LLM의 도구 학습 성능을 향상하기 위해 단계별 보상과 최적화를 활용하는 강화 학습 프레임워크인 StepTool을 소개합니다.
Existing tool documentation, primarily designed for humans, often hinders LLMs from effectively utilizing external tools. DRAFT, a novel framework, addresses this challenge by dynamically refining tool documentation based on feedback from LLM-tool interactions, thereby improving LLMs' ability to understand and use tools.