How top AI labs are building RL agents in 2026 (using Karpathy's system prompt learning idea)

How top AI labs are building RL agents in 2026 (using Karpathy's system prompt learning idea)

摘要

顶级AI实验室正趋同于一种方法：将系统提示作为奖励函数，训练强化学习智能体。这一思路源自Karpathy的系统提示学习理念，Anthropic、OpenAI和DeepSeek都在推动从RLHF到RULER的进化。