LLMs Map Internal Narratives to Depressive States
Researchers quantified the internal narratives of participants to analyze the structure and dynamics of depressive states using large-language-model representations. In two studies involving 1,257 participants, they discovered that verbal accounts of specific symptoms provided detailed insights that could predict self-reported depression levels. Maintaining the covariance among symptoms was essential for ensuring construct validity, suggesting that high-dimensional text representations reflect the underlying geometry of depression. In Study 2, the temporal dynamics were examined as participants interacted with emotional narratives, revealing that changes in quantified internal narratives corresponded with shifts in self-reported affect.
Key facts
- Study involved 1257 participants across two studies.
- Large-language-model representations were used to parameterise internal narratives.
- Verbal descriptions of symptom-specific thoughts predicted depression scores.
- Preserving covariance between symptoms was essential for construct validity.
- High-dimensional text representations mirror latent geometry of depression.
- Study 2 examined temporal dynamics of narrative-affect relationship.
- Changes in internal narratives led to changes in self-reported affect.
- Research published on arXiv with ID 2502.09487.
Entities
Institutions
- arXiv