Towards Unified Alignment Between Agents, Humans, and Environment

• DOI: 10.48550/arxiv.2402.07744
• Publication Date: 2024-02-12
• Authors: Zonghan Yang; An Liu; Zijun Liu; Kaiming Liu; Fangzhou Xiong; Yile Wang; Zeyuan Yang; Qingyuan Hu; Xinrui Chen; Zhenhe Zhang; Fuwen Luo; Zhicheng Guo; Peng Li; Yang Liu

The rapid progress of foundation models has led to the prosperity autonomous agents, which leverage universal capabilities conduct reasoning, decision-making, and environmental interaction. However, efficacy agents remains limited when operating in intricate, realistic environments. In this work, we introduce principles $\mathbf{U}$nified $\mathbf{A}$lignment for $\mathbf{A}$gents ($\mathbf{UA}^2$), advocate simultaneous alignment with human intentions, dynamics, self-constraints such as limitation monetary budgets. From perspective $\mathbf{UA}^2$, review current agent research highlight neglected factors existing benchmarks method candidates. We also proof-of-concept studies by introducing features WebShop, including user profiles demonstrate personalized reranking complex runtime cost statistics reflect self-constraints. then follow $\mathbf{UA}^2$ propose an initial design our agent, benchmark its performance several candidate baselines retrofitted WebShop. extensive experimental results further prove importance $\mathbf{UA}^2$. Our sheds light on next steps improved general problem-solving abilities.