Reinforcement learning (RL) for search agents often suffers from evidence fabrication, a failure mode where agents bypass external information ($E_{ext}$) an…
机构：微软
来源：arXiv 2610.13157 | AI4Papers 论文推荐平台