Large language models as windows on the mental structure of psychopathology
Keywords: LLMs, psychopathology, mental structure, computational psychiatry
Abstract: How people represent the world determines how they act on it, as these internal representations bias what information is retrieved from memory, the inferences that are made and which actions are preferred. The structure of these representations are built through experience by extracting relevant information from the environment. Recent research has demonstrated that representational structure can also respond to the internal motives of agents, such as their aversion to uncertainty, which impacts their behavior. This opens the possibility to directly target internal structures to cause behavioral change in psychopathologies, one of the tenets of cognitive-behavioral therapy. For this purpose, it is crucial to understand how internal structures differ across psychopatologies. In this work, we show that Large Language Models (LLMs) could be viable tool to infer structural differences linked to distinct psychopathologies. We first demonstrate that we can reliably prompt LLMs to generate (verbal) behavior that can be detected as psychopathological by standard clinical assessment questionnaires. Next, we show that such prompting can capture correlational structure between the scores of diagnostic questionnaires observed in human data. We then analyze the lexical output patterns of LLMs (a proxy of their internal representations) induced with distinct psychopathologies. This analysis allows us to generate several empirical hypotheses on the link between mental representation and psychopathologies. Finally, we illustrate the usefulness of our approach in a case study involving data from Schizophrenic patients. Specifically, we show that these patients and LLMs prompted to exhibit behavior related to schizophrenia generate qualitatively similar semantic structures. We suggest that our novel computational framework could expand our understanding of psychopathologies by creating novel research hypotheses, which might eventually lead to novel diagnostic tools.
Primary Area: applications to neuroscience & cognitive science
Code Of Ethics: I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics.
Submission Guidelines: I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2025/AuthorGuide.
Anonymous Url: I certify that there is no URL (e.g., github page) that could be used to find authors’ identity.
No Acknowledgement Section: I certify that there is no acknowledgement section in this submission for double blind review.
Submission Number: 4563
Loading