This paper investigates whether large language models (LLMs) internally represent psychological constructs or merely generate outputs that mimic psychometric structure. We adopt a cross‑persona paradigm, evaluating 14 model variants from the Llama 3 and Qwen 2.5 families (sizes 0.5B‑14B, both base and instruct modes). Human data consist of responses from 272 Taiwanese university students on seven psychological scales; one scale is used as input for the models to predict the remaining six scales.\n\nAt the behavioral level, larger variants reproduce and amplify human cross‑scale correlations. This amplification persists after correcting the seven‑scale human correlation matrix for measurement attenuation. Contrastive direction analysis reveals that the geometry of larger instruct variants aligns closely with the human matrix, whereas base variants show little alignment. Across models, geometric strength correlates with behavioral amplification ($r=0.68$, $p=0.008$; partial $r=0.69$ when controlling for log size).\n\nTo link internal structure to output behavior, we employ a matched activation‑probing paradigm. Probe slopes vary less than behavioral slopes (1.38‑1.77 vs 0.52‑1.53). A synthetic control shows ridge probing can inflate these slopes; after subtracting a baseline derived from a reliability proxy in a separate analysis, the adjusted slopes remain strongly associated with behavior ($r=0.88$). We term the gap between probe and behavior readout attenuation.\n\nIn summary, the results demonstrate decodability and geometric alignment of LLM internal representations under the tested assays, but do not establish that the measured directions mediate cross‑persona answers.\n\nReview: The study offers a rigorous framework for probing whether LLMs capture psychometric structure internally, highlighting the pivotal role of model scale and instruction tuning in shaping geometric alignment and behavioral amplification, and setting a solid experimental foundation for future AI‑psychology interdisciplinary work.