Research map / Large language models / Alignment and preferences
Do LLMs have Consistent Values?
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Probes LLMs with the PVQ-RR human values questionnaire under five prompting strategies (notably a novel "Value Anchor" prompt) and checks whether value rankings and inter-value correlation structure match the Schwartz circular model found in human data.
Problem
Whether LLM responses to value questionnaires reproduce human value structure, both first-order (value rankings) and second-order (correlations between values within a single session), and whether LLMs can act as coherent personas.
Approach
Introduces "Value Anchoring," a prompt that asks the model to answer as a person emphasizing a given value (from the Best-Worst Refined Values scale), which shifts scores of neighboring values on the value circle and yields inter-value correlations matching human Schwartz circumplex structure; compares it against basic, demographic, persona and name prompts using MDS and Procrustes analysis.
Key results
- Value Anchor prompt yields the best match to human inter-value correlation structure, with sum of squared MDS differences of about 0.10-0.18 across models at temperature 0, versus up to 0.92 for basic prompt on GPT-4.
- Most prompt-model combinations reach Spearman rank correlations above 0.8 with the human value hierarchy; the basic prompt with GPT-4 is an exception with very low correlation.
- Value anchoring makes scores of values near the anchored value on the circle higher, following a sine-like curve, which explains the emergence of human-like correlation structure.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use a value anchor instead of the basic questionnaire prompt | GPT-4 · Human value correlations | sum of squared MDS differences: 0.92 -> 0.23 | Helped |
| Use a value anchor instead of the basic questionnaire prompt | Gemini Pro · Human value correlations | sum of squared MDS differences: 0.87 -> 0.11 | Helped |
| Use a value anchor instead of a name prompt | Llama 3.1 8B · Human value correlations | sum of squared MDS differences: 0.60 -> 0.18 | Helped |
| Use a value anchor instead of a name prompt | Llama 3.1 70B · Human value correlations | sum of squared MDS differences: 0.45 -> 0.10 | Helped |
| Use a value anchor instead of a name prompt | Gemma 2 9B · Human value correlations | sum of squared MDS differences: 0.23 -> 0.10 | Helped |
| Use a value anchor instead of a name prompt | Gemma 2 27B · Human value correlations | sum of squared MDS differences: 0.23 -> 0.16 | Helped |
| Increase temperature from 0.0 to 0.7 | Value Anchor prompting across models · Human value correlations | sum of squared MDS differences | No clear effect |
| Use a value anchor rather than a name prompt | LLMs across models · Human value correlations | alignment of value correlation structure | Helped |
| Use a basic prompt rather than a value-anchor prompt | GPT-4 · Human value ranking | Spearman rank correlation | Hurt |
| Use batch rather than sequential prompting (ablation) | Llama models · Human value correlations | sum of squared MDS differences | Mixed |
| Use batch rather than sequential prompting (ablation) | Llama models · Human value ranking | Spearman rank correlation | No clear effect |
Surprising findings
- Standard basic prompting often yields negligible variance or internally inconsistent answers, suggesting LLMs do not behave as individuals with coherent value priorities by default.
- Results depend strongly on prompt choice, with the basic prompt performing poorly on correlation structure even when rankings look human-like.
Limitations
- Limited to five prompts, two temperatures and six models, and a single questionnaire (PVQ-RR).
- Code and data available only upon request; the paper states they will be released later.
- Cronbach's alpha analyses are largely in commented-out sections and not reported in the final results.
Experimental setup
Datasets: Portrait Value Questionnaire-Revised (PVQ-RR) Best-Worst Refined Values (BWVr) scale Human value data from 49 cultural groups (53,472 participants) Names from prior study (500 names, five ethnic groups) World Values Survey occupations Activity Card Sort hobbies
Benchmarks: PVQ-RR
Metrics: Spearman rank correlation sum of squared differences of Procrustes-aligned MDS embeddi Cronbach's alpha
Baselines: Basic prompt Demographic prompt Generated Persona prompt Names prompt
Builds on: Schwartz Theory of Basic Human Values Portrait Value Questionnaire-Revised (PVQ-RR) Best-Worst Refined Values scale
Closest papers by meaning
- CLAVE: An Adaptive Framework for Evaluating Values of LLM Generated Responses (NeurIPS 2024 · similarity 0.91)
- Measuring Human and AI Values Based on Generative Psychometrics with Large Language Models (AAAI 2025 · similarity 0.89)
- Can DPO Learn Diverse Human Values? A Theoretical Scaling Law (NeurIPS 2025 · similarity 0.87)
- Aligning to Thousands of Preferences via System Message Generalization (NeurIPS 2024 · similarity 0.87)
- ALI-Agent: Assessing LLMs' Alignment with Human Values via Agent-based Evaluation (NeurIPS 2024 · similarity 0.87)
- Bridging Human and LLM Judgments: Understanding and Narrowing the Gap (NeurIPS 2025 · similarity 0.87)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.