Tag
GAGPO proposes a critic-free RL method that uses a non-parametric grouped value proxy for step-level credit assignment in multi-turn agentic tasks, outperforming strong baselines on ALFWorld and WebShop.