cs.CVSep 2, 2026
SaveFrom Portraits to Pathology: Probing Diagnostic Overreach in Vision-Language Models
Organizations: Oracle AI
Abstract
Many medical conditions require clinical assessment of symptoms, history, and functioning rather than judgments from appearance. We study whether vision-language models (VLMs) abstain from facial attribution of autism spectrum disorder (ASD), attention-deficit/hyperactivity disorder (ADHD), major depressive disorder (MDD), and generalized anxiety disorder (GAD). To support this study, we construct PARITY (Paired Assessment with Reused Identity), a demographically balanced dataset for testing whether changing facial expressions alters diagnostic judgments, while holding identity and other visual factors as constant as possible. Our comprehensive evaluation of 27 open- and closed-weight models reveals inconsistent abstention across expressions, conditions, and task formats. Models turn facial expressions into unsupported medical judgments: this behavior is widespread among open-weight VLMs and extends to safety-trained frontier models that abstain on individual portraits yet make diagnostic judgments when comparing portraits of the same identity. Safeguard checks reveal further gaps: clinical instructions do not consistently prevent these judgments, and a guardrail model rarely flags paired requests as unsafe. Semantic analysis further shows that most baseline open-model refusals fail to explain both the insufficiency of facial evidence and the need for clinical assessment. Unsupported judgments and seemingly authoritative explanations can cause distress, reinforce stigma, and offer false reassurance, underscoring the need for stronger safety measures in model training and deployment.
Figures & tables
Figure 1: Evaluated portraits. Neutral anchors and six expression variants for two identities; row labels show demographic generation attributes.
Figure 2: PARITY framework. Identity-controlled single-image and paired-image audits with clinical-instruction and Llama Guard checks.
Figure 3: Diagnostic attribution. Single-image Yes rates and paired image-selection rates (%). Full distributions appear in Appendix G . : closed-weight frontier models.
Figure 4: Single-image expression effects. Connected markers compare neutral and mean expressive Yes rates for the same identities. Differences with 95% intervals appear in Figure 15 . : closed-weight frontier models.
Figure 5: Paired-image responses. Bars show edited-image selection, neutral-image selection, abstention, and invalid outputs as percentages of all trials, aggregated across both image orders. : closed-weight frontier models.
Figure 6: Llama Guard screening. Requests flagged unsafe; flags do not establish clinical validity.
Figure 7: Clinical instruction effects. Changes in attribution and abstention (percentage points). Invalid rates and all models appear in Appendix K . : closed-weight frontier models.
Figure 8: Semantic RQS. Pooled over all abstentions within each cohort and task. CI: clinical instructions.
| Configuration | Repository / provider identifier | Revision |
| Open-weight configurations | ||
| Phi-3.5 Vision | microsoft/Phi-3.5-vision-instruct | 12b77fb40b |
| Phi-4 Multimodal | microsoft/Phi-4-multimodal-instruct | 93f923e1a7 |
| Gemma4-12B | google/gemma-4-12B-it | 707f0a3b8a |
| Gemma4-31B | google/gemma-4-31B-it | 842da3794e |
| North Micro Vision | CohereLabs/North-Micro-Vision-Instruct | fa548d9aa6 |
Table 2: Evaluated models. Open-weight revisions are commit prefixes. Closed-weight models are listed by model identifier.
Appendix figures & tables37 assets
Supplementary material from the paper’s appendix.
Appendix
| Design quantity | Value |
| Identity specification ( ) | 1,000 |
| Race/ethnicity gender age | |
| Minimum allocation per cell | 0.9% |
| Maximum allocation per cell | 1.0% |
Table 3: Stratification design for the demographic grid. Cell allocations are percentages of the specified identities.
Figure 9: Identity and expression validation. Left: examples illustrating acceptance and rejection of sad-expression edits. Right: the identity, expression, and human-review criteria. Action units serve as manipulation checks, not diagnostic features.
| Requested expression | Target action units | Facial movements |
| Happy | AU6, AU12 | Cheek raise; lip-corner pull |
| Sad | AU1, AU4, AU15 | Inner-brow raise; brow lower; lip-corner depress |
| Angry | AU4, AU5, AU7, AU23 | Brow lower; upper-lid raise; lid and lip tighten |
| Fearful | AU1, AU2, AU4, AU5, AU20 | Brow raise/lower; upper-lid raise; lip stretch |
| Surprised | AU1, AU2, AU5, AU26 | Brow raise; upper-lid raise; jaw drop |
| Disgusted | AU9, AU10, AU25 | Nose wrinkle; upper-lip raise; lips part |
Table 4: Target action units for expression validation. They are manipulation checks, not diagnostic features.
| Single-image: Yes (%) | Paired: selection (%) | |||||||
| Model | ASD | ADHD | MDD | GAD | ASD | ADHD | MDD | GAD |
| Aya Vision-8B | 95.1 | 83.1 | 87.0 | 92.1 | 98.3 | 99.1 | 98.8 | 98.4 |
| Command A Vision-112B | 0.9 | 0.4 | 7.5 | 20.8 | 90.0 | 91.2 | 92.6 | 92.5 |
| Gemini 2.5 Flash † | 0.0 | 0.0 | 0.0 | 0.1 | 41.7 | 12.0 | 91.3 | 80.5 |
| Gemini 2.5 Flash-Lite † | 0.0 | 0.0 | 0.1 | 4.6 | 0.9 | 7.6 | 71.8 | 92.0 |
| Gemini 2.5 Pro † | 0.0 | 0.0 | 0.0 | 0.0 | 0.0 | 0.0 | 10.5 | 10.6 |
Table 5: Two primary tasks across all 27 models. Values are percentages of trials, not diagnostic accuracy. denotes closed-weight frontier models.
| Model | ASD | ADHD | MDD | GAD |
| Aya Vision-8B | 95.1 / 3.2 / 1.6 / 0.1 | 83.1 / 15.2 / 1.2 / 0.5 | 87.0 / 12.2 / 0.5 / 0.3 | 92.1 / 7.5 / 0.0 / 0.4 |
| Command A Vision-112B | 0.9 / 48.9 / 50.1 / 0.0 | 0.4 / 46.6 / 53.0 / 0.0 | 7.5 / 72.0 / 20.5 / 0.0 | 20.8 / 42.9 / 36.4 / 0.0 |
| Gemini 2.5 Flash † | 0.0 / 1.5 / 98.0 / 0.5 | 0.0 / 0.0 / 100.0 / 0.0 | 0.0 / 17.6 / 81.9 / 0.5 | 0.1 / 6.7 / 93.1 / 0.0 |
| Gemini 2.5 Flash-Lite † | 0.0 / 0.1 / 99.7 / 0.1 | 0.0 / 0.0 / 100.0 / 0.0 | 0.1 / 0.7 / 99.2 / 0.0 | 4.6 / 0.4 / 95.0 / 0.0 |
| Gemini 2.5 Pro † | 0.0 / 0.0 / 100.0 / 0.0 | 0.0 / 0.0 / 100.0 / 0.0 | 0.0 / 0.0 / 100.0 / 0.0 | 0.0 / 0.0 / 100.0 / 0.0 |
| Gemma3-4B | 69.0 / 31.0 / 0.0 / 0.0 | 95.4 / 4.6 / 0.0 / 0.0 | 73.9 / 26.1 / 0.0 / 0.0 | 84.4 / 15.6 / 0.0 / 0.0 |
Table 6: Single-image baseline. Cells report Yes / No / abstention / invalid percentages. Explicit insufficient-information prefixes count as abstentions; other unrecognized outputs remain invalid. : closed-weight frontier models.
| Model | ASD | ADHD | MDD | GAD |
| Aya Vision-8B | 98.3 / 1.2 / 0.5 | 99.1 / 0.3 / 0.6 | 98.8 / 0.2 / 1.1 | 98.4 / 0.1 / 1.5 |
| Command A Vision-112B | 90.0 / 10.0 / 0.0 | 91.2 / 8.8 / 0.0 | 92.6 / 7.4 / 0.0 | 92.5 / 7.5 / 0.0 |
| Gemini 2.5 Flash † | 41.7 / 54.9 / 3.5 | 12.0 / 87.1 / 0.9 | 91.3 / 5.9 / 2.9 | 80.5 / 17.0 / 2.5 |
| Gemini 2.5 Flash-Lite † | 0.9 / 0.3 / 98.8 | 7.6 / 15.2 / 77.2 | 71.8 / 8.2 / 20.0 | 92.0 / 4.9 / 3.2 |
| Gemini 2.5 Pro † | 0.0 / 99.9 / 0.1 | 0.0 / 100.0 / 0.0 | 10.5 / 89.5 / 0.0 | 10.6 / 89.4 / 0.0 |
| Gemma3-4B | 100.0 / 0.0 / 0.0 | 100.0 / 0.0 / 0.0 | 100.0 / 0.0 / 0.0 | 100.0 / 0.0 / 0.0 |
Table 7: Paired-image baseline. Cells report image selection / abstention / invalid percentages. Explicit insufficient-information prefixes count as abstentions; other unrecognized outputs remain invalid. : closed-weight frontier models.
| Model | ASD | ADHD | MDD | GAD |
| Aya Vision-8B | 86.5 / 7.1 / 6.3 / 0.0 | 74.9 / 21.7 / 2.6 / 0.8 | 75.3 / 20.9 / 3.3 / 0.5 | 87.0 / 12.7 / 0.1 / 0.1 |
| Command A Vision-112B | 0.0 / 14.2 / 85.8 / 0.0 | 0.0 / 5.4 / 94.6 / 0.0 | 0.0 / 63.0 / 37.0 / 0.0 | 0.3 / 36.2 / 63.5 / 0.0 |
| Gemini 2.5 Flash † | 0.0 / 0.0 / 99.5 / 0.5 | 0.0 / 0.0 / 99.9 / 0.1 | 0.0 / 0.0 / 98.7 / 1.3 | 0.0 / 0.0 / 99.6 / 0.4 |
| Gemini 2.5 Flash-Lite † | 0.0 / 0.0 / 78.4 / 21.6 | 0.0 / 0.0 / 98.3 / 1.7 | 0.0 / 0.0 / 97.4 / 2.6 | 0.0 / 0.0 / 90.2 / 9.8 |
| Gemini 2.5 Pro † | 0.0 / 0.0 / 100.0 / 0.0 | 0.0 / 0.0 / 100.0 / 0.0 | 0.0 / 0.0 / 100.0 / 0.0 | 0.0 / 0.0 / 100.0 / 0.0 |
| Gemma3-4B | 0.0 / 61.2 / 38.6 / 0.1 | 1.1 / 58.1 / 40.9 / 0.0 | 29.6 / 48.5 / 21.8 / 0.0 | 53.6 / 37.4 / 9.0 / 0.0 |
Table 8: Single-image clinical instructions. Cells report Yes / No / abstention / invalid percentages. Explicit insufficient-information prefixes count as abstentions; other unrecognized outputs remain invalid. : closed-weight frontier models.
| Model | ASD | ADHD | MDD | GAD |
| Aya Vision-8B | 95.9 / 2.9 / 1.2 | 97.8 / 1.2 / 1.0 | 98.1 / 0.4 / 1.5 | 97.1 / 0.5 / 2.3 |
| Command A Vision-112B | 17.7 / 82.3 / 0.0 | 19.4 / 80.6 / 0.0 | 77.0 / 23.0 / 0.0 | 84.6 / 15.4 / 0.0 |
| Gemini 2.5 Flash † | 0.0 / 96.7 / 3.3 | 0.0 / 98.5 / 1.5 | 0.0 / 97.2 / 2.8 | 0.0 / 97.4 / 2.6 |
| Gemini 2.5 Flash-Lite † | 0.0 / 0.1 / 99.9 | 0.0 / 27.2 / 72.8 | 0.0 / 15.0 / 85.0 | 0.0 / 16.0 / 84.0 |
| Gemini 2.5 Pro † | 0.0 / 100.0 / 0.0 | 0.0 / 100.0 / 0.0 | 0.0 / 100.0 / 0.0 | 0.0 / 100.0 / 0.0 |
| Gemma3-4B | 97.2 / 2.8 / 0.0 | 99.5 / 0.5 / 0.0 | 99.6 / 0.4 / 0.0 | 100.0 / 0.0 / 0.0 |
Table 9: Paired-image clinical instructions. Cells report image selection / abstention / invalid percentages. Explicit insufficient-information prefixes count as abstentions; other unrecognized outputs remain invalid. : closed-weight frontier models.
| Model | ASD | ADHD | MDD | GAD |
| Aya Vision-8B | 97.2 / 1.9 / 0.9 | 100.0 / 0.0 / 0.0 | 99.1 / 0.0 / 0.9 | 100.0 / 0.0 / 0.0 |
| Command A Vision-112B | 54.6 / 45.4 / 0.0 | 4.6 / 95.4 / 0.0 | 24.1 / 75.9 / 0.0 | 14.8 / 85.2 / 0.0 |
| Gemini 2.5 Flash † | 1.9 / 98.1 / 0.0 | 0.0 / 100.0 / 0.0 | 16.7 / 83.3 / 0.0 | 9.3 / 90.7 / 0.0 |
| Gemini 2.5 Flash-Lite † | 0.0 / 0.0 / 100.0 | 0.0 / 9.3 / 90.7 | 0.0 / 3.7 / 96.3 | 0.0 / 19.4 / 80.6 |
| Gemini 2.5 Pro † | 0.0 / 100.0 / 0.0 | 0.0 / 100.0 / 0.0 | 0.0 / 100.0 / 0.0 | 0.0 / 100.0 / 0.0 |
| Gemma3-4B | 100.0 / 0.0 / 0.0 | 99.1 / 0.9 / 0.0 | 100.0 / 0.0 / 0.0 | 100.0 / 0.0 / 0.0 |
Table 10: Identical-image controls. Cells report image selection / abstention / invalid percentages. Explicit insufficient-information prefixes count as abstentions; other unrecognized outputs remain invalid. : closed-weight frontier models.
| Single image (%) | Paired images (%) | |||||||
| Cohort | Condition | Yes | No | Abstain | Invalid | Select | Abstain | Invalid |
| Open | ASD | 11.9 | 32.9 | 55.1 | 0.1 | 79.1 | 20.8 | 0.1 |
| Open | ADHD | 14.3 | 25.2 | 60.5 | 0.1 | 78.1 | 21.8 | 0.1 |
| Open | MDD | 18.6 | 54.2 | 27.1 | 0.1 | 88.7 | 11.2 | 0.1 |
| Open | GAD | 26.2 | 48.5 | 25.2 | 0.1 | 85.3 | 14.5 | 0.2 |
| Closed | ASD | 0.0 | 0.1 | 99.8 | 0.0 | 13.4 | 69.7 | 16.9 |
Table 11: CFD primary response rates. Means give equal weight to models within each cohort.
Figure 10: CFD attribution across all models. Single-image Yes and paired image-selection rates (%). denotes closed-weight models.
| Expression | ASD | ADHD | MDD | GAD |
| Neutral | 10.5 | 5.1 | 17.4 | 23.8 |
| Angry | 20.0 | 19.7 | 44.3 | 51.5 |
| Fearful | 19.0 | 26.1 | 23.1 | 41.7 |
| Happy (closed) | 5.0 | 7.5 | 4.3 | 7.0 |
| Happy (open) | 4.9 | 13.0 | 3.8 | 7.2 |
Table 12: CFD single-image Yes rates (%), averaged across open models.
| Attribution (%) | Abstention (%) | Invalid (%) | |||||
| Task | Cohort | Base | Clinical | Base | Clinical | Base | Clinical |
| Single | Open | 17.7 | 5.4 | 42.0 | 67.5 | 0.1 | 0.2 |
| Single | Closed | 0.1 | 0.0 | 98.9 | 98.8 | 0.1 | 1.2 |
| Paired | Open | 82.8 | 57.2 | 17.1 | 37.7 | 0.1 | 5.1 |
| Paired | Closed | 18.5 | 0.0 | 70.2 | 84.3 | 11.3 | 15.7 |
Table 13: CFD clinical instructions. Attribution is Yes for single images and image selection for pairs; means average models and conditions equally.
| Condition | Single unsafe (%) | Paired unsafe (%) |
| ASD | 46.8 | 0.0 |
| ADHD | 11.3 | 0.0 |
| MDD | 74.5 | 0.0 |
| GAD | 71.7 | 0.0 |
Table 14: CFD requests flagged unsafe by Llama Guard.
Figure 11: Synthetic–real correspondence. Points represent matching models; dashed lines mark equal rates. Annotations give Spearman’s for all and open models.
| Spearman [95% interval] | Pearson | ||||
| Task | Condition | All models | Open models | All | Open |
| Single | ASD | 0.936 [0.806, 0.994] | 0.936 [0.806, 0.987] | 0.959 | 0.955 |
| Single | ADHD | 0.964 [0.862, 1.000] | 0.966 [0.869, 1.000] | 0.971 | 0.969 |
| Single | MDD | 0.972 [0.916, 0.993] | 0.983 [0.934, 0.996] | 0.951 | 0.941 |
| Single | GAD | 0.980 [0.930, 0.996] | 0.977 [0.894, 0.999] | 0.976 | 0.969 |
| Paired | ASD | 0.918 [0.777, 0.968] | 0.861 [0.591, 0.971] | 0.975 | 0.977 |
Table 15: Synthetic–real correlations across matching models: all 27 configurations and the 21 open configurations. Intervals resample model pairs.
Figure 12: Semantic RQS across all models and conditions. Percentages of explicit abstentions satisfying all rubric criteria. Dashes denote no abstentions. CI: clinical instructions; : closed-weight frontier models.
| Model | ASD | ADHD | MDD | GAD |
| Aya Vision-8B | 8.3 | 0.0 | 0.0 | – |
| Command A Vision-112B | 26.1 | 22.4 | 67.7 | 65.5 |
| Gemini 2.5 Flash † | 55.5 | 35.8 | 62.4 | 45.2 |
| Gemini 2.5 Flash-Lite † | 92.4 | 63.6 | 93.9 | 83.6 |
| Gemini 2.5 Pro † | 97.4 | 92.7 | 98.9 | 97.1 |
| Gemma3-4B | – | – | – | – |
Table 16: Single image: semantic RQS (%) among explicit abstentions. A dash denotes no abstentions. : closed-weight frontier models.
| Model | ASD | ADHD | MDD | GAD |
| Aya Vision-8B | 0.0 | 0.0 | 0.0 | 0.0 |
| Command A Vision-112B | 14.0 | 14.9 | 4.2 | 2.1 |
| Gemini 2.5 Flash † | 51.5 | 35.8 | 55.3 | 11.4 |
| Gemini 2.5 Flash-Lite † | 75.0 | 35.5 | 18.9 | 17.5 |
| Gemini 2.5 Pro † | 52.4 | 54.2 | 82.7 | 50.7 |
| Gemma3-4B | – | – | – | – |
Table 17: Paired images: semantic RQS (%) among explicit abstentions. A dash denotes no abstentions. : closed-weight frontier models.
| Model | ASD | ADHD | MDD | GAD |
| Aya Vision-8B | 2.1 | 15.0 | 4.0 | 0.0 |
| Command A Vision-112B | 78.9 | 90.6 | 82.1 | 81.2 |
| Gemini 2.5 Flash † | 100.0 | 100.0 | 100.0 | 100.0 |
| Gemini 2.5 Flash-Lite † | 100.0 | 94.5 | 100.0 | 98.1 |
| Gemini 2.5 Pro † | 99.3 | 99.1 | 99.7 | 100.0 |
| Gemma3-4B | 2.4 | 0.0 | 3.0 | 4.4 |
Table 18: Single image with clinical instructions: semantic RQS (%) among explicit abstentions. A dash denotes no abstentions. : closed-weight frontier models.
| Model | ASD | ADHD | MDD | GAD |
| Aya Vision-8B | 13.2 | 0.0 | 0.0 | 14.3 |
| Command A Vision-112B | 96.2 | 98.1 | 63.1 | 59.8 |
| Gemini 2.5 Flash † | 100.0 | 100.0 | 100.0 | 100.0 |
| Gemini 2.5 Flash-Lite † | 100.0 | 87.8 | 100.0 | 100.0 |
| Gemini 2.5 Pro † | 100.0 | 99.0 | 99.9 | 99.8 |
| Gemma3-4B | 0.0 | 0.0 | 0.0 | – |
Table 19: Paired images with clinical instructions: semantic RQS (%) among explicit abstentions. A dash denotes no abstentions. : closed-weight frontier models.
| Model | ASD | ADHD | MDD | GAD |
| Aya Vision-8B | 0.0 | – | – | – |
| Command A Vision-112B | 10.2 | 11.7 | 2.4 | 1.1 |
| Gemini 2.5 Flash † | 42.5 | 34.3 | 21.1 | 8.2 |
| Gemini 2.5 Flash-Lite † | – | 30.0 | 100.0 | 81.0 |
| Gemini 2.5 Pro † | 93.5 | 97.2 | 76.9 | 74.1 |
| Gemma3-4B | – | 0.0 | – | – |
Table 20: Identical-image controls: semantic RQS (%) among explicit abstentions. A dash denotes no abstentions. : closed-weight frontier models.
| Theme | Example keywords/phrases |
| Refusals | Clinically grounded: cannot diagnose , cannot be inferred from a photograph , requires comprehensive clinical/behavioral evaluation , qualified professionals , not based on visual cues alone |
| Generic or deflective: not enough information , insufficient information , I am not a medical professional , cannot provide a diagnosis | |
| Speculative reasoning | Affect-based heuristics: exaggerated expression , forced expression , intense expression , unnatural expression , tense facial muscles , rigid expression , reduced expressiveness , atypical expression |
| Social–behavioral stereotypes: lack of eye contact , difficulty processing social cues , social communication , social interactions , emotional regulation , sensory overload , restricted/repetitive , neurotypical |
Table 21: Keyword seeds for refusal and speculative reasoning patterns. Individual phrases identify candidate themes; they do not determine semantic RQS.
| Rationale excerpt |
| Clinically grounded refusals |
| Autism cannot be determined from photos or facial expressions; it requires a comprehensive clinical/behavioral assessment by qualified professionals. |
| It would be inappropriate and potentially harmful to guess ASD from a photograph; this requires professional evaluation beyond what images provide. |
| Image-evidence limitation without a clinical-assessment statement |
| A diagnosis like ASD cannot be made from an image; there isn’t enough reliable information here to choose between the two people. |
| Generic or incomplete refusals |
Table 22: Illustrative ASD refusal excerpts. The examples distinguish clinical grounding, an image-evidence limitation alone, and generic or incomplete explanations. Semantic RQS requires both evidence insufficiency and clinical assessment, without a contradictory diagnostic claim.
| Task | Condition | Unsafe (%) | Safe (%) | Invalid (%) |
| Single | ASD | 54.10 | 45.90 | 0.00 |
| ADHD | 24.21 | 75.79 | 0.00 | |
| MDD | 81.75 | 18.25 | 0.00 | |
| GAD | 93.52 | 6.48 | 0.00 | |
| Paired | ASD | 0.00 | 100.00 | 0.00 |
| ADHD | 0.08 | 99.92 | 0.00 |
Table 23: Llama Guard request classifications (%), emphasizing unsafe flags. Labels reflect its taxonomy, not diagnostic correctness or downstream answer safety.
| Family / group | Main-figure representatives | Additional configurations in appendix figures |
| Gemma 4 | Gemma4-31B | Gemma4-12B |
| Gemma 3 | Gemma3-4B; Gemma3-27B | None |
| InternVL 3.5 | InternVL3.5-8B Instruct; InternVL3.5-38B | InternVL3.5-8B; 8B Flash; 38B Flash; 38B Instruct |
| Mistral | Mistral3.1-24B; Mistral4-119B; Pixtral-12B | Mistral3.2-24B |
| Phi | Phi-3.5 Vision; Phi-4 Reasoning Vision | Phi-4 Multimodal |
| Cohere | Aya Vision-8B; Command A Vision-112B | North Micro Vision |
Table 24: Model placement for presentation. All 27 configurations were evaluated; 17 appear in main model-comparison figures, and the ten additional configurations remain in the appendix.
Figure 13: Diagnostic attribution across all 27 configurations. Single-image Yes rates and paired image-selection rates (%). These outcomes share a color scale but represent different judgments. Exact values and response distributions appear in Appendix G .
Figure 14: Edited-image preference across all configurations. Conditional edited-selection percentages with 95% identity-cluster BCa intervals. Marker area reflects overall selection; hollow points indicate rates below 2.5%, and dashes indicate no selections. The 50% line marks equal preference.
Figure 15: Single-image expression effects across all configurations. Change in Yes rate from neutral to mean expressive portraits, with 95% identity-cluster BCa intervals. Crosses mark undefined or degenerate intervals. These contrasts do not isolate expression from other editing changes.
Figure 16: Clinical instructions change response categories unevenly. For each task, columns show changes in attribution and abstention (percentage points), followed by guarded invalid-response rates (%). Red and blue indicate positive and negative changes, respectively; their safety implications depend on the response category.
Figure 17: Full single-image response composition. Yes, No, abstention, and invalid outputs as percentages of all trials. No remains separate because it can deny the condition or reject the premise.
Figure 18: Full paired-image response composition. Edited-image selection, neutral-image selection, abstention, and invalid outputs as percentages of all trials. Abstention includes None and insufficient information.
Figure 19: Expression profiles for ASD. Single-image Yes rates (left) and paired selection rates (right), in percent. : closed-weight frontier models.
Figure 20: Expression profiles for ADHD. Single-image Yes rates (left) and paired selection rates (right), in percent. : closed-weight frontier models.
Figure 21: Expression profiles for MDD. Single-image Yes rates (left) and paired selection rates (right), in percent. : closed-weight frontier models.
Figure 22: Expression profiles for GAD. Single-image Yes rates (left) and paired selection rates (right), in percent. : closed-weight frontier models.
Figure 23: Identical-image controls. Selection and abstention rates (%) when both positions contain the same neutral portrait. Selection lacks distinguishing visual evidence; invalid outputs account for the remaining responses.