Generative Psycho-Lexical Approach for Constructing Value Systems in Large Language Models

TL;DR

The paper addresses the gap that Large Language Models (LLMs) lack psychologically grounded intrinsic values. It introduces the Generative Psycho-Lexical Approach (GPLA), an automated, agent-centric pipeline that uses three LLMs (Perception Parser, Value Generator, Value Evaluator) and five steps to extract perceptions, map them to values, filter redundancies, measure value orientations non-reactively, and construct a five-factor LLM value system. The authors validate GPLA through three benchmarking tasks—Confirmatory Factor Analysis for structure validity, LLM Safety Prediction for predictive validity, and LLM Value Alignment for representation power—demonstrating that the proposed system yields superior structure fit, safety prediction accuracy, and alignment performance compared to Schwartz’s human-centered values. The work emphasizes GPLA’s advantages over prior psycho-lexical approaches in terms of full automation, broader lexicon coverage, and non-reactive measurement, while noting current language limitations and suggesting future multilingual extensions and more flexible alignment targets to enhance practical deployment. Overall, GPLA offers a scalable, adaptable framework for constructing and validating psychologically informed LLM value systems with meaningful implications for safety and alignment in real-world AI systems.

Abstract

Values are core drivers of individual and collective perception, cognition, and behavior. Value systems, such as Schwartz's Theory of Basic Human Values, delineate the hierarchy and interplay among these values, enabling cross-disciplinary investigations into decision-making and societal dynamics. Recently, the rise of Large Language Models (LLMs) has raised concerns regarding their elusive intrinsic values. Despite growing efforts in evaluating, understanding, and aligning LLM values, a psychologically grounded LLM value system remains underexplored. This study addresses the gap by introducing the Generative Psycho-Lexical Approach (GPLA), a scalable, adaptable, and theoretically informed method for constructing value systems. Leveraging GPLA, we propose a psychologically grounded five-factor value system tailored for LLMs. For systematic validation, we present three benchmarking tasks that integrate psychological principles with cutting-edge AI priorities. Our results reveal that the proposed value system meets standard psychological criteria, better captures LLM values, improves LLM safety prediction, and enhances LLM alignment, when compared to the canonical Schwartz's values.

Figures (4)

• Figure 1: Generative Psycho-lexical Approach (GPLA) for Constructing LLM Value System.
• Figure 2: Dendrogram of our value system. Values with a "*" are negatively loaded.
• Figure 4: Scree plot of the eigenvalues of our factors.
• Figure 5: Correlation heatmap derived from lexicon co-occurrence frequency biedma2024beyond, after Min-Max normalization.