跳到主要内容Skip to main content

← 返回AI 论文← Back to AI Papers

Agent 与工具Agents & tools

Generative Agents: Interactive Simulacra of Human BehaviorGenerative Agents: Interactive Simulacra of Human Behavior

问题:如何用 LLM 构建可信的人类行为代理,让多个 agent 在沙盒里自主社交与计划。
证据位置:摘要给出记忆流、反思、动态检索与计划架构;沙盒 25 个 agent;情人节派对案例显示邀请、约会与到场是自发涌现;消融显示观察、计划、反思各自贡献可信度。
作者自陈局限:沙盒小城与 25 个 agent;可信度由作者评估,不是真实人类行为验证;记忆与反思有成本;未覆盖大规模社会系统。
它没主张的:不是社会预测工具;不是真实人类复制;不是生产级多 agent 架构。
Problem: how to build believable human-behaviour proxies with LLMs, so multiple agents can socialise and plan in a sandbox.
Where the evidence sits: the abstract describes a memory stream, reflection and dynamic retrieval/planning architecture; a sandbox of 25 agents; the Valentine's party case shows invitations, dates and attendance emerging autonomously; ablation shows observation, planning and reflection each contribute to believability.
What the authors say about limits: a small-town sandbox with 25 agents; believability is author-evaluated, not validated against real humans; memory and reflection cost compute; large-scale social systems are not covered.
What it does not claim: not a social forecasting tool; not a copy of real humans; not a production multi-agent architecture.

关键参数Key specs

证据位置 / evidence
摘要:25 agents、记忆/反思/计划、派对案例
沙盒规模 / sandbox
25 agents
复现状态 / reproduction
未复现(只读摘要)

官方原文怎么说What the source says

computational software agents that simulate believable human behavior
1 https://arxiv.org/abs/2304.03442
store a complete record of the agent's experiences using natural language, synthesize those memories over time into higher-level reflections, and retrieve them dynamically
2 https://arxiv.org/abs/2304.03442
a small town of twenty five agents
3 https://arxiv.org/abs/2304.03442

资料与许可Source and licence

资料时间Source checked
2026-11-19
许可 / 类型Licence / kind
arxiv-preprint
许可说明Licence evidence
https://arxiv.org/abs/2304.03442
类目Category
Agent 与工具Agents & tools

论文里的数字怎么读How to read the numbers

先分清测什么、怎么算、和谁比、有没有数据污染。拿不准就回原文看实验设置。Before citing a score, separate what it measures, how it is scored, what it compares against and whether contamination is controlled. If unsure, go back to the experiment setup in the paper. 看评测基准See benchmark cards · 看数据集卡See dataset cards

去官网看原页Open the official page →

← 返回AI 论文← Back to AI Papers