Researchers have developed the Developmental Sentence Completion Test (DSCT), a 20-item instrument designed to assess how well large language models can recognize and respond to different stages of human cognitive development.

The study, published on arXiv by researchers Xiao Xiao, Hayoun Noh, and Mar Gonzalez-Franco, addresses a gap in conversational AI personalization. While current systems adapt to user preferences and knowledge, they largely ignore how users interpret and construct meaning from AI outputs.

The research draws on Robert Kegan's constructive-developmental theory, which describes how people make sense of their reality at different cognitive stages. Traditional assessment methods require expert interviews or lengthy proprietary tests that don't scale for AI evaluation.

Testing Three Response Types

The researchers tested their DSCT across three scenarios: simulated personas, real human responses, and default model outputs. Top frontier models showed high accuracy when recovering intended labels from simulated personas.

Human-LLM agreement proved more challenging, with fair correlation overall but stronger agreement within developmental neighborhoods rather than exact stage matches. When models answered DSCT prompts without persona conditioning, responses showed stable stage-like differences across model families.

Larger and newer models consistently generated text rated at higher developmental stages. The findings suggest that stage-conditioned signals appear cleaner in synthetic responses compared to human-written text.

The research indicates that building stage-aware conversational AI faces constraints beyond classifier accuracy alone. The core challenge lies in extracting reliable developmental signals from user-generated text.

This work could inform future AI systems that adapt not just to what users know, but to how they fundamentally process and construct meaning from information. The DSCT offers a scalable alternative to existing developmental assessment methods for AI evaluation contexts.