Evaluating Developmental Cognition Capabilities of LLMs

Conversational AI typically treats users as cognitively homogeneous, overlooking differences in how people make sense of what models say. We bring Robert Kegan's constructive-developmental theory into LLM evaluation as a complementary user-modeling lens, distinct from preferences, expertise, or demographics. Existing assessment methods rely either on expert interviews that do not scale or on sentence-completion instruments that are proprietary, lengthy, or invasive.

We introduce the Developmental Sentence Completion Test (DSCT), a 20-item instrument designed to elicit developmental signal in self-administered text, and use it to ask how much of that signal can be recovered by LLMs across three elicited response regimes: simulated personas, real human respondents, and default model-generated answers.

On simulated personas, top frontier models recover simulator-intended stage labels with high accuracy, but agreement with trained human raters is lower, indicating that simulator and classifiers may share priors about how stages should appear in text. On real human DSCT responses, human-LLM agreement is fair, with much stronger within-neighborhood than exact agreement, suggesting that developmental signal is recoverable but noisy even under structured elicitation. Finally, when LLMs answer DSCT prompts without persona-conditioning, their responses exhibit stable stage-like differences across model families, with larger and newer models tending to generate higher-stage-seeming text. These results suggest that the core constraint for stage-aware conversational AI is not classifier accuracy alone, but the recoverability of developmental signal from elicited text. DSCT offers a first step toward benchmarking that possibility.