ZJU-REAL/SDAR

Self-retiring on-policy distillation for agentic RL RetireOPD lets the student drop the teacher once their discrepancy stops shrinking, then trains with RL alone. Improves ALFWorld success by 14-19%…

🚨ZJU × Meituan × Tsinghua propose #SDAR, a new framework for multi-turn #LLM agents. It combines RL and OPSD via token-level adaptive gating to improve training stability. It achieves substantial…

Paper: Code:

智能体在训练时自己挑出做得好的轨迹当示范教材，自己教自己，不用人标注就能把复杂任务的成功率拉上去。 SDAR 是浙大 REAL 实验室提出的智能体强化学习方法。训练过程中自动筛选高分轨迹作为自产示范，再对策略做自我蒸馏，在 ALFWorld、WebShop 和 Search-QA 三个基准上全部超过标准 RL 基线。

Self-Distilled Agentic Reinforcement Learning (SDAR) SDAR stabilizes multi-turn LLM agent training by gating self-distillation signals within GRPO, yielding +9.4% gains on ALFWorld and significant…