智能体
x.com ·
@omarsar0
论文：100 个 LLM agents 运行城镇经济 26 周，货币流通趋于停滞
一项 arXiv 论文将 100 个 LLM agents 置于基于真实 Pokhara Lakeside 地理数据的城镇经济中运行 26 个模拟周，共 91 次运行、244 万次决策。结果显示货币流动停滞：12 倍旅游冲击使企业收入上升 4.62 倍，但工资仅变动 1.03 倍，3981 个菜单项中仅 0.3% 曾被重新定价；现金转移在 311 步后仍有 96.7% 未支出。作者还发现更换底层 LLM 会改变所有测量结果，而删除 agents 记忆没有可检测影响，且 1 至 2 周短周期结果具有误导性，财富排名到第 12 周才开始变化。
AI推荐理由 该帖介绍了一篇 arXiv 论文的实验设计：100 个 LLM agents 在真实地理数据上运行城镇经济 26 个模拟周，覆盖 91 次运行和 244 万次决策。给出的具体发现包括货币流通停滞、12 倍旅游冲击仅带来 4.62 倍企业收入而工资仅 1.03 倍、3981 个菜单项中仅 0.3% 被重新定价、96.7% 的现金转移在 311 步后仍未支出，以及更换底层 LLM 会改变全部测量结果而删除记忆无显著影响。这些量化结果对构建 agent 仿真有直接参考价值，但帖子为论文摘要转述，未提供同行评审或复现信息。
78
浏览 835
点赞 5
评论 2
转发 0
正文
What happens if you agents to run a town's economy?
This super interesting paper provides some insights:
They put 100 LLM agents in charge of a town economy for 26 simulated weeks, and they find that money stops moving.
The agents earn wages, run businesses and set prices on real Pokhara Lakeside geography, across 91 runs and 2.44M decisions.
A 12x tourist shock raises business revenue 4.62x. Wages move 1.03x, and only 0.3% of 3,981 menu items are ever repriced.
A cash transfer shows the same pattern. 96.7% of it is still unspent 311 steps later.
Here is the interesting part for anyone building agent simulations. Swapping the underlying LLM changed every outcome they measured. Deleting the agents' memory changed none of them detectably.
Results at the usual 1 to 2 week horizon also mislead. Wealth rankings look frozen at 2 weeks and only start to move by week 12.
Paper: [arxiv.org/abs/2609.11108](https://arxiv.org/abs/2609.11108)
Chat with Paper: [academy.dair.ai/papers/but-how…](https://academy.dair.ai/papers/but-how-would-ai-agents-run-a-towns-economy-2609.11108)
