PRISM:作为AI对齐多视角框架的综合与调解视角推理
PRISM: Perspective Reasoning for Integrated Synthesis and Mediation as a Multi-Perspective Framework for AI Alignment
AI总结:
本文提出多视角框架PRISM,通过七种基础世界观和帕累托优化方案解决AI对齐中的价值观冲突与规范博弈,系统记录并调解伦理权衡。
摘要:
在这项工作中,我们提出了综合与调解视角推理(PRISM),这是一个多视角框架,旨在解决AI对齐中的持续挑战,如人类价值观冲突和规范博弈。PRISM以认知科学和道德心理学为基础,将道德关切组织为七个“基础世界观”,每个世界观被假设为捕捉人类道德认知的一个独特维度,从以生存为中心的反射到高阶整合视角。然后,它应用受帕累托启发的优化方案来调和竞争优先级,而不将其简化为单一指标。在可靠上下文验证以实现稳健使用的假设下,该框架遵循结构化工作流程,引出特定观点的响应,将其综合为平衡结果,并以透明和迭代的方式调解剩余冲突。通过参考认知科学、道德心理学和神经科学中的分层道德认知方法,PRISM阐明了不同道德驱动如何相互作用,并系统记录和调解伦理权衡。我们通过工作原型生成的真实输出展示了其有效性,将PRISM应用于公共卫生政策、工作场所自动化和教育等领域的经典对齐问题。通过将AI审议锚定在这些人类视角中,PRISM旨在限制可能漂移到非人类或以机器为中心领域的解释性跳跃。我们简要概述了未来方向,包括现实世界部署和形式验证,同时保持对多视角综合和冲突调解的核心关注。
英文摘要:
In this work, we propose Perspective Reasoning for Integrated Synthesis and Mediation (PRISM), a multiple-perspective framework for addressing persistent challenges in AI alignment such as conflicting human values and specification gaming. Grounded in cognitive science and moral psychology, PRISM organizes moral concerns into seven "basis worldviews", each hypothesized to capture a distinct dimension of human moral cognition, ranging from survival-focused reflexes through higher-order integrative perspectives. It then applies a Pareto-inspired optimization scheme to reconcile competing priorities without reducing them to a single metric. Under the assumption of reliable context validation for robust use, the framework follows a structured workflow that elicits viewpoint-specific responses, synthesizes them into a balanced outcome, and mediates remaining conflicts in a transparent and iterative manner. By referencing layered approaches to moral cognition from cognitive science, moral psychology, and neuroscience, PRISM clarifies how different moral drives interact and systematically documents and mediates ethical tradeoffs. We illustrate its efficacy through real outputs produced by a working prototype, applying PRISM to classic alignment problems in domains such as public health policy, workplace automation, and education. By anchoring AI deliberation in these human vantage points, PRISM aims to bound interpretive leaps that might otherwise drift into non-human or machine-centric territory. We briefly outline future directions, including real-world deployments and formal verifications, while maintaining the core focus on multi-perspective synthesis and conflict mediation.
机构、总结、中文摘要均由AI分析而来,可能存在错误,请以论文正文为准。