LLM Architecture Detects Human Values in Text
A new paper on arXiv (2605.27373) introduces an LLM-based architecture for identifying and quantifying human values in text. The system overcomes prior limitations tied to specific value theories or complex prompt engineering. It comprises three modules: one generates structured value specifications from any theoretical framework's foundational texts; another labels texts using these specifications. The approach aims to align autonomous systems' decisions with ethical and moral considerations, moving beyond traditional utility-maximisation models.
Key facts
- arXiv paper 2605.27373
- LLM-based architecture identifies human values in text
- Three-module system
- Overcomes limitations of previous approaches
- Aims to align autonomous systems with ethical considerations
- Detects explicit and implicit values
- Quantifies intensity of values
- Not tied to specific value theory
Entities
Institutions
- arXiv