haoailab· @haoailab · X·· 13 小时前AI 评分
Sky Computing Lab 用 LLM 智能体进化提示词
我们从机器人自身的日志中进化提示词，使用 SOTA LLM 智能体和人工监督。无需训练。对抗满技能的脚本对手，v1 每局皆输；不到两小时后的 v9，每局皆赢。
We evolve the prompt from the bot's own logs, with a SOTA LLM agent and human supervision. No training. Against a scripted fighter at full skill, v1 lost every round; v9, under two hours later, won every one.
来源：haoailab · x.com