Non-Mirror Models in Depression Detection

• Non-Mirror models are language-AI systems that infer depression scores from life history interviews rather than direct symptom checklists.
• They aim to reduce criterion contamination by separating predictor content from the structured diagnostic assessment.
• Empirical results show lower effect sizes compared to Mirror models, suggesting a trade-off between reduced bias and increased interpretative ambiguity.

Non-Mirror models are language-AI models of depression developed from language that does not structurally mirror the depression assessment they are used to predict. In the operationalization introduced in "Mirror Language AI Models of Depression are Criterion-Contaminated" (Li et al., 7 Aug 2025), the predicted target is a depression symptom total derived from a structured diagnostic interview, while the predictive input for the Non-Mirror condition is a separate life history interview. The term is defined in direct contrast to Mirror models, which predict a structured interview score from the transcript of that same structured interview. The distinction is intended to isolate criterion contamination, evaluate generalizability, and test whether broader, more naturalistic language retains clinically meaningful signal.

1. Definition and conceptual boundary

Within the reported study, Non-Mirror models are language-AI systems that predict structured diagnostic interview depression scores from transcript material that does not directly answer the same symptom questions used to construct the target score (Li et al., 7 Aug 2025). The target assessment is based on 10 DSM-5 major depressive episode items, scored $0/1$ and summed into a depression symptom total. The Non-Mirror input is a life history interview that is semi-structured, broader, and not designed to directly assess depression.

Mirror models, by contrast, use the structured diagnostic interview transcript itself to predict the depression score derived from that same interview. The distinction is therefore not merely one of prompt wording or architecture; it is a distinction in the relation between predictors and criterion-generating information. In the study’s framing, Non-Mirror models are designed so that the model must infer depression-related severity from discourse that is adjacent to, rather than duplicative of, the formal assessment.

This contrast is methodologically central. A Mirror model can perform well because its input language already contains the content that defines the label. A Non-Mirror model must instead operate on broader semantic cues, narrative structure, contextual descriptions, and indirect symptom-relevant material. This suggests that Non-Mirror models are intended as a stricter test of out-of-assessment inference.

2. Operationalization in the depression-prediction study

The empirical study used $N = 110$ participants recruited from a university setting, and each participant completed two interviews (Li et al., 7 Aug 2025). The first was a structured diagnostic interview based on 10 DSM-5 major depressive episode items, covering depressed mood, anhedonia, appetite/weight change, sleep, psychomotor change, fatigue, guilt/worthlessness, concentration, suicidality, and impairment. The second was a life history interview covering education and employment, relationships and support, sleep, eating, exercise, physical and mental health history, strengths, coping, and future plans.

Three models were evaluated: GPT-4, GPT-4o, and LLaMA3-70B. They were prompted to perform the same sequence of operations across conditions: read an interview transcript, determine whether each of the 10 DSM-5 depression symptoms was present in the past two weeks, output Yes/No, provide a confidence score, extract supporting utterances, and identify the most relevant interviewer question. The prompt framework was held constant while the transcript source changed.

The Non-Mirror condition therefore asked the models to predict the structured interview depression score from the life history interview transcript. The Mirror condition asked the same models to predict that score from the structured diagnostic interview transcript itself. This establishes a head-to-head comparison in which the target remains fixed while the degree of structural overlap between input and criterion changes.

3. Criterion contamination and the rationale for Non-Mirror modeling

The study argues that Mirror models are criterion contaminated because the predicted score depends in part on predictors that already contain the criterion-generating content (Li et al., 7 Aug 2025). In plain terms, the criterion is the depression score from the structured interview, and the predictors are the participant’s responses to those same structured interview questions. The resulting predictor–criterion overlap is treated as a source of artificial effect size inflation.

The paper formulates the logic compactly as

$$\text{Criterion contamination} \approx \text{overlap between predictors and criterion-generating information}.$$

It also gives the standard expression for the coefficient of determination,

$$R^2 = 1 - \frac{\sum_i (y_i - \hat y_i)^2}{\sum_i (y_i - \bar y)^2},$$

to emphasize that if predictors already encode the target content, residual error may shrink in a way that inflates apparent predictive success.

Non-Mirror models are introduced as a methodological counterpoint to this problem. Because the life history interview is not designed to directly assess depression, the input does not mirror the score construction procedure. The study’s interpretation is that lower effect sizes in the Non-Mirror setting should not automatically be read as inferiority; rather, they may indicate reduced contamination and therefore a more realistic test of generalization. A plausible implication is that the Non-Mirror formulation shifts evaluation from symptom restatement toward latent semantic inference.

4. Empirical performance profile

The reported performance gap between Mirror and Non-Mirror conditions is substantial, especially for $R^2$, accuracy, and symptom-level classification metrics (Li et al., 7 Aug 2025). Mirror models show very large effect sizes, whereas Non-Mirror models show smaller but still nontrivial effect sizes.

Model	Mirror condition	Non-Mirror condition
GPT-4	Accuracy $= 0.97$; F1 $= 0.94$; MSE $= 0.67$; MAE $= 0.32$; $R^2 = 0.80$; Pearson $N = 110$0; Spearman $N = 110$1	Accuracy $N = 110$2; F1 $N = 110$3; MSE $N = 110$4; MAE $N = 110$5; $N = 110$6; Pearson $N = 110$7; Spearman $N = 110$8
GPT-4o	Accuracy $N = 110$9; F1 $$\text{Criterion contamination} \approx \text{overlap between predictors and criterion-generating information}.$$0; MSE $$\text{Criterion contamination} \approx \text{overlap between predictors and criterion-generating information}.$$1; MAE $$\text{Criterion contamination} \approx \text{overlap between predictors and criterion-generating information}.$$2; $$\text{Criterion contamination} \approx \text{overlap between predictors and criterion-generating information}.$$3; Pearson $$\text{Criterion contamination} \approx \text{overlap between predictors and criterion-generating information}.$$4; Spearman $$\text{Criterion contamination} \approx \text{overlap between predictors and criterion-generating information}.$$5	Accuracy $$\text{Criterion contamination} \approx \text{overlap between predictors and criterion-generating information}.$$6; F1 $$\text{Criterion contamination} \approx \text{overlap between predictors and criterion-generating information}.$$7; MSE $$\text{Criterion contamination} \approx \text{overlap between predictors and criterion-generating information}.$$8; MAE $$\text{Criterion contamination} \approx \text{overlap between predictors and criterion-generating information}.$$9; $$R^2 = 1 - \frac{\sum_i (y_i - \hat y_i)^2}{\sum_i (y_i - \bar y)^2},$$0; Pearson $$R^2 = 1 - \frac{\sum_i (y_i - \hat y_i)^2}{\sum_i (y_i - \bar y)^2},$$1; Spearman $$R^2 = 1 - \frac{\sum_i (y_i - \hat y_i)^2}{\sum_i (y_i - \bar y)^2},$$2
LLaMA3-70B	Accuracy $$R^2 = 1 - \frac{\sum_i (y_i - \hat y_i)^2}{\sum_i (y_i - \bar y)^2},$$3; F1 $$R^2 = 1 - \frac{\sum_i (y_i - \hat y_i)^2}{\sum_i (y_i - \bar y)^2},$$4; MSE $$R^2 = 1 - \frac{\sum_i (y_i - \hat y_i)^2}{\sum_i (y_i - \bar y)^2},$$5; MAE $$R^2 = 1 - \frac{\sum_i (y_i - \hat y_i)^2}{\sum_i (y_i - \bar y)^2},$$6; $$R^2 = 1 - \frac{\sum_i (y_i - \hat y_i)^2}{\sum_i (y_i - \bar y)^2},$$7; Pearson $$R^2 = 1 - \frac{\sum_i (y_i - \hat y_i)^2}{\sum_i (y_i - \bar y)^2},$$8; Spearman $$R^2 = 1 - \frac{\sum_i (y_i - \hat y_i)^2}{\sum_i (y_i - \bar y)^2},$$9	Accuracy $R^2$0; F1 $R^2$1; MSE $R^2$2; MAE $R^2$3; $R^2$4; Pearson $R^2$5; Spearman $R^2$6

The strongest reported Mirror result is GPT-4 with $R^2$7. The strongest reported Non-Mirror result is GPT-4 with $R^2$8. The paper treats the first as likely inflated by criterion contamination and the second as evidence that depression-related severity can still be inferred from broader language that does not mirror the assessment.

A further comparison uses self-reported PHQ-9 symptoms as an external criterion. For GPT-4, the Mirror-predicted structured interview scores correlated with PHQ-9 at Pearson $R^2$9 and Spearman $= 0.97$0, while the Non-Mirror-predicted scores correlated at Pearson $= 0.97$1 and Spearman $= 0.97$2. The study summarizes this as roughly similar performance, around $= 0.97$3. It also reports that PHQ-9 and DSM-based scores correlate at Pearson $= 0.97$4 and Spearman $= 0.97$5. The stated interpretation is that the striking Mirror advantage against the structured interview criterion contracts when evaluated against an independent symptom measure.

5. Semantic structure, topic modeling, and error modes

To examine what the models were using as evidence, the study applied BERTopic to utterances selected by the models as support for depressive symptoms (Li et al., 7 Aug 2025). Some thematic clusters appeared in both Mirror and Non-Mirror outputs, including sleep, appetite/eating, worthlessness/guilt, concentration/difficulty thinking, suicidal thoughts, and relationships and social context. This indicates that the two conditions share clinically recognizable semantic content, even though only one directly mirrors the formal assessment.

The conditions differ in the granularity of their topic structure. Mirror-selected utterances were described as tightly clustered and closely aligned with DSM-5 symptom dimensions such as sleep, appetite, suicide, cognition, and guilt. Non-Mirror clusters were described as broader and more diffuse, with themes including school/work stress, social relationships, confidence, change, health/life circumstances, and sleep and appetite. The study interprets this as evidence that Non-Mirror models rely not only on explicit symptom phrases but also on broader contextual correlates of depression.

The paper also compares true positives and false positives in the Non-Mirror setting. True positives tended to involve explicit emotional distress and phrases such as “scary,” “hard,” “fatigue,” “depressed,” and “can’t calm it down.” False positives often involved discussion of past history, the word need, and talk about change, improvement, or prior struggles. The authors interpret this as an error mode in which the model may confuse past depression with current depression, or general distress with DSM-level symptom endorsement in the past two weeks. This suggests that time-sensitive reasoning is a central unresolved issue for Non-Mirror inference.

A related analysis tested whether Non-Mirror performance was driven mainly by the two overlapping symptoms between interviews, sleep and appetite. The result reported is that removing any single symptom had only small effects. The largest symptom importance scores were Q1 depressed mood: $= 0.97$6, Q6 fatigue/energy loss: $= 0.97$7, Q9 suicidal ideation: $= 0.97$8, and Q4 sleep disturbance: $= 0.97$9, where importance was defined as

$= 0.94$0

The stated implication is that Non-Mirror performance was not dominated by the overlapping sleep/appetite content.

6. Interpretation, utility, and limits of the concept

The paper’s central claim is that Non-Mirror models may provide a more generalizable evaluation regime than Mirror models, precisely because they are not built from language that directly reproduces the scoring rubric (Li et al., 7 Aug 2025). In this framing, large Mirror effect sizes are not simply strong predictive results; they may partly index structural overlap between input and criterion. Non-Mirror models, despite lower $= 0.94$1, are presented as closer to realistic clinical conditions in which people speak naturally rather than answer a symptom checklist.

The paper further argues that Non-Mirror models may have clinical utility because they can operate on routine dialogue, detect depression from indirect or narrative language, scale across settings, and identify semantic features that standard structured assessments may not foreground. Topic modeling is used to support the claim that the inferred signal remains interpretable, spanning fatigue, sleep problems, appetite changes, emotional strain, relationship difficulties, and school or work stress.

At the same time, the reported error profile places clear limits on the present formulation. The confusion between past and current symptoms, and between broad distress and recent DSM-level endorsement, indicates that Non-Mirror models do not remove inferential ambiguity; they relocate it. A plausible implication is that the principal trade-off is between reduced criterion contamination and increased temporal and contextual ambiguity.

The term Non-Mirror also has a separate history in other arXiv literatures, especially mirror symmetry and Landau–Ginzburg theory, where it can denote different phenomena altogether: a full naive mirror that is “too large” for a non-nef toric phase (Ballard et al., 2013), a Landau–Ginzburg model with no projective mirror (Ballico et al., 2016), or heterotic $= 0.94$2 constructions that are genuinely beyond the old $= 0.94$3 mirror framework (Bertolini et al., 2018). In the depression-modeling literature, however, the term has a specific methodological meaning: prediction from language that does not structurally mirror the assessment being predicted.