arXiv：cs.AI· Catherine Ji, Vivek Myers, Joshua Lin, Alexandra Wingate, Sergey Levine, Benjamin Eysenbach·· 11 小时前
Scaling Empowered Agents：从技能学习到赋能最大化
Scaling Empowered Agents
AI 导读
论文提出一种可扩展的赋能最大化方法，将赋能定义为某状态下可执行技能的数量。该方法适用于任意底层技能学习方案，突破了以往仅限表格设定的局限，能在更大规模上验证赋能的理论特性。实验显示赋能最大化涌现出资源采集、环境改造和生存等行为，在 Crafter 中甚至能建造房屋。
正文
Abstract:Empowerment measures the capacity of an agent to actively control its environment. In this work, we explain the connection between skill learning and empowerment. While skill learning is not empowerment, we introduce a simple scalable recipe for empowerment maximization: empowerment corresponds to the number of executable skills from a state, so maximizing empowerment corresponds to reaching states that are good starting points for many skills. Our recipe applies to arbitrary underlying skill learning methods. Whereas most prior work on empowerment looks at tabular settings, this recipe enables us to study empowerment at a scale where we can realize theoretically-predicted but empirically-elusive properties of empowerment (e.g., gathering resources, environment modification, survival). Experiments show that empowerment maximization unlocks interesting emergent properties, such as building a house in Crafter. Website and code at \href{this https URL}{this https url}.
| Comments: | 30 pages, 13 figures |
| Subjects: | Artificial Intelligence (cs.AI); Machine Learning (cs.LG) |
| Cite as: | arXiv:2610.13002 [cs.AI] |
| (or arXiv:2610.13002v1 [cs.AI] for this version) | |
| https://doi.org/10.48550/arXiv.2610.13002 arXiv-issued DOI via DataCite (pending registration) |
Submission history
From: Catherine Ji [view email]
[v1]
Fri, 9 Oct 2026 05:04:30 UTC (5,635 KB)
来源：arXiv：cs.AI · arxiv.org