Operational Constraints Conflict Resolution中对齐动态分辨率的要求
人工智能
2025-11-19 v2
摘要
部署的自主AI系统必须在新颖或不明确的情境中评估多个可行的行动方案（扩展的行为序列）。尽管经过大量训练，这些系统不可避免地会遇到没有可行行动方案完全满足所有运营约束（如操作程序、规则、法律、规范和目标）的情境。为实现符合人类期望和价值观的目标，代理人必须超越其训练的策略，而是构建、评估和论证候选行动方案。这一过程需要可能超出先前（策略）训练的“知识”。本文刻画了代理人在这些情境下决策的要求。还识别了代理人需要的类型知识，以便在代理人目标和符合人类期望之间做出决策。drawing on both analysis and empirical case studies, we examine how agents need to integrate normative, pragmatic, and situational understanding to select and then to pursue more aligned courses of action in complex, real-world environments.
引用
@article{arxiv.2511.10952,
title = {Requirements for Aligned, Dynamic Resolution of Conflicts in Operational Constraints},
author = {Steven J. Jones and Robert E. Wray and John E. Laird},
journal= {arXiv preprint arXiv:2511.10952},
year = {2025}
}
备注
8 pages + references and technical appendix (accepted for oral presentation at AAAI26)