Joint reinforcement learning and self-distillation combines task rewards with dense teacher guidance for agent training, but their gradients can partly cance…
机构:腾讯
来源:arXiv 2610.29965 | AI4Papers 论文推荐平台