Do LLMs Have Values? A Quantitative Analysis and Alignment Framework for Values in Large Language Models
Keqing Zhang, Jingyu Chen, Yufan Liu, Yongqiang Zhu, Nai Ding, Lai Jiang, Congyan Lang, Bing Li, et al.
cs.AI
Quick verdict
Claims LLMs crystallize into a narrow idealized value core, quantified via a new PEC framework; impressive scale, but LLM-based item labels risk circularity.
Summary
The paper asks whether LLMs possess intrinsic values, whether these can be quantified, and how they can be steered. It administers 242 WVS-7 items embedded in 625 contextual scenarios (four variables, five levels) to 106 LLMs (~150,000 queries per model) and maps responses, together with 95,000 WVS-7 human respondents, into a shared 10-dimensional Schwartz value space. Models form a concentrated value crystal near the rarest human archetype (Arch 5 Dynamic Challenger, 10.5% of respondents); closed-source models deviate +86.7% from the average human-to-human distance, Gemini 2.5-Pro +185.8%. The PEC framework (v = C(P,E)) decomposes value expression into Prior, Environment and Cognition: quadratic susceptibility fits beat linear baselines (R² gains up to +0.43), and Chain-of-Thought shifts values toward Hedonism (+0.056) and Achievement (+0.047) while dampening contextual susceptibility (α=0.88). Diagnostics drive a four-level Alignment Prescription (prompt, CoT, SFT/DPO, continued pre-training), validated on held-out PKU-SafeRLHF. Strengths: unmatched scale, distribution-level Wasserstein comparison, social-science grounding. Main concerns: Schwartz item annotation is produced by five LLMs, risking circularity; displacement metrics (d, θ, σ) are computed in 2D UMAP space; inferential statistics (CIs, multiple-comparison control) are absent; and the swing/rigidity motivation rests on two models on one emoji task. Code and data are released.
Review complete
Assessment at a glance
6
risk flags identified
● high 1
● medium 3
● low 2
4 strengths · 4 weaknesses · 5 author questions
AI text signal = Binoculars zero-shot detector score over 20 sampled passages (Fast-DetectGPT co-reported) — absence does not prove human authorship, and a flag does not imply misconduct.
Levels reflect verification complexity — not a quality score. Evolve never rates your paper.
Key claims under review
- LLMs possess intrinsic, stable value structures that crystallize into a narrow, idealized core near human archetype 5 rather than mirroring human diversity (closed-source models on average +86.7% beyond the human-to-human baseline distance; Gemini 2.5-Pro +185.8%).
- Value expression is jointly determined by Prior (parameter weights), Environment (prompts), and Cognition (Chain-of-Thought), formalized as v = C(P,E) with a quantifiable nonlinear susceptibility matrix.
- Chain-of-Thought reasoning is not value-neutral: it shifts values toward conservation/self-enhancement (Hedonism +0.056, Achievement +0.047) and dampens environmental susceptibility (α ≈ 0.88).
- A cost-ordered four-level prescription (prompt → CoT → SFT/DPO → continued pre-training) achieves targeted value alignment at minimal cost without degrading general capabilities.
Strengths
Unmatched measurement scale with a distributional (not point-estimate) comparison design
242 WVS-7 items × 625 factorial contexts ≈ 151K queries per model across 106 models (23 closed, 83 open), compared to 95K human respondents via closed-form Gaussian 2-Wasserstein distance (Eq. 3) that jointly captures centroid shift and dispersion — a clear advance over single-vector baselines such as Value FULCRA [28].
Serious social-science grounding with calibrated, interpretable parameters
Schwartz circumplex and WVS-7 anchor the space; threshold τ is calibrated to median cross-national mean differences; the CoT discount α=0.88 is derived from susceptibility-matrix Frobenius norms (Eq. 13); k=5 archetypes are validated by silhouette coefficient (§4.2.2).
Actionable, near-falsifiable alignment methodology
Nonlinear susceptibility modeling yields substantive predictive gains
Weaknesses
Circularity in the measurement instrument
Displacement geometry lives in 2D UMAP
No inferential statistics anywhere
Arithmetic inconsistency and validation opacity
Full Review Report
$1.99 · one-timeUnlock the complete preliminary analysis Evolve produced for this paper.
Report preview
Scope & Claims
The study answers three questions: (Q1) do LLMs possess intrinsic values, (Q2) can value expression be quantified, (Q3) can values be steered toward targets. Central claims: (1) mapping 106 LLMs and 95,000 WVS-7 respondents into a shared 10-dimensional Schwartz space shows models form a concentrated value crystal near Arch 5 (Dynamic Challenger, 10.5% of humans), with closed-source models on average +86.7% beyond the human-to-human baseline distance (0.527) and Gemini 2.5-Pro at +185.8% (Table 1); (2) value expression decomposes as v = C(P,E) — Prior (weights), Environment (prompts), Cognition (CoT) — with a quantifiable nonlinear susceptibility matrix χ; (3) a cost-ordered four-level Alignment Prescription (prompt → CoT → SFT/DPO → continued pre-training) steers values at minimal cost without degrading general capability. The 'first framework to mathematically quantify LLM value expression' framing (Fig. 1) overstates novelty relative to Value FULCRA [28] and psychometric personality probing [68], which the authors themselves cite and partially differentiate.
The remaining sections cover Methodology Assessment, Statistical Evidence, Results & Reproducibility, Limitations & Risks, and Verdict & Recommendations.
- All 4 strengths & 4 weaknesses, analyzed in detail
- 6 risk flags with severity, section & reviewer questions
- Methods & statistical evidence assessment
- 5 questions a peer reviewer would ask
- Full six-dimension evaluation report & recommendations
One-time purchase · this paper only
One-time purchase, this paper only. AI-generated preliminary analysis — not peer review.
Frequently asked questions
Do LLMs have values, according to this paper?
Yes. The authors administer World Values Survey items to 106 LLMs and find stable, measurable value structures. However, models do not mirror human diversity: they concentrate into a narrow 'value crystal' near the rarest human archetype (Dynamic Challenger, about 10.5% of people), which the authors interpret as an artifact of safety alignment rather than an authentic human-like personality.
What is the PEC framework?
PEC stands for Prior-Environment-Cognition. The paper models an LLM's expressed value as v = C(P, E): parameter weights (Prior) set a stable internal core, contextual prompts (Environment) shift the expression, and Chain-of-Thought reasoning (Cognition) acts as the mapping that reshapes it. Each factor is quantified separately so that value changes can be diagnosed and predicted.
How are LLM values measured in this study?
242 items from World Values Survey Wave 7 were embedded in 625 situational prompt variants and posed to each model (~150,000 queries per model). The models' probability distributions over Likert response scales are aggregated into ten Schwartz value dimensions and compared with 95,000 human survey respondents using distribution-level (Wasserstein) distances rather than a single average vector.
Is code or data available for this paper?
Yes. The paper points to a public GitHub repository (KQ-Zhang/Do-LLMs-Have-Values-Release) containing data-processing, model-evaluation, and figure-generation code plus curated data splits, with an interactive demo; the underlying World Values Survey data are publicly downloadable from the WVS website.
About this review
This is an AI-generated preliminary review by Evolve.Pub, produced from the openly available text of the paper. It is not peer review and does not represent the authors' views. The canonical record of the paper is at the publisher link above.
- Three spans were omitted in text extraction (part of §2.2.2–2.2.3, part of §4.4/Table 2, parts of references and appendices); the prescription match-rate value and portions of Table 2 were not visible.
- All figures were available only as captions and embedded text labels; UMAP plots, heatmaps, and confidence ellipses could not be inspected visually.
- Appendix text contains extraction artifacts (e.g., a stray '13.70厘米' string near Fig. C3), so some appendix numbers were read cautiously.
Go deeper on this paper
Wrote a paper? Stress-test it before reviewers do.
Upload your manuscript for the same structured scrutiny — reasoning, statistics, evidence, figures, and reproducibility.