Agent trajectories often contain both successful decisions and failed attempts, yet optimization with sparse trajectory-level rewards struggles to identify w…
机构:中科院
来源:arXiv 2611.06543 | AI4Papers 论文推荐平台