arXiv cs.AI· Elinor Poole-Dayan, Jillian Fisher, Jacob Andreas, Mitchell Gordon, Michiel A. Bakker·· 4 小时前AI 评分26
让多元价值对齐研究真正落地：通向高影响力研究的路线图
A Roadmap to Impactful Pluralistic Alignment Research
AI 导读
研究者审计主要前沿实验室和开源权重开发者后，未发现公开证据表明多元价值对齐已影响其模型的训练或评估。论文提出三大障碍：现有论证多为规范性或理论性、社区尚未明确多元行为何时适用及如何落地、现有评测与方法未达部署标准，并针对每项给出研究路线图。该路线图也适用于具有类似变革理论的其他负责任 AI 研究领域。
正文 · 原文
Abstract:Much research on responsible AI aims to change the behavior of AI systems that billions of people use. Pluralistic value alignment---the goal of building AI systems that represent and serve diverse human values and perspectives---has emerged as one such active research agenda. Yet, in an audit of the major frontier labs and open-weight developers, we find no public evidence that it has shaped how their models are trained or evaluated. Pluralistic alignment requires adoption to achieve its main goals, so we identify three barriers and lay out a roadmap of research to address each: 1. The justifications for pluralistic alignment so far are mostly normative or theoretical. Studies are needed that measure the benefits and drawbacks of pluralistic AI for users and society. 2. The pluralistic alignment research community has not settled when pluralistic behavior is warranted or what it should look like in practice. Research is needed to specify this behavior concretely enough for developers to operationalize. 3. Existing pluralism evaluations and methods are not ready for deployment. Evaluations and methods are needed that report their effects on other model properties and meet the requirements of production systems. These barriers are not unique to pluralistic alignment, so our roadmap applies to other areas of responsible AI research with similar theories of change. As people hand more decisions to AI systems and AI agents act on their behalf, we hope our roadmap helps this research reach the deployed models making those decisions.
| Subjects: | Artificial Intelligence (cs.AI) |
| Cite as: | arXiv:2607.22305 [cs.AI] |
| (or arXiv:2607.22305v2 [cs.AI] for this version) | |
| https://doi.org/10.48550/arXiv.2607.22305 arXiv-issued DOI via DataCite |
Submission history
From: Elinor Poole-Dayan [view email]
[v1]
Fri, 24 Jul 2026 13:46:03 UTC (254 KB)
[v2]
Tue, 3 Nov 2026 18:03:31 UTC (1,492 KB)
来源：arXiv cs.AI · arxiv.org