Multi-turn agents are commonly evaluated with LLM-based simulated users, which can be overly cooperative and can therefore overestimate an agent's performanc…
机构:帝国理工
来源:arXiv 2610.35426 | AI4Papers 论文推荐平台