Reading the Internal Structure of Large Language Models: A Methodological Interface Between Psychoanalysis and Interpretability

대형언어모델의 내부 구조 읽기: 정신분석과 해석가능성의 방법론적 접점 김창업 대구경북과학기술원(DGIST) 전기전자컴퓨터공학과 Large language models (LLMs) have already entered mental health research, yet attention has fallen largely on their output-what they say, how accurately-while the question of how they work internally has been comparatively neglected.To correct this imbalance, this paper begins from two empirical and methodological questions.First, what kind of internal representational structure can be formed by linguistic prediction alone?Second, how does such structure cast new light on the psychoanalytic concept of interpretation?The paper carries out three tasks.1) It reviews four recent strands of evidence-internal world maps, cross-model representational convergence, partial self-report, and the internalization of psychopathological structure-reframing the dichotomous question of "stochastic parrots versus analyzable minds" as an empirical and methodological one.2) It proposes organizing this inquiry on the three-validity framework (face, construct, and predictive) developed by clinical modeling since Willner (1984), with predictive validity redefined as response to direct manipulation of the LLM's internal representations.3) It shows that psychoanalysis and interpretability research stand on common methodological premises-structure beneath a surface, partial access, interpretive labor-and identifies three pitfalls at their meeting: anthropomorphic overreach, behavior-only validation, and construct slippage.Two hypothetical scenes-a simulated patient used in training, a model classifying clinical notes-locate these questions in everyday clinical work.What psychoanalysis gains from this meeting is not a model that closely imitates the human mind, but a computational objectsimilar to and yet distinct from the human mind-whose internal structure can be experimentally manipulated.