Persona Lenses: What Makes a Vision-Language Model See Through Its Persona's Eyes?
Organizations: LG Uplus
Abstract
Vision-language models have been used to simulate where a user looks on a web page. However, a persona in the prompt reaches the simulated gaze only partly. We measure this reach as the difference in one model's gaze between two MatrAIx personas that differ in a single attribute. We introduce persona lenses, which place the persona between the model and the page: a text lens (a two-step prompt in which the model first writes what the persona would look at), a pixel lens (marks on the region matching the persona's interests) and a weight lens (a dwell reward gated to keep the scanpath natural). Using Qwen3-VL (4B and 32B) and Qwen3.5-9B on generated and live pages, we compare the lenses with one another and with scale, the 32B model with no lens. It has been confirmed that the text lens separates the personas strongly but is not robust to profile changes that should not move gaze. The pixel lens separates the personas less but is unaffected by attributes unrelated to gaze. At 4B and 9B, the weight lens separates the personas on interests held out from training, with no extra step on a new page. The 32B model separates the personas with the base prompt alone. Therefore, the choice of a lens depends on cost and model size.
Figures & tables
| Interest | Model | True sect. | Pixel lens (rec.) | Pixel rand. | Text lens |
|---|---|---|---|---|---|
| Sports | 4B | (.18) | |||
| 9B | (.66) | ||||
| 32B | (.19) | ||||
| Fashion | 4B | (.43) | |||
| 9B | (.44) | ||||
| 32B | (.21) |
| Both sets (87 cells) | Later set (43 cells) | |||||
|---|---|---|---|---|---|---|
| Condition | 4B | 9B | 32B | 4B | 9B | 32B |
| Text lens | 70 (80) | 76 (87) | 65 (75) | 35 (81) | 38 (88) | 34 (79) |
| Pixel lens | 34 (39) | 30 (34) | 28 (32) | 21 (49) | 17 (40) | 13 (30) |
| and above random | 26 (30) | 21 (24) | 27 (31) | 14 (33) | 12 (28) | 13 (30) |
| True-section marker | 57 (66) | 65 (75) | 67 (77) | 31 (72) | 35 (81) | 32 (74) |
| Text + pixel lens | 74 (85) | 80 (92) | 67 (77) | 37 (86) | 39 (91) | 30 (70) |
| Held-out | |||
| Variant | Fash. | Trav. | Conc. |
| Base 4B | |||
| + text lens † | |||
| Strong reward | – | ||
| Relaxed gate | – | ||
| Base 9B | |||
| Model | Trav. | ScanMatch | Img.-spec. |
|---|---|---|---|
| Human viewers | — | ||
| Joint model, 4B | |||
| Joint model, 9B | |||
| Joint model, 32B |
| App. | Contents | Main |
|---|---|---|
| A | System prompt, task sentences, answer length, text lens, generated pages, live-page cells | Secs. 3.1, 4.1, 4.3, 4.5 |
| B | Perceptual and cognitive traits, declarative probe, input format | — |
| C | Pixel lens, blob detector, page-level tests, pages without topic names, other model family and encoder, further interests, earlier lens | Secs. 3.2, 4.3, 4.5 |
| D | Retrained adapters, training settings, gate and its relaxation, text lens in training, irrelevant attributes | Secs. 3.3, 4.4 |
| E | Registered scale run, live portal pages, text size | Sec. 4.5 |
| F | Attention measure, target attention by model, masks, 12 registered predictions | Sec. 5 |
Appendix figures & tables11 assets
Supplementary material from the paper’s appendix.
Appendix
| Setting | Interest | Pixel lens | True section | Rec. |
|---|---|---|---|---|
| InternVL3.5-8B | Sports | |||
| Fashion | ||||
| Finance | ||||
| 4B, CLIP | Sports | |||
| Fashion | ||||
| Finance |
| Target attention | |||
| Condition | Fash. | Trav. | Dwell vs. base |
| Base 4B | — | ||
| Strong reward | / | ||
| Relaxed gate | / | ||
| Base 9B | — | ||
| relaxed gate | / | ||
| 4B | 32B | ||||
| Masked layers | True-section marker | Text lens | Strong | Gate | Base |
| Unmasked, fash. | |||||
| Unmasked, trav. | |||||
| First quarter | / | — | — | ||
| Second quarter | / | — | — | ||
| Third quarter | / | — | / | ||
| Model | Seeds | Six-int. sum | Fashion | Travel | ScanMatch | Image-specific | Dwell (ms) |
| Human viewers (UEyes) | — | — | — | — | 279 | ||
| Base 4B, no training | 1 | — | — | — | |||
| Human-gaze model, 4B | 1 | ref. | ref. | 297 | |||
| Joint model, 4B | 3 | 263 | |||||
| Human-gaze model, 9B | 1 | ref. | ref. | — | |||
| Joint model, 9B | 3 | 262 ∗ |
| Layers | |||||
|---|---|---|---|---|---|
| Interest | Condition | 1–9 | 10–18 | 19–27 | 28–36 |
| Fashion | Text lens | ||||
| Pixel lens | |||||
| True-section marker | |||||
| Relaxed gate | |||||
| Strong reward | |||||
| Text lens | Pixel lens | True-section marker | Pixel lens on | ||||||
| Model | Interest | Attn. (seeds) | Dwell | Attn. (seeds) | Dwell | Attn. | Dwell | text lens | |
| Five pages of the weight-lens evaluation set | |||||||||
| 4B | Fash. | (3/5) | (5/5) | ||||||
| Trav. | (4/5) | (0/5) | |||||||
| 9B | Fash. | (3/5) | (1/5) | ||||||
| Trav. | (5/5) | (5/5) | |||||||
| Model | Set | CC: base text lens | ScanMatch: base text lens |
|---|---|---|---|
| 4B | All 108 | ||
| [ , ] | [ , ] | ||
| Web 27 | |||
| [ , ] | [ , ] | ||
| 9B | All 108 | ||
| [ , ] | [ , ] |
| Setting | gpt-6.1-terra | gpt-6-luna |
|---|---|---|
| Model name returned | The requested name in all 31,813 successful calls | The requested name in all 31,866 successful calls |
| Access | Institutional API gateway that forwards requests to a third-party relay, for both models | |
| Date | 11 November 2026. Design calls on the generated-page set within 129 minutes | 11 November 2026. Design calls on the generated-page set within 179 minutes, including 65 minutes with no call around a restart |
| Price per million tokens | 5 output, the rates of the gateway’s cost report | 0.50 output, published prices not checked against the gateway |
| Reasoning | Requested off. Reasoning tokens in 15 of 28,800 gaze calls, 0 of 2,760 note calls and 1 of 13 calls of the marker-reading check, 5,460 tokens in all | Requested off. No reasoning tokens in any call |
| Route check | One input-token count (33) and 24 different answers in 24 identical requests. Distinct values in 24 random-integer requests: 7 at temperature 0, 10 at 2.0, 8 with no setting and 1 at top- | One input-token count (32) and 24 different answers in 24 identical requests. Distinct values in 24 random-integer requests: 6 at temperature 0, 3 at 2.0, 5 with no setting and 1 at top- |
| Interest | Pages | Pairs | Base prompt [95% CI] | True sect. | Pixel lens (rec.) | Pixel rand. | Text lens | Text pixel |
|---|---|---|---|---|---|---|---|---|
| gpt-6.1-terra | ||||||||
| Sports | With topic names | 480 | [.442, .479] | (.79) | ||||
| Without topic names | 480 | [.355, .385] | (.61) | |||||
| Without with | 480 | |||||||
| Fashion | With topic names | 480 | [.272, .312] | (.63) | ||||
| generated only | 420 | [.275, .319] | (.68) | |||||