📄 中文摘要
代理强化学习（ARL）作为一种有前景的训练代理以解决复杂多步骤交互任务的范式，近年来受到广泛关注。然而，ARL的训练过程往往不稳定，容易导致训练崩溃，这限制了其在更大环境和更长交互时间上的可扩展性，同时也约束了对算法设计选择的系统性探索。ARLArena的提出为解决这一问题提供了稳定的训练方案和系统分析框架，能够在可控和可重复的环境中检验训练的稳定性。ARLArena首先构建了一个干净且标准化的测试平台，然后将策略梯度分解为四个核心设计维度，以评估其性能。
ARLArena：稳定的代理强化学习统一框架
出处: ARLArena: A Unified Framework for Stable Agentic Reinforcement Learning
发布: 2026年2月26日
📄 English Summary
ARLArena: A Unified Framework for Stable Agentic Reinforcement Learning
Agentic reinforcement learning (ARL) has emerged as a promising paradigm for training agents to tackle complex, multi-step interactive tasks. However, the instability of ARL often leads to training collapse, which limits scalability to larger environments and longer interaction horizons, while constraining systematic exploration of algorithmic design choices. The proposed ARLArena offers a stable training recipe and a systematic analysis framework to examine training stability in a controlled and reproducible setting. ARLArena constructs a clean and standardized testbed and decomposes policy gradient into four core design dimensions to assess performance across various configurations.