Summary
A new arXiv paper (2605.27373) by Eduardo de la Cruz Fernández, Marcelo Karanik, and Sascha Ossowski introduces an LLM-based architecture for detecting and quantifying the intensity of human values in text, whether explicit or implicit. Unlike prior approaches tied to specific value theories or dependent on complex prompt engineering, the proposed architecture is modular and tailorable: one module generates structured value specifications from the foundational texts of any theoretical framework, a second module annotates texts using these specifications, and a third assigns support/opposition ratings based on rhetorical and semantic evidence. This design separates the conceptualization of values from the detection task, yielding a scalable, reproducible pipeline that can adapt to different value theories. Experiments on the ValueEval dataset show the pipeline achieves good detection performance. The work supports building ethically-aware decision-making mechanisms for increasingly autonomous intelligent systems, moving beyond traditional utility-maximization models.
Paper Overview
Field: AI
Authors: Eduardo de la Cruz Fernández, Marcelo Karanik, Sascha Ossowski
Published: 2026-05-28
arXiv: 2605.27373
Summary
As intelligent systems become more autonomous, the research community is working on decision-making mechanisms that incorporate ethical and moral considerations, moving beyond traditional utility-maximization models. A key element of this effort is assessing how well decisions align with human values.
This paper introduces an LLM-based architecture that detects and quantifies the intensity of human values in text — whether explicit or implicit. The approach avoids two limitations of previous methods: being tied to a specific value theory, and relying on complex prompt engineering.
The architecture consists of three coordinated modules:
- Value specification module: generates structured value specifications from the foundational texts of any theoretical framework.
- Annotation module: uses these specifications to label texts with human values.
- Rating module: assigns support/opposition levels based on rhetorical and semantic evidence.
This modular design separates the conceptualization of values from the detection task, producing a scalable, reproducible pipeline that can be adapted to different theories. Experiments on the ValueEval dataset demonstrate that the pipeline achieves good detection performance.
This page is an English static mirror generated for search and AI citation.
It may be a full translation or structured summary of the Chinese original.
Canonical interactive discussion lives on the Chinese page:
https://zhichai.net/topic/177980501