EUVIMEDEUVIMED entdecken
Crossref

Moral Foundations Reward Modeling for Pluriversal LLM Alignment

Zhangfei Yang

Proceedings of the AAAI Symposium Series · 06.11.2026 · 10 (1) · 48-55

DOI: 10.1609/aaaiss.v10i1.43390 ↗

Kurz gefasst

Fachartikel in der Zeitschrift Proceedings of the AAAI Symposium Series, online erschienen am 06.11.2026. Verfasst von Zhangfei Yang.

AbstractBibliografische AngabenZitierenVerwandte Studien

Abstract (Originalsprache)

Prevailing approaches to value alignment in large language models (LLMs), including reinforcement learning from human feedback (RLHF) and direct preference optimization (DPO), rest on a Bradley--Terry assumption of a single scalar preference order over responses. When annotator populations disagree along cultural, political, or religious lines, this assumption does more than simplify the learning problem: it aggregates heterogeneous preferences into a single ordering, creating familiar social-choice trade-offs between collective consistency and the representation of minority viewpoints. We argue that this scalar aggregation structure can systematically privilege majority and Western, Educated, Industrialized, Rich, and Democratic (WEIRD) value distributions while obscuring minority viewpoints and epistemologies from the Global South and from China, whose AI governance model imposes a distinct, state-codified form of alignment rather than an implicit statistical one. Using the rise of open-weights Chinese models such as DeepSeek as a case study, we argue that neither the implicit-convergence paradigm of Western frontier labs nor the explicit-bifurcation paradigm practiced in China resolves the deeper problem of representing genuine, irreducible value pluralism. Rather than positing an abstract, axiomatic multi-valued logic to represent this pluralism, we ground our solution in an empirically validated account of moral psychology, Moral Foundations Theory, and its associated survey instrument, the Moral Foundations Questionnaire. We develop a Moral Foundations Reward Model (MFRM) that scores a response along a shared set of moral-foundation dimensions and then aggregates that score using a community-specific, empirically elicited distribution of foundation weights, so that a single response yields a distinct, uncertainty-quantified judgment for each cultural cluster rather than one averaged verdict. We specify a concrete two-stage elicitation pipeline, Polis-style opinion clustering followed by within-cluster Moral Foundations Questionnaire administration, a closed-form decision and routing rule based on the resulting cluster-level mean and variance, and a worked numerical illustration. We connect this proposal to existing empirical work on cross-cultural LLM opinion representation and moral reasoning, and to the pluralistic alignment taxonomy of Overton, steerable, and distributional pluralism. We close by discussing the framework's limitations, including the fact that Moral Foundations Theory was itself developed and predominantly validated in WEIRD populations, and the risk that a mechanism designed to represent bottom-up cultural pluralism could be misappropriated to launder top-down political censorship; we argue that these risks must be named rather than assumed away.

Zitieren

APA

Yang, Z. (2026). Moral Foundations Reward Modeling for Pluriversal LLM Alignment. Proceedings of the AAAI Symposium Series, 10(1), 48-55. https://doi.org/10.1609/aaaiss.v10i1.43390

Vancouver

Yang Z. Moral Foundations Reward Modeling for Pluriversal LLM Alignment. Proceedings of the AAAI Symposium Series. 2026;10(1):48-55. doi:10.1609/aaaiss.v10i1.43390

BibTeX
@article{Yang2026,
  author = {Yang, Zhangfei},
  title = {Moral Foundations Reward Modeling for Pluriversal LLM Alignment},
  journal = {Proceedings of the AAAI Symposium Series},
  year = {2026},
  volume = {10},
  number = {1},
  pages = {48--55},
  doi = {10.1609/aaaiss.v10i1.43390}
}
Kostenlos starten →