Researchers have proposed SAPO (Single-Rollout Autoregressive Policy Optimization), a new framework for agentic reinforcement learning that addresses three limitations of existing methods. SAPO uses a single autoregressive backbone to predict policy and value functions at distinct causal boundaries, allowing it to share parameters and reduce memory overhead. The method also includes a trajectory-level generalized advantage estimator to robustly estimate the contribution of each turn. Experiments show that SAPO outperforms PPO and GRPO on two tasks while reducing memory cost and runtime. --- Why it matters: This matters because existing agentic reinforcement learning methods have limitations, including lack of explicit value generalization and potential advantage collapse in long-horizon complex tasks. SAPO addresses these issues, making it a more efficient and effective framework for training large language models. Source: https://arxiv.org/abs/2608.19842