting mechanism. Each box represents a prompt that is passed to LLM agents in a round-robin fashion.
Figure 1: Autonomous dialog with back-and-forth prompting mechanism. Each box represents a prompt that is passed to LLM agents in a round-robin fashion.

实验结果

研究问题

  • RQ1提示工程如何让自治LLM代理模拟可信的人类互动?
  • RQ2哪些类别(单对单、单对多、多对多)最能描述LLM驱动的仿真及其局限性?
  • RQ3限制大规模、接近人类现实的LLM仿真的主要瓶颈是什么(例如上下文窗口)?
  • RQ4人设表述是否会产生与明确目标不同的涌现策略?

主要发现

卖家目标买家目标结果最终提示标记
Sell for over $20Negotiate for lowest priceSold for $25522
Sell for around $20Negotiate for lowest priceSold for $17369
Sell for over $20Buy for under $20No deal472
  • 两方讨价还价仿真最终价格为25美元,超过卖方目标20美元。
  • 六方谋杀悬疑仿真展示了一对多的对话并使用记忆流在代理之间维持上下文。
  • 提示的人物设定可以产生涌现的谈判策略,如卖方最初提高价格。
  • 简单场景的最终提示标记数在约300到600标记之间,远低于gpt-3.5-turbo的4,096标记上限。
  • 三种情景变体展示改变代理目标如何改变结果(例如以高于20美元成交 → 25美元;以约20美元成交 → 17美元;在对立约束下没有交易)。
  • 研究讨论了约束并提出改进建议,如更长的上下文窗口和记忆流增强的检索以支持更大规模的仿真。
Figure 2: This illustration depicts interactions between agents, where each response from the captain is conditioned on the memory stream.
Figure 2: This illustration depicts interactions between agents, where each response from the captain is conditioned on the memory stream.

更好的研究,从现在开始

从论文设计到论文写作,大幅缩短您的研究时间。

无需绑定信用卡

本解读由 AI 生成,并经人工编辑审核。