SEED: Self-Evolving On-Policy Distillation for Agentic Reinforcement Learning https://arxiv.org/abs/2607.14777 https://www.alphaxiv.org/ru/overview/2607.14777