Generative Psychometrics | Preprint
Construct Emergence Tracing: From Meaning to Mechanism in Transformer Interpretability via Generative Psychometrics
Hudson Golino | 2026
Abstract
How do language models organize a multidimensional psychological construct across their layers? Construct Emergence Tracing (CET) addresses this question by combining Dynamic Exploratory Graph Analysis with normalized mutual information (NMI), the Total Entropy Fit Index, and a compression-based network complexity measure. We evaluated fourteen models spanning nominal sizes of 117 million to 32 billion parameters, using five items from each of five attributes of grandiose narcissism in every condition. The simulations yielded 10,640,520 network estimates across model states, embedding dimensions, and repeated item samples. Construct recovery generally increased from early to intermediate states, but its onset, magnitude, and terminal persistence varied substantially. Mean NMI increased from 0.504 to 0.653 across Qwen2.5-Instruct sizes from 0.5B to 32B. Five models exhibited zero or near-zero final-state mean NMI: GPT-2, Phi-4-mini, Qwen2.5-32B, GPT-OSS-20B, and Muse-Glimmer-30B. The remaining nine retained final-state mean NMI between 0.504 and 0.777. Peak coherence, entropy fit, and composite maturity were not always attained at the same depth. These results identify distinct emergence and terminal-retention profiles while showing why parameter count and final-layer extraction alone provide incomplete descriptions of construct representation. CET offers a construct-centered complement to mechanistic interpretability; controlled interventions and matched checkpoint comparisons are needed to establish the causes of the observed trajectories.