Interpretable Symptom Vectors for Depression in a Large Language Model
Fangyi Zhu, Ajay Subramanian, Allison Constant, Camille Wang, Ravish Gupta, Corey J. Keller
cs.CL · cs.AI · cs.LG · q-bio.NC
Quick verdict
Maps clinician-aligned depression symptom vectors in Gemma-3 activations (held-out AUC 0.789); rigorous stats but small n, LLM-rewrite confound, no baselines.
Summary
The paper asks whether depression symptoms are linearly encoded in an LLM's residual stream. The authors extract activations from Gemma-3-27B-PT for 51 clinical-instrument excerpts annotated into mood, somatic, and suicidality groups, and use PERMANOVA gated by PERMDISP across 63 layers and eight distance metrics to select layer 21 as the point of maximal group separation (pseudo-F up to 6.91). Symptom Vectors built from group centroids at that layer, decorrelated via a Gram-matrix pseudoinverse (condition number ≈ 14,612), yield three coefficients per passage. On a held-out naturalistic corpus (423 excerpts from Darkness Visible, a clinical handbook, and ReDSM5), the matching annotation group attains the highest median coefficient on each axis, preserving clinician rank order, though the mood axis separates weakly. A single Depression Vector separates held-out depressive text from HappyDB happy moments with AUC 0.789 — a population-level, not per-excerpt, effect. Strengths include the PERMDISP eligibility gate and honest in-sample/held-out bookkeeping; weaknesses include a tiny suicidality group (n = 8), rewriting of every excerpt by Claude Opus 4 as a potential confound, absence of supervised-probing or embedding baselines, and no clinical outcome validation. Code and rewritten corpora are released.
Review complete
Assessment at a glance
6
risk flags identified
● high 1
● medium 3
● low 2
5 strengths · 5 weaknesses · 5 author questions
AI text signal = Binoculars zero-shot detector score over 20 sampled passages (Fast-DetectGPT co-reported) — absence does not prove human authorship, and a flag does not imply misconduct.
Levels reflect verification complexity — not a quality score. Evolve never rates your paper.
Key claims under review
- Depression symptom groups (mood/emotional/cognitive, somatic, suicidality) are geometrically separable in the residual stream of Gemma-3-27B-PT, with maximal separability at layer 21 across three raw distance metrics.
- Gram-pseudoinverse-decorrelated projection onto centroid-based Symptom Vectors yields per-symptom coefficients that preserve clinician-annotated rank ordering on held-out naturalistic text.
- A single layer-21 Depression Vector separates held-out depressive text from HappyDB happy moments with AUC = 0.789, functioning as a population-level emotional valence gate for symptom projection.
- The resulting decorrelated, clinician-aligned signal offers a mechanistic foundation for interpretable depression assessment, though the authors disclaim any current clinical validity.
Strengths
Dispersion-controlled layer selection
The PERMDISP eligibility gate (§2.2.2) prevents within-group dispersion differences from masquerading as centroid separation; the three raw distance metrics converge independently on layer 21 (pseudo-F 6.91/6.49/2.83 with PERMDISP P = 0.43/0.65/0.47), and the weakly eligible later maxima (layers 35–40) are disclosed rather than hidden (Table 3, Fig. 2).
Honest in-sample vs held-out bookkeeping
Core Clinical projections are explicitly labeled an in-sample consistency check (§2.3), the in-sample AUC of 1.000 is 'not evidence of generalization' (§3.3), and only Naturalistic-vs-HappyDB (AUC 0.789, P = 2.51e-48) is treated as the generalization test — an unusual level of candor.
Clinically grounded, dual-annotated stimuli
Principled handling of collinearity
Reproducibility artifacts released
Weaknesses
Fragile statistical base
Universal LLM rewriting is an un-ablated confound
Weak mood-axis evidence vs the framing
No baselines
Population-level effect only
Full Review Report
$1.99 · one-timeUnlock the complete preliminary analysis Evolve produced for this paper.
Report preview
Scope & Claims
This arXiv preprint (2609.01832; Stanford psychiatry with academic and industry coauthors) asks whether a clinically validated depression symptom taxonomy is linearly encoded inside a frozen LLM. Three claims: (1) mood/emotional/cognitive, somatic, and suicidality text groups are geometrically separable in Gemma-3-27B-PT's residual stream, maximally at layer 21 of 62; (2) projecting held-out naturalistic text onto centroid-based Symptom Vectors, decorrelated via a Gram-pseudoinverse, preserves clinician-annotated rank ordering; (3) a single Depression Vector separates held-out depressive from non-depressive text (AUC = 0.789) and can act as a valence gate for symptom projection. The authors explicitly disclaim clinical validity.
The remaining sections cover Methodology Assessment, Statistical Evidence, Results & Reproducibility, Limitations & Risks, and Verdict & Recommendations.
- All 5 strengths & 5 weaknesses, analyzed in detail
- 6 risk flags with severity, section & reviewer questions
- Methods & statistical evidence assessment
- 5 questions a peer reviewer would ask
- Full six-dimension evaluation report & recommendations
One-time purchase · this paper only
One-time purchase, this paper only. AI-generated preliminary analysis — not peer review.
Frequently asked questions
What are 'Symptom Vectors' for depression in an LLM?
They are directions in the model's internal activation space, computed as the average activation of text describing each of three symptom groups (mood/emotional/cognitive, somatic, suicidality) drawn from validated clinical instruments. Projecting a passage onto these directions yields one coefficient per symptom, intended as an interpretable per-symptom signal, with a Gram-matrix correction to decorrelate the three overlapping axes.
How accurate is the depression detection (AUC 0.789)?
The AUC of 0.789 means a randomly chosen depressive text scored higher than a randomly chosen happy-moment text about 79% of the time in a held-out comparison (Naturalistic corpus vs HappyDB). The authors stress this is a population-level effect with heavily overlapping score distributions (cosines between 0.18 and 0.25) and is not reliable for classifying individual excerpts.
Which model and layer did the study use?
Google's pre-trained (not instruction-tuned) Gemma-3-27B. Activations were recorded at all 63 hook points of the 62-block model, and layer 21 was selected because three raw distance metrics (Centroid Euclidean, Energy, Earth Mover's) all peaked there with homogeneous within-group dispersion.
Is this ready for clinical use on real patients?
No. The authors explicitly state that agreement with clinician text annotations is not clinical validity: nothing has been validated against diagnoses, treatment response, or longitudinal outcomes, and the method should not be construed as a clinical instrument.
About this review
This is an AI-generated preliminary review by Evolve.Pub, produced from the openly available text of the paper. It is not peer review and does not represent the authors' views. The canonical record of the paper is at the publisher link above.
- Figure-embedded text (pairwise significance brackets in Fig. 3, axis annotations in Fig. 2) is partially garbled by PDF extraction; significance calls were cross-read from the surrounding figure text and Table 4 medians, which extracted cleanly. The document ends with the supplementary PROCEED symptom descriptions — a natural end, with no truncation detected.
Go deeper on this paper
Wrote a paper? Stress-test it before reviewers do.
Upload your manuscript for the same structured scrutiny — reasoning, statistics, evidence, figures, and reproducibility.