From Self-Distillation to Self-Practice: Privileged Information for Multi-Turn Agents
Privileged Self-Practice (PSP), which keeps the PI and moves it from the loss to the sampler and is the only method that consistently outperforms plain GRPO, improving task-goal completion on AppWorld and the resolved rate by up to 61% on SWE-bench Verified.
Xing-Yu Su, Abhishek Kumar, Qing Ping et al.
· 0 citations