SAPO：面向智能体强化学习的单轨自回归策略优化

View PDF HTML (experimental)

Abstract:Agentic reinforcement learning (RL) has emerged as an important post-training approach for enhancing the capabilities of Large Language Models (LLMs). However, existing methods face a trade-off between policy performance and resource efficiency. Conventional Proximal Policy Optimization (PPO) implementations incur substantial memory overhead from a separate critic, whereas critic-free group-relative methods require multiple rollouts and face potential learning bottlenecks on long-horizon tasks. In this work, we propose Single-rollout Autoregressive Policy Optimization (SAPO), an efficient PPO-style framework that unifies policy optimization and value learning within a single causal language model. SAPO exploits the autoregressive structure of LLMs to sequentially generate action and value estimation at distinct causal boundaries with shared parameters, and then jointly optimizes the PPO objectives and an auxiliary on-policy SARSA objective with turn-level generalized advantage estimation, where the latter is designed to facilitate value learning. Extensive experiments on ALFWorld and WebShop with Qwen2.5-1.5B/7B and Qwen3-14B demonstrate that SAPO reduces peak GPU memory usage by 24.8% and per-iteration runtime by 23.1% over strong PPO baseline, while matching or slightly improving task success rate. Our experiments also show that SAPO outperforms Group Relative Policy Optimization (GRPO) and recent cutting-edge variants in both task success and training stability.

Submission history

From: Dayang Liang [view email]
[v1] Thu, 20 Aug 2026 09:43:47 UTC (315 KB)
[v2] Sun, 30 Aug 2026 04:13:02 UTC (315 KB)
[v3] Wed, 30 Sep 2026 15:16:29 UTC (944 KB)
[v4] Mon, 26 Oct 2026 13:13:30 UTC (1,079 KB)