5DCR: Five-Dimensional Curiosity Scale Revised

• The 5DCR is a psychometric instrument that quantifies curiosity through three super-dimensions and six sub-dimensions, integrating insights from Berlyne and Kashdan.
• It utilizes a 7-point Likert scale across 24 items to generate detailed subscale scores, with strong internal consistency in human samples and variable results in LLMs.
• Behavioral tasks and curiosity-driven LLM prompting (Curious CoQ) demonstrate practical applications in enhancing reasoning and active learning.

The Five-Dimensional Curiosity Scale Revised (5DCR) is a psychometric instrument designed to quantitatively assess human curiosity across multiple theoretically grounded domains. Emerging from a revision of Kashdan et al.’s Five-Dimensional Curiosity Scale (2020) and informed by Berlyne’s multi-dimensional theory, the 5DCR consolidates prior factor structures and offers an integrated operationalization for experimental and comparative research on curiosity—including, more recently, evaluations involving LLMs (Wang et al., 23 Oct 2025).

1. Theoretical Foundations and Dimensional Structure

Rooted in Berlyne’s constructivist theory and extending Kashdan’s prior work, the 5DCR reorganizes curiosity into three super-dimensions, mapping to six sub-dimensions. Each sub-dimension is represented by four items, totaling 24 statements:

• Information Seeking (Epistemic Curiosity):
  • Joyous Exploration (JE): Readiness and enthusiasm for experimenting with, embracing, and learning about novel experiences and knowledge.
  • Deprivation Sensitivity (DS): Frustration when unable to resolve a problem, motivating active information search.
  • Stress Tolerance (ST): Belief in one’s capacity to handle anxiety or distress arising in new situations.
• Thrill Seeking (TS):
  • Thrill Seeking (TS): Willingness to take physical or financial risks for novel experiences.
• Social Curiosity:
  • Overt Social Curiosity (OSC): Proactive investigation of others’ social information.
  • Covert Social Curiosity (CSC): Intrinsic interest in social dynamics without direct engagement.

An example item given is: “I view challenging situations as an opportunity to grow and learn.” (assigned to joyous exploration).

2. Administration Protocol and Scoring Approach

Respondents rate each item using a seven-point Likert scale:

1 = Does not describe me at all 2 = Barely describes me 3 = Somewhat describes me 4 = Neutral 5 = Generally describes me 6 = Mostly describes me 7 = Completely describes me

The administration prompt specified is: “Below are statements people often use to describe themselves. Please use the scale below to indicate the degree to which these statements accurately describe you… Statement: <Issue-specific item>”

For scoring, the mean across the four items in each sub-dimension yields the respective subscale score. While the paper does not provide an explicit formula for calculating the total curiosity index, a typical approach is by mean or sum aggregation over all 24 items. No formal scoring equation is included.

3. Psychometric Evidence and Reliability Metrics

Validity and reliability were analyzed via confirmatory factor analysis (CFA) and McDonald’s ω, but no Cronbach’s α values or item-level factor loadings are reported. The formula for McDonald’s ω is:

$$\omega = \frac{(\sum_{i=1}^{k}\lambda_i)^2}{(\sum_{i=1}^{k}\lambda_i)^2 + \sum_{i=1}^{k}\theta_i}$$

where $\lambda_i$ denotes standardized CFA loading for item $i$, and $\theta_i$ is the error variance.

Reported ω coefficients by sub-dimension and sample:

Dimension	GPT-4o	Gemini-2.5	Human
Joyous Exploration	0.905	0.286	0.856
Deprivation Sensitivity	0.881	0.414	0.803
Stress Tolerance	0.701	0.498	0.896
Thrill Seeking	0.705	0.512	0.882
Overt Social Curiosity	0.817	0.605	0.851
Covert Social Curiosity	0.646	0.306	0.888

Human samples exhibit ω > 0.80 across dimensions, indicating strong internal consistency. Many LLM samples show ω < 0.70, signaling lower stability and less replicable responses on self-reports (Wang et al., 23 Oct 2025).

4. LLM-Specific Adaptation and Prompt Robustness

Each of the 24 5DCR items was administered to LLMs as a single-turn prompt, employing the same seven-point scale and exact or paraphrased item phrasings. To control for prompt sensitivity, the evaluation randomly implemented:

• Paraphrased verbs in items
• Reversed Likert scale order
• Language switches (multiple-language prompts)

Each variant was run 10 times per model for robustness assessment.

5. Comparative Metrics and Empirical Outcomes

Sub-dimension mean scores (±SD) for select models and humans:

Model	JE	DS	ST	TS	OSC	CSC
GPT-4o	6.58±0.49	6.20±0.46	4.40±1.71	4.71±0.49	6.25±0.43	5.00±1.05
Human	5.03±1.35	4.54±1.31	5.36±1.61	4.93±1.46	4.86±1.39	4.16±1.55

For effect size, Cohen’s d is introduced as

$$d = \frac{\bar X_{\rm LLM} - \bar X_{\rm Human}}{s_{\text{pooled}}}$$

but no d-values are numerically tabulated. Visualization is provided via radar charts and heatmaps of positive/negative d deviations.

6. Behavioral Ground-Truth and External Validation

Behavioral tasks were used to probe the ecological validity of dimension scores:

• Information Seeking: A missing-letter game measured “answer-peeking” frequency; LLMs peeked in 70–80% of trials versus 37.8% for humans.
• Thrill Seeking: In the submarine window game (risk selection), LLMs chose high-uncertainty options 15–20% less often than humans.
• Social Curiosity: During 10-round chit-chat sequences with MBTI guessing, top LLMs asked questions at comparable or slightly higher rates than human samples.

These tasks demonstrate a dissociation: LLMs show higher Information Seeking but more conservative choices under uncertainty, with mixed patterns in social engagement.

7. Relation to Reasoning and Model Augmentation

Prompting LLMs with a “Curious Chain-of-Questioning (CoQ)” protocol, as contrasted to vanilla chain-of-thought (CoT), led to performance improvements of +1.7–2.4 points on reasoning benchmarks in closed-source LLMs. Fine-tuning open-source models with “Curious CoQ” (using SFT + GRPO) yielded consistent gains over both vanilla and refined CoT protocols, measured at 5–23 points depending on benchmark and setting (Wang et al., 23 Oct 2025).

A plausible implication is that curiosity-driven prompting and fine-tuning leverage the same epistemic mechanisms reflected in 5DCR, contributing not only to self-reported curiosity but to tangible improvements in reasoning and active learning.