On October 26, 2026, a new study published on arXiv (arXiv:2610.31683v1) formally introduced the Schema Coherence Inventory (SCI), aiming to address a key challenge faced by persona-conditioned language models when simulating patients or emotional agents: how to distinguish whether the model’s behavior is genuinely guided by internalized beliefs or merely reflects superficially consistent responses driven by emotional arousal.
SCI comprises 96 forced-choice items, covering the eight levels of Beck’s cognitive model. Each item offers six response options, corresponding to three types of negative self-schemas and their positive counterparts. Scoring is based on length-normalized log-likelihood values, with no text generation or parsing required throughout the process. Its core criterion is “target specificity”—that is, a condition specifically constructed for a particular belief must activate that belief more frequently than other similar schema conditions.
The study conducted 107 conditional tests across five different model sizes (1.7B–14B) using six validation experiments; nine clinically trained raters were able to identify predefined beliefs with 79.5% accuracy (compared to a chance level of 16.7%); all 30 belief items showed a significant target-specific improvement (+17.7 points); the SCI profile correctly matched 24 out of 30 free‑response items (far exceeding the 5 matches expected by chance). Additionally, the 48-item short form demonstrated a correlation coefficient of r = 0.73 with the original scale.
In the real-world application of six published patient simulators, none proved more effective at conveying their claimed core beliefs than a single-sentence statement. The study highlights that this validation approach separates cognitive constructs from accompanying emotions and uses clinical judgment as the gold standard, providing a scalable template for measuring emotion and belief in language models. The authors have simultaneously made available all items, scoring code, and a brief version of the scale.