Language models score psychiatric symptoms by wording, not psychology
Source published ·Modelwire updated
Original coverage: arXiv cs.CL ↗·How Modelwire adds context
The development
Researchers have identified a critical failure mode in language models scoring psychiatric assessments: LLM outputs correlate with questionnaire wording structure rather than actual psychological symptoms. COMPASS 2.0 introduces psychometric representational similarity analysis to distinguish genuine clinical signal from statistical artifacts induced by item phrasing. Across 275 clinical interviews, speech-derived symptom geometry tracked item wording more closely than participant self-report, with no psychological structure detected beyond linguistic patterns. This work exposes a fundamental validation gap in clinical AI applications and suggests that apparent model agreement with human assessments may be illusory, forcing a reckoning around deployment of language models in mental health contexts.
Modelwire’s AI-generated summary of coverage from arXiv cs.CL.
Modelwire analysis
ExplainerOur AI-generated reading of the wider context and the next developments to watch.
The critical finding isn't just that LLMs fail at psychiatric assessment, but that existing validation methods cannot detect this failure. COMPASS 2.0 reveals that models can appear to agree with human clinicians while actually tracking questionnaire phrasing rather than patient psychology, meaning standard accuracy metrics provide false confidence.
This work completes a trilogy of validation gaps documented here over the past week. The October 1st rubric review identified missing dimensions in clinical reasoning evaluation. The Alzheimer's speech study showed acoustic confounds masquerade as pathology. Now COMPASS 2.0 demonstrates that linguistic structure itself becomes a confound in psychiatric contexts. Together, these papers establish that clinical AI deployment requires explicit deconfounding layers, not just benchmark scores. The common thread: apparent model-human agreement masks systematic artifact capture.
If COMPASS 2.0's psychometric RSA method is adopted by any major mental health AI vendor or healthcare system for pre-deployment validation within the next 18 months, that signals the field is taking deconfounding seriously. If psychiatric LLM products continue shipping without this or equivalent structural validation, the gap between research and practice has widened further.
This interpretation is generated from the summary above and the archive coverage cited below. Our methodology · Report an error
Coverage behind this analysis
These archive entries ground the connection in our analysis. They are ordered by source publication date, with links to our coverage and the original sources.
·arXiv cs.CL
Clinical LLM evaluation lacks unified framework for reasoning assessment
Clinical deployment of LLMs hinges on reasoning quality, not test scores alone. This structured review identifies a critical gap in evaluation methodology: existing benchmarks and assessment rubrics fail to measure how well models integrate evidence over time, revise diagnoses, or justify clinical decisions. The analysis spans medical education frameworks, recent LLM benchmarks, and text-generation evaluation…
MentionsCOMPASS 2.0 · Language models · Psychometric representational similarity analysis
How this coverage is produced
Modelwire uses AI to generate summaries and context from source headlines, snippets, and selected archive coverage. Automated checks do not verify every claim, and items are not routinely reviewed by a person before publication. Zacaria Solis operates the site. Read the linked source for the full evidence and report errors through our corrections process.
Modelwire summarizes, we don’t republish. arXiv cs.CL originally reported this story as “COMPASS 2.0: psychometric representational similarity analysis distinguishes symptom structure from personal signal”. The full content lives on arxiv.org. If you’re a publisher and want a different summarization policy for your work, see our takedown page.