LLM agents combine models with execution harnesses, so agentic benchmark performance reflects their joint behavior. We argue that harnesses should therefore …
机构:上智院
来源:arXiv 2610.32483 | AI4Papers 论文推荐平台