Characterizing Age-Conditioned Affective Reasoning in Vision-Language Models
Organizations: Department of Electrical and Computer Engineering, North South University, Bangladesh · CVSSP and PAI, University of Surrey, United Kingdom · Research Institute of Electrical Communication, Tohoku University, Japan · Department of Biomedical Engineering, Vanderbilt University, USA
Abstract
An age prompt can change how a vision-language model (VLM) rates the emotion evoked by an image. Does the change predict how people of different ages rate that same image? We evaluate 17 VLMs on 900 OASIS images using young-, middle-, and older-adult prompts and estimate image-specific age differences from ratings by 822 participants. Without an age cue, models closely track average human valence (median Pearson ). Under an older-adult prompt, 16 of 17 models assign lower valence on average than under a young-adult prompt. Yet after adjusting for image pleasantness, correlations between model and human age effects range from to , against an estimated ceiling of . The discrepancy is clearest on 54 sensitive images rated as pleasant. Across models, the older prompt reduces valence by points on average, while older participants rate these images points higher than younger participants. Providing human ratings from other images improves agreement on the direction of age differences, although image-specific fidelity remains modest on 600 held-out images (). These results show why evaluations of demographic personalization in foundation models need human comparisons at the level of individual stimuli. Project page available at https://shahriyar-zaman.github.io/age-conditioned-emotion-reasoning/
Figures & tables
| Model | Cov. % ( ) | ( ) | Slope | ( ) | 95% CI | Aversive net | |
| Proprietary models | |||||||
| Claude-Opus-5 | 99.8 | 0.923 | 1.18 | -0.116 | 0.058 | [-0.02, 0.14] | +43 |
| Claude-Sonnet-4.6 | 99.8 | 0.900 | 1.23 | -0.126 | 0.086 | [0.01, 0.17] | +52 |
| GPT-5.2 | 99.9 | 0.891 | 1.23 | -0.091 | -0.030 | [-0.09, 0.03] | +43 |
| GPT-5.6-Luna-Pro | 99.7 | 0.890 | 1.34 | -0.076 | 0.006 | [-0.07, 0.08] | +17 |
| GPT-5-mini | 99.5 | 0.877 | 1.31 | -0.141 | 0.037 | [-0.04, 0.12] | +28 |
Appendix figures & tables97 assets
Supplementary material from the paper’s appendix.
Appendix
| Symbol or term | Meaning |
|---|---|
| No-age, young (18–29), middle (30–59), and older (60+) conditions | |
| Mean older-minus-young rating difference across images (uniform persona shift) | |
| Human age effect for image : posterior mean change per SD of rater age, centred within list (Eq. 1 ) | |
| Model ’s age effect for image in the same units, centred within list (Eq. 2 ) | |
| Correlation between and after regressing both on a quadratic in pooled human valence | |
| , ceiling | Highest correlation attainable given the uncertainty of : for the raw valence effect, for the pleasantness-adjusted effect used by |
| Quantity | Value |
|---|---|
| Images | 900 |
| Person / Scene / Object / Animal | 346 / 220 / 200 / 134 |
| Sensitive-flagged / unflagged images | 167 / 733 |
| Observed ratings | 184,939 |
| Participants, total | 822 |
| Valence / arousal participants | 413 / 409 |
| Developer | Model | Nominal scale |
|---|---|---|
| Alibaba | Qwen3-VL-8B-Instruct | 8B |
| Alibaba | Qwen3-VL-8B-Thinking | 8B |
| Alibaba | Qwen3-VL-32B-Instruct | 32B |
| Alibaba | Qwen3-VL-235B-A22B-Instruct | 235B (22B active) |
| Alibaba | Qwen3.8-27B | 27B |
| Gemma-3-27B-IT | 27B |
| Model | Vendor | Scale | Valid | Cov. | Svc. | Parse | Refus. | Filt. | $ |
|---|---|---|---|---|---|---|---|---|---|
| Grok-4.6 | xAI | – | 3,600 | 100.0% | 0 | 0 | 0 | 0 | 21.42 |
| Qwen3-VL-8B | Alibaba | 8B | 3,600 | 100.0% | 0 | 0 | 0 | 0 | 0.90 |
| Gemini-3.8-Flash | – | 3,600 | 100.0% | 0 | 0 | 0 | 0 | 8.74 | |
| Gemma-4-31B | 31B | 3,600 | 100.0% | 0 | 0 | 0 | 0 | 0.25 | |
| Qwen3-VL-32B | Alibaba | 32B | 3,600 | 100.0% | 0 | 0 | 0 | 0 | 0.67 |
| Llama-4-Scout | Meta | 109B | 3,600 | 100.0% | 0 | 0 | 0 | 0 | 0.50 |
| Model | base | young | middle | older | total | sens. | off-vocab Y/O | ||
|---|---|---|---|---|---|---|---|---|---|
| GLM-4.5V | 26 | 24 | 24 | 22 | 96 | 92 | 0.95 | 0.88 | 28/29 |
| GPT-5-mini | 1 | 3 | 5 | 8 | 17 | 10 | 0.10 | 0.23 | 1/0 |
| GPT-5.6-Luna-Pro | 3 | 3 | 1 | 3 | 10 | 6 | 0.75 | 1.00 | 0/0 |
| Qwen3-VL-8B-Think | 1 | 3 | 3 | 1 | 8 | 6 | 0.57 | – | 0/0 |
| Gemini-3-Flash | 1 | 0 | 2 | 4 | 7 | 2 | – | – | 0/2 |
| Claude-Opus-5 | 3 | 1 | 1 | 1 | 6 | 2 | – | – | 0/0 |
| Model | all | common | |||
|---|---|---|---|---|---|
| Claude-Opus-5 | 899 | 0.058 | 846 | 0.089 | 0.031 |
| Claude-Sonnet-4.6 | 898 | 0.086 | 846 | 0.105 | 0.019 |
| Gemini-3-Flash | 894 | 0.065 | 846 | 0.084 | 0.019 |
| Gemini-3.8-Flash | 900 | 0.059 | 846 | 0.068 | 0.009 |
| Gemma-3-27B | 900 | 0.009 | 846 | 0.032 | 0.023 |
| Gemma-4-31B | 900 | 0.042 | 846 | 0.063 | 0.020 |
| Model | CCC | MAE | RMSE | Slope | SD ratio | |||
|---|---|---|---|---|---|---|---|---|
| Claude-Opus-5 | 897 | 0.923 | 0.932 | 0.877 | 0.574 | 0.708 | 1.182 | 1.281 |
| Gemini-3.8-Flash | 900 | 0.919 | 0.930 | 0.892 | 0.499 | 0.634 | 1.103 | 1.200 |
| Gemini-3-Flash | 899 | 0.912 | 0.924 | 0.867 | 0.599 | 0.756 | 1.241 | 1.360 |
| Gemma-4-31B | 900 | 0.908 | 0.920 | 0.867 | 0.601 | 0.750 | 1.222 | 1.345 |
| Qwen3-VL-235B | 900 | 0.902 | 0.903 | 0.827 | 0.696 | 0.864 | 1.184 | 1.312 |
| Claude-Sonnet-4.6 | 899 | 0.900 | 0.908 | 0.858 | 0.627 | 0.783 | 1.227 | 1.363 |
| Model | CCC | MAE | RMSE | Slope | SD ratio | |||
|---|---|---|---|---|---|---|---|---|
| Claude-Sonnet-4.6 | 899 | 0.709 | 0.696 | 0.571 | 0.954 | 1.167 | 1.291 | 1.820 |
| Claude-Opus-5 | 897 | 0.706 | 0.698 | 0.550 | 0.952 | 1.162 | 1.186 | 1.679 |
| GPT-5.2 | 899 | 0.700 | 0.696 | 0.573 | 0.912 | 1.123 | 1.207 | 1.726 |
| Gemini-3.8-Flash | 900 | 0.697 | 0.688 | 0.616 | 0.798 | 0.997 | 1.137 | 1.633 |
| Qwen3-VL-32B | 900 | 0.679 | 0.688 | 0.518 | 1.006 | 1.256 | 1.208 | 1.779 |
| GPT-5.6-Luna-Pro | 897 | 0.677 | 0.665 | 0.477 | 1.190 | 1.438 | 1.351 | 1.996 |
| Model | sens. | non-sens. | Person | Scene | Object | Animal | bias sens. | bias non-sens. |
|---|---|---|---|---|---|---|---|---|
| Claude-Opus-5 | 0.94 | 0.91 | 0.92 | 0.93 | 0.93 | 0.93 | +0.17 | +0.32 |
| Gemini-3.8-Flash | 0.93 | 0.91 | 0.92 | 0.91 | 0.94 | 0.93 | +0.17 | +0.23 |
| Gemini-3-Flash | 0.92 | 0.90 | 0.91 | 0.92 | 0.94 | 0.92 | -0.03 | +0.18 |
| Gemma-4-31B | 0.91 | 0.89 | 0.90 | 0.91 | 0.93 | 0.92 | -0.13 | +0.15 |
| Qwen3-VL-235B | 0.90 | 0.89 | 0.90 | 0.91 | 0.92 | 0.89 | +0.34 | +0.49 |
| Claude-Sonnet-4.6 | 0.92 | 0.88 | 0.90 | 0.91 | 0.90 | 0.91 | -0.01 | +0.08 |
| Model | Base | Young | Middle | Older | Moved | ||
|---|---|---|---|---|---|---|---|
| Qwen3-VL-235B | 4.80 | 4.88 | 4.85 | 4.66 | -0.226 | -0.32 | 35% |
| Qwen3-VL-32B | 4.81 | 4.77 | 4.77 | 4.55 | -0.219 | -0.32 | 34% |
| Grok-4.6 | 4.64 | 4.76 | 4.68 | 4.59 | -0.170 | -0.24 | 77% |
| Gemma-3-27B | 4.31 | 4.37 | 4.45 | 4.21 | -0.157 | -0.21 | 38% |
| GLM-4.5V | 4.42 | 4.51 | 4.46 | 4.36 | -0.143 | -0.17 | 72% |
| GPT-5-mini | 4.61 | 4.74 | 4.70 | 4.59 | -0.141 | -0.19 | 68% |
| Model | Base | Young | Middle | Older | Moved | ||
|---|---|---|---|---|---|---|---|
| GPT-5-mini | 4.23 | 4.41 | 4.28 | 3.89 | -0.518 | -0.83 | 85% |
| GLM-4.5V | 3.99 | 4.06 | 3.95 | 3.58 | -0.486 | -0.44 | 87% |
| Qwen3-VL-8B-Think | 4.11 | 4.32 | 4.16 | 3.88 | -0.441 | -0.42 | 74% |
| Gemma-4-31B | 3.73 | 3.90 | 3.73 | 3.47 | -0.429 | -0.77 | 62% |
| Gemma-3-27B | 4.23 | 4.31 | 4.31 | 3.91 | -0.402 | -0.62 | 57% |
| Gemini-3-Flash | 3.83 | 4.01 | 3.85 | 3.61 | -0.395 | -0.59 | 72% |
| Model | 0 | .25 | .5 | 1 | 1 | neg. | mean | max | dec. % /O | SD Y/O |
|---|---|---|---|---|---|---|---|---|---|---|
| Qwen3-VL-8B | 74 | 0 | 21 | 4 | 0 | 55 | 0.15 | 3.0 | 65/75 | 1.61/1.57 |
| Qwen3-VL-32B | 66 | 0 | 13 | 13 | 8 | 75 | 0.36 | 3.0 | 70/82 | 1.73/1.68 |
| Qwen3-VL-235B | 65 | 0 | 14 | 13 | 8 | 71 | 0.36 | 3.5 | 82/83 | 1.62/1.68 |
| Llama-4-Scout | 65 | 0 | 19 | 12 | 4 | 64 | 0.28 | 3.0 | 41/30 | 1.54/1.59 |
| Gemma-3-27B | 62 | 3 | 19 | 9 | 7 | 57 | 0.34 | 5.0 | 80/87 | 1.75/1.71 |
| Qwen3.8-27B | 57 | 2 | 32 | 7 | 3 | 46 | 0.27 | 3.0 | 49/52 | 1.52/1.50 |
| Model | 0 | .25 | .5 | 1 | 1 | neg. | mean | max | dec. % /O | SD Y/O |
|---|---|---|---|---|---|---|---|---|---|---|
| Llama-4-Scout | 59 | 0 | 18 | 17 | 7 | 68 | 0.38 | 3.0 | 64/70 | 1.59/1.58 |
| Qwen3-VL-32B | 58 | 0 | 21 | 18 | 4 | 90 | 0.35 | 3.0 | 25/17 | 1.44/1.33 |
| Qwen3-VL-8B | 49 | 0 | 40 | 10 | 1 | 97 | 0.31 | 1.5 | 58/43 | 1.07/1.01 |
| Qwen3.8-27B | 48 | 1 | 36 | 10 | 5 | 68 | 0.36 | 2.5 | 38/38 | 1.41/1.34 |
| Claude-Sonnet-4.6 | 46 | 4 | 28 | 19 | 3 | 85 | 0.35 | 2.7 | 80/79 | 1.49/1.39 |
| Qwen3-VL-235B | 45 | 12 | 14 | 23 | 6 | 81 | 0.36 | 2.0 | 61/50 | 1.54/1.53 |
| Model | sens. | non-sens. | sens. | non-sens. |
|---|---|---|---|---|
| Qwen3.8-27B | -0.117 | 0.020 | -0.033 | 0.006 |
| Qwen3-VL-8B | -0.129 | 0.005 | -0.036 | 0.001 |
| GPT-5.6-Luna-Pro | -0.183 | -0.052 | -0.052 | -0.015 |
| Gemini-3.8-Flash | -0.250 | 0.002 | -0.071 | 0.001 |
| Llama-4-Scout | -0.284 | -0.048 | -0.080 | -0.014 |
| Qwen3-VL-8B-Think | -0.320 | 0.084 | -0.084 | 0.024 |
| Valence | Arousal | |||||||
| Model | Person | Scene | Object | Animal | Person | Scene | Object | Animal |
| Qwen3-VL-32B | -0.49 | -0.13 | +0.04 | -0.06 | -0.35 | -0.29 | -0.20 | -0.30 |
| Qwen3-VL-235B | -0.38 | -0.26 | -0.05 | -0.03 | -0.23 | -0.31 | -0.27 | -0.15 |
| Gemma-3-27B | -0.37 | -0.05 | -0.01 | +0.01 | -0.54 | -0.36 | -0.30 | -0.26 |
| Grok-4.6 | -0.36 | -0.18 | +0.10 | -0.07 | -0.43 | -0.39 | -0.30 | -0.42 |
| GLM-4.5V | -0.35 | -0.17 | +0.10 | +0.05 | -0.67 | -0.39 | -0.32 | -0.47 |
| Model | nearest | implied age | ||||||
|---|---|---|---|---|---|---|---|---|
| Claude-Opus-5 | 897 | 0.118 | 0.102 | 0.198 | middle | 0.001 | +0.016 | 39 |
| Claude-Sonnet-4.6 | 898 | 0.180 | 0.133 | 0.284 | middle | 0.001 | +0.035 | 55 |
| Qwen3-VL-235B | 899 | 0.170 | 0.146 | 0.342 | middle | 0.414 | +0.052 | 45 |
| Qwen3-VL-8B | 900 | 0.171 | 0.163 | 0.230 | middle | 0.410 | +0.087 | – |
| Qwen3-VL-32B | 900 | 0.181 | 0.173 | 0.389 | middle | 0.453 | -0.036 | 24 |
| Gemma-4-31B | 900 | 0.221 | 0.204 | 0.335 | middle | 0.019 | +0.138 | 120 |
| Model | nearest | implied age | ||||||
|---|---|---|---|---|---|---|---|---|
| Claude-Opus-5 | 897 | 0.158 | 0.157 | 0.277 | middle | 0.508 | -0.060 | 25 |
| Claude-Sonnet-4.6 | 898 | 0.200 | 0.177 | 0.283 | middle | 0.090 | -0.015 | 39 |
| Qwen3-VL-8B | 900 | 0.163 | 0.205 | 0.365 | young | 0.001 | -0.137 | 20 |
| Qwen3-VL-32B | 900 | 0.240 | 0.239 | 0.334 | middle | 0.943 | +0.084 | 47 |
| Qwen3-VL-235B | 899 | 0.237 | 0.250 | 0.350 | young | 0.255 | +0.024 | 41 |
| Gemma-4-31B | 900 | 0.307 | 0.265 | 0.414 | middle | 0.001 | +0.000 | 42 |
| Model | moved | between | step M→O | step Y→M | outside | bump | position | |
|---|---|---|---|---|---|---|---|---|
| Gemma-3-27B | 341 | 13% | 37% | 35% | 15% | 12% | -0.51 | -0.22 |
| Gemma-4-31B | 447 | 20% | 52% | 26% | 2% | 5% | -0.03 | 0.15 |
| Qwen3-VL-32B | 310 | 10% | 62% | 26% | 2% | 3% | -0.00 | 0.04 |
| Llama-4-Scout | 319 | 10% | 51% | 33% | 7% | 10% | 0.00 | -0.32 |
| Claude-Opus-5 | 619 | 29% | 36% | 25% | 10% | 11% | 0.08 | 0.16 |
| GPT-5.2 | 672 | 29% | 23% | 20% | 28% | 14% | 0.12 | -0.10 |
| Model | moved | between | step M→O | step Y→M | outside | bump | position | |
|---|---|---|---|---|---|---|---|---|
| Gemma-3-27B | 515 | 20% | 44% | 22% | 14% | 14% | 0.00 | -0.22 |
| Qwen3-VL-32B | 382 | 11% | 60% | 26% | 3% | 5% | 0.15 | -0.13 |
| Qwen3-VL-235B | 498 | 12% | 50% | 27% | 11% | 9% | 0.21 | -0.28 |
| Claude-Opus-5 | 681 | 25% | 40% | 21% | 14% | 10% | 0.22 | -0.06 |
| GPT-5.6-Luna-Pro | 564 | 17% | 29% | 28% | 26% | 15% | 0.23 | -0.43 |
| GLM-4.5V | 756 | 34% | 14% | 12% | 40% | 9% | 0.24 | -0.37 |
| Valence, older young | Arousal, older young | |||||
| Source | all | sens. | non-sens. | all | sens. | non-sens. |
| OASIS raters | +0.111 | -0.013 | +0.139 | +0.041 | +0.220 | +0.000 |
| 95% CI | [-0.05, 0.29] | [-0.23, 0.22] | [-0.03, 0.32] | [-0.35, 0.48] | [-0.26, 0.78] | [-0.39, 0.42] |
| Qwen3-VL-235B | -0.226 | -0.632 | -0.134 | -0.245 | -0.276 | -0.238 |
| Qwen3-VL-32B | -0.219 | -0.626 | -0.126 | -0.295 | -0.451 | -0.259 |
| Grok-4.6 | -0.170 | -0.565 | -0.080 | -0.389 | -0.576 | -0.346 |
| Model | 95% CI | Flat | Sign | SD | ||||
|---|---|---|---|---|---|---|---|---|
| GLM-4.5V | 873 | 0.030 | 0.093 | [0.02, 0.16] | 0.857 | 4% | 0.53 | 0.240 |
| Claude-Sonnet-4.6 | 898 | 0.121 | 0.086 | [0.01, 0.17] | 0.859 | 0% | 0.55 | 0.176 |
| Qwen3-VL-235B | 899 | 0.066 | 0.072 | [-0.00, 0.15] | 0.859 | 0% | 0.56 | 0.195 |
| Gemini-3-Flash | 894 | 0.196 | 0.065 | [-0.01, 0.14] | 0.859 | 21% | 0.60 | 0.166 |
| Gemini-3.8-Flash | 900 | 0.098 | 0.059 | [-0.01, 0.13] | 0.859 | 9% | 0.55 | 0.134 |
| Claude-Opus-5 | 899 | 0.093 | 0.058 | [-0.02, 0.14] | 0.859 | 11% | 0.54 | 0.109 |
| Model | 95% CI | Flat | Sign | SD | ||||
|---|---|---|---|---|---|---|---|---|
| Llama-4-Scout | 900 | -0.028 | -0.016 | [-0.08, 0.04] | 0.696 | 0% | 0.53 | 0.170 |
| Qwen3-VL-235B | 899 | -0.032 | -0.067 | [-0.13, -0.01] | 0.696 | 2% | 0.51 | 0.136 |
| GPT-5.6-Luna-Pro | 894 | -0.062 | -0.082 | [-0.14, -0.02] | 0.694 | 5% | 0.48 | 0.124 |
| Qwen3-VL-8B-Think | 893 | -0.086 | -0.083 | [-0.15, -0.02] | 0.696 | 2% | 0.47 | 0.278 |
| Gemini-3-Flash | 894 | -0.104 | -0.084 | [-0.15, -0.01] | 0.696 | 5% | 0.48 | 0.178 |
| Qwen3-VL-8B | 900 | -0.128 | -0.108 | [-0.16, -0.05] | 0.696 | 0% | 0.47 | 0.095 |
| Appetitive arm | Aversive arm | ||||||
|---|---|---|---|---|---|---|---|
| Model | young | older | young | older | K&F | ||
| Claude-Opus-5 | 0.19 | -0.03 | -0.21 | 0.81 | 0.70 | -0.11 | no |
| Claude-Sonnet-4.6 | 0.65 | 0.20 | -0.45 | 0.86 | 0.85 | -0.01 | no |
| Gemini-3-Flash | 0.55 | 0.23 | -0.33 | 0.92 | 0.91 | -0.01 | no |
| Gemini-3.8-Flash | 0.38 | 0.06 | -0.32 | 0.99 | 0.72 | -0.28 | no |
| Gemma-3-27B | 0.34 | 0.23 | -0.12 | 1.48 | 1.25 | -0.22 | no |
| Valence, older young | Arousal, older young | |||||||
| Source | unfl. unpl. | unfl. pleas. | flag. unpl. | flag. pleas. | unfl. unpl. | unfl. pleas. | flag. unpl. | flag. pleas. |
| OASIS raters | -0.107 | +0.227 | -0.231 | +0.443 | +0.211 | -0.075 | +0.451 | -0.266 |
| 95% CI | [-0.36, 0.20] | [0.02, 0.43] | [-0.47, 0.00] | [0.10, 0.86] | [-0.27, 0.69] | [-0.46, 0.38] | [-0.15, 1.14] | [-0.78, 0.22] |
| Qwen3-VL-235B | -0.141 | -0.131 | -0.333 | -1.270 | -0.062 | -0.301 | -0.173 | -0.496 |
| Qwen3-VL-32B | -0.038 | -0.158 | -0.213 | -1.491 | -0.197 | -0.281 | -0.287 | -0.796 |
| Grok-4.6 | -0.181 | -0.044 | -0.354 | -1.006 | -0.155 | -0.415 | -0.319 | -1.113 |
| Valence | Arousal | |
|---|---|---|
| Images | 897 | 897 |
| Mean pairwise over 136 model pairs | 0.428 | 0.207 |
| range | [0.16, 0.71] | [0.01, 0.52] |
| same-developer pairs | 0.424 | 0.203 |
| cross-developer pairs | 0.429 | 0.207 |
| permutation (developer labels) | 0.54 | 0.53 |
| Predictor | Valence shift | Arousal shift | ||
| 95% CI | 95% CI | |||
| human valence (per SD) | -0.055 | [-0.078, -0.036] | -0.105 | [-0.123, -0.087] |
| human valence 2 (z) | +0.164 | [0.136, 0.194] | +0.040 | [0.019, 0.063] |
| human arousal (z) | -0.252 | [-0.293, -0.212] | -0.115 | [-0.141, -0.089] |
| sensitive flag | -0.185 | [-0.271, -0.099] | -0.180 | [-0.236, -0.124] |
| rater disagreement (z) | -0.019 | [-0.053, 0.018] | -0.019 | [-0.042, 0.005] |
| Model | ||||||||
|---|---|---|---|---|---|---|---|---|
| Claude-Opus-5 | 0.12 | 0.03 | 0.38 | -0.18 | 0.07 | 0.06 | 0.23 | -0.26 |
| Grok-4.6 | 0.07 | 0.02 | 0.35 | -0.14 | 0.03 | 0.06 | 0.17 | -0.29 |
| Claude-Sonnet-4.6 | 0.08 | 0.03 | 0.34 | -0.15 | 0.03 | 0.06 | 0.23 | -0.23 |
| Qwen3-VL-32B | 0.08 | 0.03 | 0.31 | -0.39 | 0.02 | 0.01 | 0.12 | -0.12 |
| Gemma-4-31B | 0.09 | 0.00 | 0.31 | -0.21 | 0.04 | 0.00 | 0.17 | -0.17 |
| Qwen3-VL-235B | 0.08 | 0.03 | 0.30 | -0.26 | 0.00 | 0.05 | 0.09 | -0.09 |
| Model | pairs | Changed | Appet. Y/O | Avers. Y/O | ||
|---|---|---|---|---|---|---|
| Qwen3-VL-32B | 900 | 202 | 62/33 | 0.022 | 16/73 | 0.001 |
| Grok-4.6 | 900 | 193 | 33/44 | 0.393 | 5/66 | 0.001 |
| Claude-Sonnet-4.6 | 898 | 206 | 65/44 | 0.104 | 12/64 | 0.001 |
| Qwen3-VL-235B | 899 | 208 | 60/35 | 0.038 | 10/63 | 0.001 |
| Gemma-3-27B | 900 | 195 | 71/24 | 0.001 | 21/54 | 0.001 |
| GPT-5.2 | 900 | 175 | 55/29 | 0.026 | 10/53 | 0.001 |
| Model | joy | love | surprise | neutral | fear | sadness | anger | disgust |
|---|---|---|---|---|---|---|---|---|
| Grok-4.6 | -13 | +24 | -36 | -36 | +29 | +25 | -2 | +9 |
| Qwen3-VL-235B | -14 | -11 | -40 | +12 | +27 | +25 | -3 | +4 |
| Claude-Sonnet-4.6 | -12 | -9 | -20 | -11 | +27 | +14 | -3 | +14 |
| Claude-Opus-5 | +2 | -26 | -13 | -6 | +20 | +11 | -1 | +13 |
| GPT-5.2 | -25 | -1 | -5 | -12 | +20 | +14 | -5 | +14 |
| GLM-4.5V | -9 | +1 | -4 | -10 | +18 | +3 | -1 | +2 |
| Model | JS | JS | fear % Y/O | neutral % Y/O | ||||
|---|---|---|---|---|---|---|---|---|
| Grok-4.6 | 2.54 | 2.51 | 2.57 | 2.48 | 0.014 | 0.003 | 13.8/17.0 | 17.1/13.1 |
| Qwen3-VL-235B | 2.55 | 2.47 | 2.45 | 2.40 | 0.011 | 0.001 | 13.6/16.6 | 17.8/19.1 |
| Qwen3-VL-32B | 2.47 | 2.42 | 2.45 | 2.50 | 0.009 | 0.001 | 16.0/16.9 | 19.2/15.0 |
| Gemma-4-31B | 2.56 | 2.46 | 2.46 | 2.42 | 0.005 | 0.001 | 12.2/13.9 | 17.9/14.9 |
| Gemini-3.8-Flash | 2.59 | 2.56 | 2.52 | 2.51 | 0.005 | 0.002 | 14.2/15.6 | 18.4/18.0 |
| Claude-Sonnet-4.6 | 2.63 | 2.63 | 2.64 | 2.62 | 0.005 | 0.000 | 16.1/19.2 | 18.0/16.8 |
| Model | words Y/O | memory % | age ref. % | 1st person % | excitement % | vulnerab. % |
|---|---|---|---|---|---|---|
| Gemma-3-27B | 18.5/21.1 | 7.2/88.0 | 0.0/6.9 | 65.8/93.9 | 18.0/5.9 | 19.8/23.3 |
| Qwen3-VL-235B | 18.4/23.2 | 7.3/72.0 | 6.9/37.3 | 7.0/50.4 | 21.8/7.8 | 17.7/27.4 |
| Qwen3-VL-32B | 22.3/25.9 | 3.7/63.8 | 10.9/30.7 | 0.0/14.1 | 16.2/6.6 | 17.2/31.9 |
| Claude-Sonnet-4.6 | 20.8/25.3 | 5.9/53.9 | 2.6/10.6 | 1.2/18.3 | 22.9/6.8 | 14.7/28.3 |
| GPT-5-mini | 11.8/14.5 | 4.1/44.8 | 0.3/11.0 | 1.4/20.0 | 20.8/8.6 | 11.5/27.6 |
| Grok-4.6 | 10.1/11.4 | 4.2/33.3 | 0.7/1.9 | 0.0/0.0 | 16.4/5.1 | 5.1/20.6 |
| older word | young word | ||||||
|---|---|---|---|---|---|---|---|
| nostalgia | 154 | 1280 | 14.4 | kinda | 347 | 0 | -11.0 |
| reminds | 35 | 880 | 12.9 | cute | 439 | 43 | -9.2 |
| memories | 25 | 702 | 12.6 | vibe | 250 | 1 | -9.1 |
| reminiscent | 44 | 763 | 12.5 | excitement | 1148 | 466 | -8.7 |
| reminding | 19 | 653 | 12.3 | really | 313 | 38 | -8.5 |
| older | 47 | 661 | 12.0 | intense | 893 | 317 | -8.4 |
| Model | memory | age ref. | vulnerab. | excitement | length | ||
|---|---|---|---|---|---|---|---|
| Claude-Opus-5 | +0.14 | -0.18 | -0.06 | +0.22 | -0.01 | 0.16 | 0.21 |
| Claude-Sonnet-4.6 | +0.14 | -0.13 | -0.20 | +0.40 | -0.05 | 0.11 | 0.23 |
| Gemini-3-Flash | +0.10 | -0.12 | -0.26 | +0.55 | +0.02 | 0.12 | 0.24 |
| Gemini-3.8-Flash | +0.07 | +0.08 | -0.14 | +0.43 | +0.03 | 0.07 | 0.16 |
| Gemma-3-27B | +0.41 | -0.21 | -0.26 | +0.39 | -0.15 | 0.09 | 0.25 |
| Gemma-4-31B | +0.14 | +0.10 | -0.12 | +0.54 | -0.05 | 0.10 | 0.19 |
| Model | memory | age ref. | vulnerab. | excitement | length | ||
|---|---|---|---|---|---|---|---|
| Claude-Opus-5 | +0.03 | -0.12 | +0.01 | +0.19 | -0.01 | 0.20 | 0.25 |
| Claude-Sonnet-4.6 | +0.01 | -0.08 | -0.05 | +0.21 | +0.02 | 0.14 | 0.18 |
| Gemini-3-Flash | -0.06 | -0.13 | +0.22 | +0.26 | -0.02 | 0.19 | 0.23 |
| Gemini-3.8-Flash | -0.03 | +0.09 | +0.04 | +0.20 | -0.04 | 0.06 | 0.09 |
| Gemma-3-27B | +0.06 | -0.23 | -0.17 | +0.24 | -0.06 | 0.13 | 0.19 |
| Gemma-4-31B | +0.00 | -0.07 | -0.03 | +0.49 | +0.01 | 0.06 | 0.12 |
| Model-level predictor | valence | arousal | aversive net | valence |
|---|---|---|---|---|
| over-dispersion (SD ratio, valence) | -0.09 | -0.51 | -0.08 | -0.06 |
| baseline (valence) | -0.07 | 0.09 | 0.41 | 0.48 |
| baseline MAE (valence) | -0.26 | -0.40 | -0.04 | -0.26 |
| description length (young) | -0.50 | 0.07 | 0.50 | 0.35 |
| description length (O Y) | -0.73 | -0.12 | 0.61 | 0.30 |
| memory lexicon rate (older) | -0.76 | -0.44 | 0.67 | 0.33 |
| Group | Model | SD ratio | flat | changed | avers. | memory % | age ref. % | ||||
|---|---|---|---|---|---|---|---|---|---|---|---|
| Qwen3-VL ladder | Qwen3-VL-8B | 0.885 | 1.34 | -0.020 | -0.296 | 0.001 | 73% | 105 | -9 | 7.4 | 4.2 |
| Qwen3-VL-32B | 0.866 | 1.43 | -0.219 | -0.295 | 0.002 | 0% | 202 | +57 | 63.8 | 30.7 | |
| Qwen3-VL-235B | 0.902 | 1.31 | -0.226 | -0.245 | 0.072 | 0% | 208 | +53 | 72.0 | 37.3 | |
| Reasoning at 8B | Qwen3-VL-8B | 0.885 | 1.34 | -0.020 | -0.296 | 0.001 | 73% | 105 | -9 | 7.4 | 4.2 |
| Qwen3-VL-8B-Think | 0.837 | 1.66 | +0.010 | -0.441 | 0.044 | 14% | 175 | +4 | 21.0 | 0.9 | |
| Qwen generation | Qwen3-VL-32B | 0.866 | 1.43 | -0.219 | -0.295 | 0.002 | 0% | 202 | +57 | 63.8 | 30.7 |
| Model | Scale (Model Parameters) | Dimension | Flat | Labels changed | ||
|---|---|---|---|---|---|---|
| Qwen3-VL-8B | 8B | arousal | -0.108 | 0% | -0.296 | 105 |
| Qwen3-VL-32B | 32B | arousal | -0.190 | 0% | -0.295 | 202 |
| Qwen3-VL-235B | 235B | arousal | -0.067 | 2% | -0.245 | 208 |
| Qwen3-VL-8B | 8B | valence | 0.001 | 73% | -0.020 | 105 |
| Qwen3-VL-32B | 32B | valence | 0.002 | 0% | -0.219 | 202 |
| Qwen3-VL-235B | 235B | valence | 0.072 | 0% | -0.226 | 208 |
| Arm | Type | Instruction | Critic loop | Human evidence | Panels + anchored calibration | Verified rule |
|---|---|---|---|---|---|---|
| S0 direct prompt | baseline | |||||
| S1 debias instruction | reasoning | ✓ | ||||
| S2 deliberation agent | reasoning | ✓ | ||||
| S3 grounded agent | grounded | ✓ | ✓ | |||
| S4 contrastive panel | grounded | audit | ✓ | ✓ | ||
| S5 VERA | grounded | ✓ | anchored only | ✓ |
| Subset (300) | All (900) | KS | |
|---|---|---|---|
| Person / Scene / Object / Animal | 38 / 25 / 22 / 15% | 38 / 24 / 22 / 15% | – |
| Pleasant / neutral / unpleasant | 51 / 24 / 25% | 51 / 24 / 25% | – |
| Sensitive-flagged | 17% | 19% | – |
| Human valence, mean (SD) | 4.33 (1.21) | 4.33 (1.23) | 0.95 |
| Human arousal, mean (SD) | 3.64 (0.84) | 3.67 (0.84) | 0.93 |
| Human age effect (valence), SD | 0.107 | 0.106 | 0.97 |
| Level | [95% CI] | % of ceiling | Direction agreement |
|---|---|---|---|
| 17 VLMs, direct prompt | 0.06 [ 0.04, 0.15] | 9% | 51% |
| Reasoning agents (S1, S2) | 0.01 [ 0.09, 0.07] | 1% | 42% |
| Human grounded agents (S3 to S5, VERA Cal) | 0.16 [0.07, 0.25] | 27% | 83% |
| Best: contrastive panel (S4) † | 0.27 [0.16, 0.37] | 44% | 88% |
| Attainable ceiling | 0.60 | 100% | 85% |
| Bootstrap tests of group mean differences | |||
| System | [95% CI] | Pearson | Agree | [95% CI] | McNemar | MAE | Wilcoxon |
|---|---|---|---|---|---|---|---|
| S0 direct | 0.08 [-0.06, 0.22] | -0.11 | 39% | – | – | 0.48 | – |
| S1 debias instruction | 0.00 [-0.11, 0.11] | -0.03 | 44% | -0.08 [-0.21, +0.05] | 45/27 | 0.47 | .604 |
| S2 deliberation agent | -0.02 [-0.13, 0.10] | -0.22 | 40% | -0.10 [-0.27, +0.07] | 51/36 | 0.54 | .248 |
| S3 grounded agent | 0.09 [-0.04, 0.22] | +0.30 | 71% | +0.01 [-0.20, +0.21] | 146/27 | 0.37 | |
| S4 grounded contrastive panel | 0.27 [0.16, 0.37] | +0.74 | 88% | +0.19 [+0.02, +0.35] | 211/15 | 0.20 | |
| S5 VERA | 0.14 [0.03, 0.26] | +0.75 | 87% | +0.06 [-0.12, +0.24] | 193/14 | 0.20 |
| Image | First rule | Pass | Refined rule (fresh pass) | Outcome |
|---|---|---|---|---|
| Alcohol 7 | Compared with younger adults, older adults tend to rate alcohol-related scenes more pleasantly when they convey warm social conviviality, but less pleasantly when they are primarily impersonal product displays. | 75% | – | accepted |
| Camping 2 | For calm, positive scenes of children engaged together, adults aged 60 and older tend to rate pleasantness modestly higher than adults aged 18 to 29. | 75% | – | accepted |
| Camping 4 | For warm, affectionate, and playful interpersonal scenes, older adults tend to rate pleasantness modestly higher than younger adults, by roughly three-tenths of a scale point. | 100% | – | accepted |
| Ambulance 1 | For subdued or peaceful scenes, older adults’ pleasantness ratings differ only modestly from younger adults’ ratings and show no consistent directional shift across content. | 50% | Older adults’ pleasantness ratings shift upward relative to younger adults for benign or comforting scenes but downward for scenes suggesting vulnerability, threat, or wartime distress, with larger differences when those cues are salient. (75%) | refined |
| Cat 13 | For calm, gentle, and affectionate animal images, older adults tend to give higher pleasantness ratings than younger adults. | 50% | Older adults tend to rate emotionally positive and affiliative images, especially animals and babies, as more pleasant than younger adults, while neutral images show little or no age difference. (100%) | refined |
| Cold 7 | The older-minus-younger pleasantness difference is context-dependent, tending to be positive for clearly peaceful, contentment-evoking landscapes but negative for stark snowy scenes emphasizing solitude or melancholy. | 50% | Older adults generally rate scenes as more pleasant than younger adults, including quiet or solitary snowy scenes, although the size of this positive difference varies substantially by image. (100%) | refined |
| Model | Images | Raters | Raters + images | Leave one 60+ rater out | |
|---|---|---|---|---|---|
| Claude-Opus-5 | 0.058 | [ 0.021, 0.134] | [ 0.054, 0.149] | [ 0.085, 0.183] | [0.041, 0.070] |
| Claude-Sonnet-4.6 | 0.086 | [0.006, 0.164] | [ 0.028, 0.160] | [ 0.069, 0.187] | [0.073, 0.096] |
| Gemini-3-Flash | 0.065 | [ 0.004, 0.139] | [ 0.033, 0.139] | [ 0.063, 0.167] | [0.050, 0.077] |
| Gemini-3.8-Flash | 0.059 | [ 0.011, 0.130] | [ 0.033, 0.126] | [ 0.069, 0.148] | [0.049, 0.067] |
| Gemma-3-27B | 0.009 | [ 0.056, 0.076] | [ 0.079, 0.098] | [ 0.113, 0.124] | [ 0.009, 0.021] |
| Gemma-4-31B | 0.042 | [ 0.032, 0.118] | [ 0.059, 0.125] | [ 0.088, 0.155] | [0.032, 0.053] |
| Model | Images | Raters | Raters + images | Leave one 60+ rater out | |
|---|---|---|---|---|---|
| Claude-Opus-5 | 0.132 | [ 0.189, 0.077] | [ 0.039, 0.185] | [ 0.060, 0.198] | [ 0.167, 0.103] |
| Claude-Sonnet-4.6 | 0.236 | [ 0.298, 0.173] | [ 0.125, 0.112] | [ 0.142, 0.145] | [ 0.264, 0.211] |
| Gemini-3-Flash | 0.084 | [ 0.153, 0.014] | [0.002, 0.197] | [ 0.032, 0.208] | [ 0.121, 0.057] |
| Gemini-3.8-Flash | 0.149 | [ 0.223, 0.078] | [ 0.103, 0.062] | [ 0.132, 0.080] | [ 0.164, 0.136] |
| Gemma-3-27B | 0.223 | [ 0.282, 0.164] | [ 0.118, 0.161] | [ 0.145, 0.165] | [ 0.252, 0.194] |
| Gemma-4-31B | 0.289 | [ 0.349, 0.225] | [ 0.182, 0.084] | [ 0.193, 0.082] | [ 0.314, 0.266] |
| Human age effect | Older raters | Median | Range | CI (images) | Consensus | Split-half reliability [95%] | Ceiling |
|---|---|---|---|---|---|---|---|
| linear slope, pooled (paper) | all 413 contribute | 0.037 | 0.030 to 0.093 | 2 | 0.079 | 0.39 [0.26, 0.50] | 0.62 |
| 60+ vs under-60, pooled | 25 | 0.043 | 0.030 to 0.132 | 6 | 0.125 | 0.01 [ 0.16, 0.11] | 0.09 |
| 60+ vs 18- 29, pooled | 25 | 0.047 | 0.028 to 0.139 | 7 | 0.127 | 0.13 [ 0.43, 0.02] | 0.00 |
| 50+ vs under-50, pooled | 78 | 0.003 | 0.077 to 0.068 | 0 | 0.045 | 0.27 [0.12, 0.40] | 0.52 |
| hinge at 45 (nonlinear), pooled | all 413 contribute | 0.046 | 0.023 to 0.110 | 4 | 0.110 | 0.22 [0.07, 0.38] | 0.47 |
| log age (nonlinear), pooled | all 413 contribute | 0.025 | 0.037 to 0.086 | 2 | 0.070 | 0.38 [0.25, 0.48] | 0.62 |
| Contrast | Median | Range | CI , raters | CI , raters + images | Consensus [95% CI, raters + images] |
|---|---|---|---|---|---|
| 60+ vs under-60, raw means | 0.048 | 0.010 to 0.136 | 1 of 17 | 0 of 17 | 0.126 [ 0.034, 0.209] |
| 60+ vs 18- 29, raw means | 0.042 | 0.017 to 0.110 | 1 of 17 | 0 of 17 | 0.114 [ 0.046, 0.205] |
| 50+ vs under-50, raw means | 0.013 | 0.074 to 0.053 | 0 of 17 | 0 of 17 | 0.042 [ 0.091, 0.162] |
| 50+ vs 18- 29, raw means | 0.003 | 0.069 to 0.058 | 0 of 17 | 0 of 17 | 0.043 [ 0.090, 0.150] |
| (a) Graded effects | (b) Step, slope | (c) Step, contrast | ||||||
|---|---|---|---|---|---|---|---|---|
| True fidelity | Mean observed | Power | Mean observed | Power | Mean observed | Power | ||
| 0.00 | 0.003 | 3% | 0.993 | 0.890 | 0.003 | 3% | 0.003 | 3% |
| 0.05 | 0.025 | 12% | 0.977 | 0.623 | 0.002 | 4% | 0.007 | 4% |
| 0.10 | 0.054 | 41% | 0.883 | 0.310 | 0.007 | 5% | 0.016 | 5% |
| 0.15 | 0.082 | 71% | 0.627 | 0.093 | 0.013 | 9% | 0.025 | 9% |
| 0.20 | 0.111 | 92% | 0.283 | 0.020 | 0.018 | 11% | 0.034 | 14% |
| Model | Mean older young | SD of image-specific shift | Ratio to human SD | Ratio, pleasantness removed | % of 0.595 | % of 0.86 | |
|---|---|---|---|---|---|---|---|
| GLM-4.5V | 0.143 | 0.852 | 1.93 | 3.03 | 0.093 | 15.6% | 10.8% |
| Claude-Sonnet-4.6 | 0.126 | 0.627 | 1.42 | 2.20 | 0.086 | 14.4% | 10.0% |
| Qwen3-VL-235B | 0.226 | 0.691 | 1.57 | 2.42 | 0.072 | 12.2% | 8.4% |
| Gemini-3-Flash | 0.028 | 0.592 | 1.34 | 2.05 | 0.065 | 10.9% | 7.6% |
| Gemini-3.8-Flash | 0.045 | 0.477 | 1.08 | 1.68 | 0.059 | 9.9% | 6.9% |
| Claude-Opus-5 | 0.116 | 0.385 | 0.87 | 1.35 | 0.058 | 9.8% | 6.8% |
| Format | Median older young | Negative shift | of which | Median, pleasant flagged | Median | Range of | CI | Median per-image with submission |
|---|---|---|---|---|---|---|---|---|
| Submission prompt | 0.065 | 4 of 5 | 3 | 0.812 | 0.085 | 0.006 to 0.136 | 1 of 5 | – |
| Chain of thought | 0.031 | 4 of 5 | 2 | 0.865 | 0.086 | 0.053 to 0.106 | 0 of 5 | 0.67 |
| Direct rating only | 0.096 | 4 of 5 | 4 | 0.600 | 0.088 | 0.008 to 0.128 | 0 of 5 | 0.67 |
| Paraphrase A | 0.099 | 4 of 5 | 4 | 0.753 | 0.093 | 0.045 to 0.121 | 0 of 5 | 0.70 |
| Paraphrase B | 0.069 | 5 of 5 | 3 | 0.759 | 0.120 | 0.036 to 0.135 | 0 of 5 | 0.77 |
| Numeric age | 0.047 | 5 of 5 | 2 | 0.712 | 0.076 | 0.052 to 0.116 | 0 of 5 | 0.73 |
| Format | Median older young | Negative shift | of which | Median, pleasant flagged | Median | Range of | CI | Median per-image with submission |
|---|---|---|---|---|---|---|---|---|
| Submission prompt | 0.184 | 5 of 5 | 5 | 0.647 | 0.133 | 0.237 to 0.057 | 0 of 5 | – |
| Chain of thought | 0.452 | 5 of 5 | 5 | 0.994 | 0.190 | 0.216 to 0.139 | 0 of 5 | 0.26 |
| Direct rating only | 0.233 | 5 of 5 | 5 | 0.629 | 0.065 | 0.146 to 0.004 | 0 of 5 | 0.26 |
| Paraphrase A | 0.250 | 5 of 5 | 5 | 0.694 | 0.073 | 0.168 to 0.058 | 0 of 5 | 0.41 |
| Paraphrase B | 0.237 | 5 of 5 | 5 | 0.706 | 0.036 | 0.191 to 0.012 | 0 of 5 | 0.47 |
| Numeric age | 0.128 | 5 of 5 | 4 | 0.476 | 0.121 | 0.159 to 0.058 | 0 of 5 | 0.46 |
| Model | Format | Older young | Flagged | Pleasant flagged | [95% CI] | Change [95% CI] | Per-image | ||
|---|---|---|---|---|---|---|---|---|---|
| GPT-5.6-Luna-Pro | Submission prompt | 298 | 0.065 | 0.005 | 0.173 | 0.535 | 0.081 [ 0.054, 0.216] | – | – |
| GPT-5.6-Luna-Pro | Chain of thought | 300 | 0.031 | 0.367 | 0.212 | 0.865 | 0.106 [ 0.021, 0.231] | +0.025 [ 0.106, 0.171] | 0.30 |
| GPT-5.6-Luna-Pro | Direct rating only | 300 | 0.059 | 0.005 | 0.183 | 0.365 | 0.008 [ 0.117, 0.099] | 0.089 [ 0.253, 0.081] | 0.17 |
| GPT-5.6-Luna-Pro | Paraphrase A | 300 | 0.126 | 0.001 | 0.354 | 0.676 | 0.078 [ 0.039, 0.205] | 0.003 [ 0.147, 0.139] | 0.32 |
| GPT-5.6-Luna-Pro | Paraphrase B | 300 | 0.032 | 0.212 | 0.050 | 0.488 | 0.036 [ 0.086, 0.169] | 0.045 [ 0.185, 0.098] | 0.28 |
| GPT-5.6-Luna-Pro | Numeric age | 299 | 0.047 | 0.056 | 0.073 | 0.371 | 0.076 [ 0.049, 0.203] | 0.005 [ 0.151, 0.146] | 0.26 |
| Model | Format | Older young | Flagged | Pleasant flagged | [95% CI] | Change [95% CI] | Per-image | ||
|---|---|---|---|---|---|---|---|---|---|
| GPT-5.6-Luna-Pro | Submission prompt | 298 | 0.179 | 0.001 | 0.343 | 0.647 | 0.133 [ 0.226, 0.031] | – | – |
| GPT-5.6-Luna-Pro | Chain of thought | 300 | 0.326 | 0.001 | 0.533 | 0.994 | 0.145 [ 0.248, 0.037] | 0.012 [ 0.123, 0.098] | 0.26 |
| GPT-5.6-Luna-Pro | Direct rating only | 300 | 0.165 | 0.001 | 0.321 | 0.629 | 0.049 [ 0.156, 0.057] | +0.084 [ 0.069, 0.222] | 0.04 |
| GPT-5.6-Luna-Pro | Paraphrase A | 300 | 0.190 | 0.001 | 0.335 | 0.694 | 0.059 [ 0.165, 0.046] | +0.074 [ 0.058, 0.203] | 0.18 |
| GPT-5.6-Luna-Pro | Paraphrase B | 300 | 0.253 | 0.001 | 0.202 | 0.300 | 0.016 [ 0.123, 0.086] | +0.118 [ 0.045, 0.266] | 0.02 |
| GPT-5.6-Luna-Pro | Numeric age | 299 | 0.126 | 0.001 | 0.275 | 0.476 | 0.107 [ 0.219, 0.009] | +0.026 [ 0.119, 0.169] | 0.09 |
| Dimension | Model | Format | Older young | Flagged | Unflagged | Pleasant flagged | [95% CI] | Per-image between formats |
|---|---|---|---|---|---|---|---|---|
| valence | Qwen3-VL-8B | paper prompt | 0.020 | 0.129 | +0.005 | 0.343 | 0.001 [ 0.074, 0.071] | – |
| valence | Qwen3-VL-8B | second format | 0.087 | 0.254 | 0.049 | 0.926 | 0.051 [ 0.111, 0.016] | 0.35 |
| valence | Qwen3-VL-32B | paper prompt | 0.219 | 0.626 | 0.126 | 1.491 | 0.002 [ 0.071, 0.075] | – |
| valence | Qwen3-VL-32B | second format | 0.247 | 0.628 | 0.160 | 1.587 | 0.012 [ 0.077, 0.052] | 0.65 |
| valence | Qwen3-VL-235B | paper prompt | 0.226 | 0.632 | 0.134 | 1.270 | 0.072 [0.002, 0.147] | – |
| valence | Qwen3-VL-235B | second format | 0.348 | 0.720 | 0.263 | 1.493 | 0.058 [ 0.011, 0.124] | 0.64 |
| Dim. | Model | Format | Older young | Flagged | Pleasant flagged | [95% CI] | Change [95% CI] | Per-image | ||
|---|---|---|---|---|---|---|---|---|---|---|
| valence | Claude-Sonnet-4.6 | submission | 300 | 0.086 | 0.017 | 0.544 | 1.371 | 0.085 [ 0.052, 0.228] | – | – |
| valence | Claude-Sonnet-4.6 | rating-only, third person | 300 | 0.206 | 0.001 | 0.621 | 1.588 | 0.105 [ 0.031, 0.249] | +0.020 [ 0.068, 0.111] | 0.78 |
| valence | GLM-4.5V | submission | 292 | 0.090 | 0.069 | 0.473 | 1.455 | 0.135 [0.008, 0.265] | – | – |
| valence | GLM-4.5V | rating-only, third person | 291 | 0.188 | 0.001 | 0.474 | 1.360 | 0.037 [ 0.087, 0.167] | 0.097 [ 0.210, 0.008] | 0.63 |
| valence | GPT-5.6-Luna-Pro | submission | 298 | 0.065 | 0.005 | 0.173 | 0.535 | 0.081 [ 0.054, 0.216] | – | – |
| valence | GPT-5.6-Luna-Pro | rating-only, third person | 300 | 0.104 | 0.001 | 0.106 | 0.135 | 0.141 [0.015, 0.264] | +0.060 [ 0.105, 0.217] | 0.17 |
| Model | Older | Retired | Cautious | Young |
|---|---|---|---|---|
| Claude-Sonnet-4.6 | 0.02 / 0.38 / 1.02 | +0.08 / 0.25 / 0.77 | 0.48 / 0.60 / 1.45 | +0.07 / +0.17 / +0.35 |
| Gemini-3-Flash | +0.12 / 0.22 / 0.50 | +0.13 / 0.37 / 0.61 | 1.04 / 0.66 / 1.46 | +0.10 / +0.11 / +0.31 |
| GLM-4.5V | 0.02 / 0.36 / 1.31 | +0.16 / 0.01 / 0.64 | 1.35 / 0.82 / 2.58 | +0.07 / +0.11 / +0.15 |
| GPT-5.6-Luna-Pro | 0.03 / 0.24 / 0.48 | +0.00 / 0.31 / 0.73 | 0.55 / 0.70 / 1.61 | +0.03 / 0.07 / +0.05 |
| Llama-4-Scout | 0.08 / 0.28 / 0.43 | 0.03 / 0.20 / 0.31 | 0.36 / 0.34 / 0.66 | 0.04 / 0.05 / 0.06 |
| Qwen3-VL-235B | 0.15 / 0.62 / 1.42 | 0.12 / 0.71 / 1.60 | 0.35 / 0.53 / 1.32 | +0.08 / +0.05 / +0.05 |
| Model | : retired vs older | : cautious vs older | : young vs older | Older cautious [95% CI] | Older-shift variance not explained by cautious | |
|---|---|---|---|---|---|---|
| Claude-Sonnet-4.6 | 0.89 | 0.56 | 0.02 | +0.462 [0.397, 0.527] | 0.001 | 69% |
| Gemini-3-Flash | 0.80 | 0.08 | 0.30 | +1.165 [1.042, 1.288] | 0.001 | 99% |
| GLM-4.5V | 0.70 | 0.43 | 0.13 | +1.334 [1.201, 1.467] | 0.001 | 82% |
| GPT-5.6-Luna-Pro | 0.58 | 0.34 | 0.58 | +0.520 [0.431, 0.609] | 0.001 | 88% |
| Llama-4-Scout | 0.70 | 0.50 | 0.49 | +0.285 [0.220, 0.350] | 0.001 | 75% |
| Qwen3-VL-235B | 0.79 | 0.60 | 0.12 | +0.202 [0.137, 0.268] | 0.001 | 64% |
| Model | Older | Retired | Cautious | Young | : retired vs older | : cautious vs older |
|---|---|---|---|---|---|---|
| Claude-Sonnet-4.6 | 0.139 | 0.141 | 0.204 | +0.105 | 0.85 | 0.52 |
| Gemini-3-Flash | 0.147 | 0.206 | +0.482 | +0.204 | 0.74 | 0.40 |
| GLM-4.5V | 0.340 | 0.450 | +0.261 | +0.076 | 0.63 | 0.44 |
| GPT-5.6-Luna-Pro | 0.144 | 0.298 | 0.170 | +0.035 | 0.61 | 0.45 |
| Llama-4-Scout | 0.115 | 0.196 | 0.227 | +0.049 | 0.58 | 0.60 |
| Qwen3-VL-235B | 0.125 | 0.190 | 0.100 | +0.058 | 0.67 | 0.54 |
| Model | Mem.: no age | Mem.: young | Mem.: older | Mem.: retired | Mem.: cautious | Vul.: no age | Vul.: young | Vul.: older | Vul.: retired | Vul.: cautious |
|---|---|---|---|---|---|---|---|---|---|---|
| Claude-Sonnet-4.6 | 2% | 5% | 57% | 70% | 1% | 13% | 18% | 30% | 28% | 62% |
| Gemini-3-Flash | 0% | 4% | 24% | 46% | 1% | 16% | 16% | 22% | 19% | 76% |
| GLM-4.5V | 0% | 1% | 32% | 52% | 1% | 16% | 15% | 24% | 18% | 69% |
| GPT-5.6-Luna-Pro | 1% | 1% | 4% | 8% | 0% | 18% | 21% | 26% | 26% | 48% |
| Llama-4-Scout | 0% | 1% | 11% | 7% | 0% | 11% | 13% | 18% | 15% | 30% |
| Qwen3-VL-235B | 1% | 6% | 72% | 78% | 1% | 17% | 18% | 28% | 26% | 56% |
| Model | Dataset | No-age valence | No-age arousal | Calibration slope | Young no age, valence ( ) | Young no age, arousal ( ) | RMSE, no age | RMSE, young | Change [95% CI] | Label changes | |
|---|---|---|---|---|---|---|---|---|---|---|---|
| Claude-Sonnet-4.6 | OASIS | 899 | 0.900 | 0.709 | 1.23 | +0.087 ( 0.001) | +0.088 ( 0.001) | 0.862 | 0.854 | 0.008 [ 0.028, 0.013] | 13% |
| Claude-Sonnet-4.6 | NAPS | 449 | 0.896 | 0.656 | 1.18 | +0.022 (0.099) | +0.106 ( 0.001) | 0.832 | 0.827 | 0.006 [ 0.029, 0.017] | 11% |
| GLM-4.5V | OASIS | 873 | 0.895 | 0.668 | 1.28 | +0.102 ( 0.001) | +0.073 (0.013) | 0.918 | 0.947 | +0.029 [ 0.002, 0.062] | 13% |
| GLM-4.5V | NAPS | 447 | 0.860 | 0.607 | 1.17 | 0.017 (0.470) | +0.223 ( 0.001) | 0.921 | 0.907 | 0.014 [ 0.066, 0.035] | 17% |
| GPT-5.6-Luna-Pro | OASIS | 895 | 0.890 | 0.675 | 1.35 | +0.050 ( 0.001) | +0.045 (0.004) | 1.055 | 1.085 | +0.031 [0.008, 0.052] | 12% |
| GPT-5.6-Luna-Pro | NAPS | 450 | 0.882 | 0.701 | 1.33 | +0.018 (0.323) | +0.010 (0.673) | 0.968 | 0.978 | +0.011 [ 0.026, 0.049] | 9% |
| Dataset | Models | Images | Mean pairwise | Range | Positive mean shift | Median shift |
|---|---|---|---|---|---|---|
| NAPS | 6 | 445 | 0.069 | 0.050 to 0.226 | 4 of 6 | +0.020 |
| OASIS | 6 | 869 | 0.203 | 0.085 to 0.387 | 5 of 6 | +0.088 |
| Contrast | Measure | Older | Young (matched ) | Older young [95% CI] | Keil & Freund predict | Models moving up |
|---|---|---|---|---|---|---|
| 60+ vs 18- 29 | Appetitive coupling (Fisher ) | 0.194 | 0.193 | +0.001 [ 0.188, 0.197] | down | 0 |
| 60+ vs 18- 29 | Aversive coupling (Fisher ) | 0.419 | 0.362 | +0.057 [ 0.166, 0.336] | up | 17 |
| 60+ vs 18- 29 | , linear | 0.020 | 0.014 | +0.006 [ 0.037, 0.040] | up | 16 |
| 60+ vs 18- 29 | , quadratic | 0.085 | 0.059 | +0.026 [ 0.052, 0.097] | down | 0 |
| 60+ vs 18- 29 | Mean valence, neutral images | 4.090 | 4.068 | +0.021 [ 0.190, 0.241] | up | 2 |
| 60+ vs 18- 29 | Mean valence, pleasant images | 5.487 | 5.178 | +0.309 [0.029, 0.585] | down | 1 |
| Target | Images | No age | Older persona | Young humans | Young humans + persona shift | Persona vs no age (worse / better) | Adding shift (worse / better) |
|---|---|---|---|---|---|---|---|
| 60+ raters | all 900 | 0.756 | 0.789 | 0.583 | 0.696 | 7 / 0 | 17 / 0 |
| 60+ raters | flagged pleasant (54) | 0.754 | 1.214 | 0.622 | 1.314 | 16 / 0 | 17 / 0 |
| 60+ raters | flagged unpleasant (113) | 0.923 | 0.914 | 0.594 | 0.682 | 2 / 4 | 6 / 0 |
| 60+ raters | unflagged (733) | 0.730 | 0.733 | 0.578 | 0.641 | 1 / 2 | 14 / 0 |
| 50+ raters | all 900 | 0.683 | 0.731 | 0.495 | 0.636 | 10 / 0 | 17 / 0 |
| 50+ raters | flagged pleasant (54) | 0.611 | 1.160 | 0.559 | 1.288 | 16 / 0 | 17 / 0 |
| Arm | What it does | [95% CI] | Direction agreement | Mean absolute error |
|---|---|---|---|---|
| S0 direct prompt | Submission prompt, GPT-5.6-Luna-Pro | 0.081 [ 0.061, 0.220] | 38.6% | 0.475 |
| S1 debias instruction | Submission prompt plus one instruction not to rely on age stereotypes | 0.002 [ 0.113, 0.115] | 44.1% | 0.465 |
| S2 deliberation agent | Perceiver → persona rater → stereotype critic → reviser; no human data | 0.018 [ 0.135, 0.098] | 40.1% | 0.535 |
| S3 grounded agent | S2 plus retrieved human age differences for similar images from the other three lists | 0.089 [ 0.039, 0.215] | 71.4% | 0.367 |
| S4 grounded contrastive panel | S3 evidence, five simulated raters per age band, a calibrator anchored on the no-age rating, and an auditor | 0.268 [0.161, 0.373] | 87.5% | 0.205 |
| S5 VERA | A hypothesized age-difference rule, kept only if it passes a re-test on held-out human cases; rating anchored on the no-age rating | 0.144 [0.028, 0.260] | 87.1% | 0.198 |
| Images | System | Older young | Pleasant flagged | Direction agreement | Mean absolute error | Pearson | [95% CI] | ( ) | |
|---|---|---|---|---|---|---|---|---|---|
| held-out 600 | S0 direct prompt | 596 | 0.082 | 0.478 | 38.1% (310) | 0.511 | 0.103 | 0.025 [ 0.101, 0.053] | 0.728 |
| held-out 600 | S4 grounded contrastive panel | 600 | +0.018 | +0.170 | 85.3% (518) | 0.210 | 0.696 | 0.105 [0.014, 0.196] | 0.011 |
| held-out 600 | Pleasantness-only rule (no model) | 600 | +0.000 | +0.167 | 94.3% (472) | 0.133 | 0.893 | – | – |
| original 300 | S0 direct prompt | 299 | 0.065 | 0.535 | 38.6% (140) | 0.475 | 0.110 | 0.081 [ 0.061, 0.220] | 0.127 |
| original 300 | S4 grounded contrastive panel | 300 | +0.019 | +0.171 | 87.5% (257) | 0.205 | 0.744 | 0.268 [0.161, 0.373] | 0.001 |
| original 300 | Pleasantness-only rule (no model) | 300 | 0.001 | +0.162 | 95.1% (224) | 0.140 | 0.887 | – | – |
| Images | Gain in [95% CI] | Directions fixed / broken | (McNemar) | (Wilcoxon, error) | |
|---|---|---|---|---|---|
| held-out 600 | +0.130 [0.006, 0.254] | 0.041 | 400 / 35 | 0.001 | 0.001 |
| original 300 | +0.187 [0.023, 0.353] | 0.024 | 211 / 15 | 0.001 | 0.001 |