'배틀쉽' 게임을 활용해 AI 에이전트에게 더 나은 질문법 가르치기

This research addresses a critical limitation in AI agents—their struggle to ask strategic questions in uncertain environments essential for scientific discovery and medical diagnosis. The breakthrough is that smaller, optimized models like Llama 4 Scout can outperform GPT-5 at just 1% of the cost, validating a more practical and efficient approach for deploying intelligent agents at scale.