Do LLMs show reproducible psychological profiles when given standardized tests?

The paper administers seven psychological instruments to nine LLMs, five repeated times per model and language, in Chinese and English. It reports that "LLMs exhibit structured, model-specific profiles despite a shared alignment-shaped pattern." Across the battery the models lean toward higher prosocial, self-regulatory and stability-related responses and lower endorsement of dominance, disengagement and harmful intent. Within that common lean, each model still has its own configuration, and repeated administrations are reproducible enough to "permit recovery of model identity." The stated aim is not to infer human-like personalities or internal psychological states. It is to test whether standardized instruments can elicit "reproducible, interpretable and model-specific" response signatures.

The second half of the claim concerns what the models do not answer. Items still unresolved after a prespecified retry procedure are kept as NA rather than dropped, and the paper analyzes scored and NA responses jointly. The NA responses are "structured rather than uniformly distributed," which the authors read as showing where an item is treated as inapplicable, refused, or cannot be mapped to a valid response option. On this reading, the boundary of where self-report applies to a model is itself part of its profile. Language condition and provider origin are both associated with profile configuration and answerability.

Set against the library, this reverses the direction of Can language summaries unlock hidden psychological patterns?, where the model is the profiler of human scores. Here the instruments are turned on the model. The shared pattern fits the case in Does preference optimization harm conversational understanding? that alignment leaves a systematic imprint on behavior, and the paper's "alignment-shaped" wording says the same. The convergence across models echoes the correlated errors in Can AI systems learn social norms without embodied experience?. The language association loosely parallels the cross-linguistic variation in Do users worldwide trust confident AI outputs even when wrong?, though this excerpt says nothing about user reliance.

The excerpt does not establish which instruments or models were used, how large the between-model differences are, or how well identity recovery works. It does not describe the human-reference analyses or the direction of the context dependence. It says only that the signatures are "context dependent" under those analyses and prompt-robustness checks. It also does not show how the alignment attribution was tested, and it disclaims any claim about internal states. The defensible reading is narrow: standardized self-report probes can characterize a deployed model's response regularities, including its refusals, as a quantitative behavioral signature. That signature should be treated as tied to the prompt, language and provider conditions under which it was measured.