Automated scoring of adult ego development from sentence completions: interpretable maturity indicators derived by evolutionary prompt optimization
TL;DR
An automated scoring approach in which an evolutionary algorithm derives a small set of interpretable maturity indicators, each scored from sentence-completion responses by a domain-fine-tuned language model, supports automated ego-development scoring at agreement levels comparable to trained raters, and the indicators offer interpretable candidate dimensions of adult development for confirmatory and population-scale research.
Abstract
Ego Development Theory (EDT) describes systematic changes in how adults make meaning, but its principal measure, the Sentence Completion Test (SCT), requires resource-intensive expert scoring. We developed an automated scoring approach in which an evolutionary algorithm derives a small set of interpretable maturity indicators, each scored from sentence-completion responses by a domain-fine-tuned language model; a classifier trained on the resulting scores then predicts developmental stage. Performance was evaluated on an expert-scored dataset never used for feature discovery ( N = 851), in a fully blinded comparison with an independently trained scorer ( N = 47), and against external criteria ( N = 446). The classifier agreed exactly with expert ratings in 86.1% of cases (quadratic-weighted κ = 0.93). By contrast, zero-shot prompting of two frontier language models with the full scoring manual yielded only 40.1–44.9% agreement, indicating that expert-comparable scoring is not achieved by prompting alone. The classifier's agreement replicated at 85.1% in the blinded comparison. In external validation, the indicators correlated with personality, coherence, and autonomy measures in theoretically expected patterns, and explained substantial additional variance in expert-rated developmental stage beyond Big Five traits (Δ R 2 = 0.59). In further analyses, they were also consistent with EDT's predicted alternation of differentiation and integration across stages. The Language Complexity indicator agreed with human-rated Integrative Complexity (ICC = 0.73 and 0.77 in two corpora). Exploratory analyses in five non-SCT corpora suggested the indicators predict domain-relevant behavior in other text genres. These results support automated ego-development scoring at agreement levels comparable to trained raters, and the indicators offer interpretable candidate dimensions of adult development for confirmatory and population-scale research.
