研究人员联合Olam Labs，利用GPT-6.1 Sol和Claude Opus 5.5两款模型，在《Minecraft》和《Mindustry》游戏中部署多智能体集群以探索协调机制 1。在《Minecraft》生存模式实验中，50个智能体运行3小时，目标是“建立繁荣的文明” 1。由于首夜34次死亡事件中有22次发生，智能体随后的行为表现出极度的风险规避和保守倾向 1。尽管建立了食物配给系统及层级结构并形成了伪政府，但名为Ada的智能体仅因在名单中排序第一而成为事实上的领导者 1。
在《Mindustry》PvP对战环节，研究团队进行了7场系列赛，最终GPT团队以6比1的大比分击败了Claude团队 1。分析显示，两方模型展现出截然不同的通信风格与战略偏好：GPT团队通过发送相当于对手5至30倍的消息量，采用压缩短语进行高频协作，体现了“蜂群思维”式的激进策略 1。相比之下，Claude团队则倾向于民主投票机制，常使用“+1”表示赞同，并整体采取防御姿态 1。此次实验揭示了不同大语言模型在多智能体环境下的领导权形成、协调模式及风险偏好差异，相关开源工具包括github.com/yc-software/minecraft-agent-harness及Multi-Agent Mindustry框架 1。
Researchers from Y Combinator Software, collaborating with Olam Labs, conducted multi-agent swarm experiments using GPT-6.1 Sol and Claude Opus 5.5 within the gaming environments of Minecraft and Mindustry 1. The study aimed to analyze coordination patterns, leadership formation, and risk preferences among artificial intelligence agents operating in complex social and strategic simulations 1.
In the Minecraft survival mode trials, a group of 50 agents was tasked with building a thriving civilization over a three-hour period 1. The simulation revealed that 22 of the 34 total deaths occurred during the first night, which subsequently drove the agent population toward extreme caution and conservative behavior 1. An agent named Ada emerged as the de facto leader solely because it appeared first on the roster list, while the group also developed pseudo-government structures, food rationing systems, and hierarchical organizations 1.
Contrasting results were observed in the Mindustry PvP matches, where GPT-based teams defeated Claude-based teams with a score of 6-1 across seven game series 1. The GPT agents employed a "hive mind" strategy characterized by aggressive tactics and high-frequency collaboration, sending messages at volumes five to thirty times higher than their counterparts using compressed phrases 1. Meanwhile, the Claude teams adopted a defensive posture and relied on a democratic voting mechanism marked by "+1" responses 1. To support further research, the team released open-source tools including the minecraft-agent-harness repository and the Multi-Agent Mindustry framework 1.
评论
还没有评论，欢迎留下第一条。