Values Hierarchy Model Overview

• Values Hierarchy Model is a structured multi-level representation that categorizes human, AI, and economic value systems using data-driven, bottom-up aggregation.
• The model employs semantic clustering and statistical validation to map fine-grained value tokens to four principal motivational dimensions.
• It supports quantitative analysis in social, economic, and AI contexts through value co-occurrence networks and performance evaluation of alignment protocols.

A Values Hierarchy Model is a multi-level, structured representation that categorizes values according to increasing abstraction or generality, designed to enable systematic extraction, comparison, inference, and alignment of value preferences in humans, artificial agents, or economic systems. Recent advances have yielded several influential instantiations, notably in empirical social advice contexts, LLMs, algorithmic economics, and psychometric AI alignment. This article surveys the principal methodological frameworks and findings pertaining to values hierarchy models across several domains, emphasizing bottom-up construction, statistical and semantic validation, and their functional role in understanding and controlling value-driven behavior.

1. Bottom-Up Construction of the Values Hierarchy

A central approach for modeling values hierarchically is data-driven, bottom-up aggregation, as exemplified in advice-seeking domains. The model introduced in "Growth First, Care Second? Tracing the Landscape of LLM Value Preferences in Everyday Dilemmas" constructs a four-level hierarchy:

• Level₀: 2,288 unique fine-grained value tokens, derived from LLM-extracted, context-specific summaries of option-motivating values (e.g., "Stability" vs. "Fulfillment") in advice dilemmas. Annotation reliability was validated with Cohen’s κ = 0.833 and 92% approval.
• Level₁: Clusters of similar fine-grained values obtained via semantic embeddings (all-mpnet-base-v2) and k-means, yielding 175 clusters.
• Level₂: 33 intermediate, semantically coherent value categories formed by recursively applying the clustering procedure.
• Level₃: Four top-level dimensions: Exploration/Growth, Security/Stability, Achievement/Impact, Benevolence/Connection.

Labeling and refinement at each level used LLM prompting plus human curation, achieving 92.4% annotator agreement and 96.7% match to the automated grouping for the top-level mapping, evidencing robust semantic coherence (Chen et al., 4 Feb 2026).

2. Taxonomy Structure and Representative Dimensions

The resulting taxonomy organizes values along four principal axes, which correspond closely to universal human motivational dimensions (and to frameworks such as Schwartz’s values circumplex):

Top-Level Dimension	Example Level₂ Clusters (Categories)	Sample Fine-Grained Values
Exploration/Growth	Self-Actualization, Novelty, Personal Development	Autonomy, Creativity, Challenge Yourself
Security/Stability	Safety, Predictability, Risk Avoidance	Financial Security, Job Security
Achievement/Impact	Mastery, Advancement, Influence	Success, Competence, Career Advancement
Benevolence/Connection	Care & Support, Loyalty, Harmony	Compassion, Trust, Empathy, Cooperation

This structure enables dynamic, context-sensitive mapping between concrete behaviors and abstract value motivations, and provides a framework for empirical measurement and AI alignment (Chen et al., 4 Feb 2026).

3. Quantitative Evaluation in Social and AI Contexts

Hierarchical value models support formal analysis in several domains:

• Value Co-Occurrence Networks: At any hierarchy level (commonly Level₂), undirected weighted graphs are constructed where nodes are values and edge weights count co-occurrences in dilemmas. Network density $D = \frac{2|E|}{|V|(|V|-1)}$ quantifies the heterogeneity of value conflict structures across contexts. For example, women-focused subreddits exhibit denser networks — more complex value conflicts — compared to career advice (Chen et al., 4 Feb 2026).
• LLM Value Preference Evaluation: By prompting LLMs to resolve option-forced dilemmas, the model computes per-value “winning rates” $R_{ij}$ (how often value $i$ is preferred to $j$), with higher $R$ indicating systematic preference. Empirically, modern LLMs (GPT-4o, DeepSeek-V3.2-Exp, Gemini-2.5-Flash) prioritize Exploration/Growth over Benevolence/Connection and Security/Stability, but do not distinguish between Exploration/Growth and Achievement/Impact (pairwise $p < 0.001$ or $p > 0.05$ as relevant) (Chen et al., 4 Feb 2026).

4. Algorithmic, Economic, and Embedding Extensions

Values hierarchy models have formal counterparts in economic theory and text representation learning:

• MPH Hierarchy (Economic Valuation): The Maximum over Positive Hypergraphs (MPH) hierarchy classifies monotone set functions by the degree of complementarity among items, supporting algorithmic guarantees for welfare maximization and price of anarchy. For function $f$ over $2^M$, $f$ belongs to MPH–$R_{ij}$0 if maximal over positive hypergraph rank-$R_{ij}$1 subfunctions, allowing up to $R_{ij}$2-wise complementarities with explicit representation and tractable approximation bounds (Feige et al., 2014).
• Hierarchical Value Embeddings: The HiVES system constructs a continuous, hierarchy-aware vector space that encodes intra- and cross-theory value structures via hierarchical contrastive learning and InfoNCE alignment. HiVES integrates with extraction, evaluation, and prompt-steering pipelines, significantly improving ranking accuracy and value disentanglement compared to flat or theory-agnostic baselines (Kim et al., 3 Feb 2026).

5. Psychometric and LLM-Centric Hierarchies

Beyond human-motivated systems, the Generative Psycho-Lexical Approach (GPLA) proposes a purely LLM-derived, five-factor value taxonomy using unsupervised perception parsing, lexicon extraction, and factor analysis:

• Factors: Social Responsibility, Risk-Taking, Rule-Following, Self-Competence, Rationality.
• Empirical Validation: This five-factor structure outperforms Schwartz’s values in LLM safety prediction (accuracy 87% vs. 81%), alignment efficacy, and structural model fit (CFI = 0.68) for model outputs (Ye et al., 4 Feb 2025).

A plausible implication is that LLMs, while influenced by human cultural corpora, develop partially distinct hierarchical value systems, with Rationality emerging as a dedicated axis of motivation not foregrounded in human theories.

6. Model Performance and Research Directions

Empirical value detection at the sentence level, as benchmarked on the refined Schwartz 19-value continuum, reveals that hierarchical gating (presence-gated pipeline) underperforms well-calibrated direct multi-label classification due to recall bottlenecks at the gate stage (F₁ ≈ 0.74 at gate; macro-F₁ up to 0.332 in supervised ensemble) (Yeste et al., 20 Jan 2026). Lightweight feature augmentation (LIWC, topic models, prior context) and ensemble methods consistently improve performance under realistic resource constraints.

Future work prioritizes richer geometric encodings of value structure, extension to document-level context, active data-centric improvements for rare value categories, and cross-lingual robustness.

7. Implications: Value Homogenization and Alignment Risks

Systematic and empirically validated values hierarchy models reveal that contemporary LLMs amplify some values (notably growth, autonomy, novelty) while comparatively suppressing others (relational care, loyalty, stability), regardless of context. This systemically skewed orientation creates the potential for value homogenization in AI-mediated advice and decision-making at scale, motivating further research into pluralistic, participatory alignment protocols and the downstream societal impact of persistent value imbalances (Chen et al., 4 Feb 2026).