← News

Models & toolsSources: arXiv · 20 Nov 2026

Persona lenses steer vision-language models' simulated web gaze

Researchers propose three ways to push a persona deeper into a vision-language model's gaze simulation, with effectiveness depending on model size.

In plain words

Vision-language models can imitate where a person looks on a web page, but a persona described in the prompt only partly changes that imitation. The authors test three methods to make the persona matter more, and find the best method depends on the model's size.

What was announced

The arXiv paper introduces persona lenses, which the authors describe as placing the persona between the model and the page. The text lens is a two-step prompt in which the model first writes what the persona would look at; the pixel lens marks the region matching the persona's interests; and the weight lens applies a dwell reward gated to keep the scanpath natural. The authors measure a persona's reach as the difference in one model's gaze between two MatrAIx personas differing in a single attribute.

Testing Qwen3-VL at 4B and 32B and Qwen3.5-9B on generated and live pages, the authors report that the text lens separates personas strongly but is not robust to profile changes that should not move gaze, while the pixel lens separates them less but is unaffected by gaze-irrelevant attributes. The weight lens separates personas on held-out interests at 4B and 9B without an extra step on a new page, and the 32B model separates personas with the base prompt alone, so lens choice depends on cost and model size.

What it does not show

  • The evidence is the paper's abstract only; full methods, baselines and caveats are not available for verification.

Sources

Source-support review completed 21 Nov 2026. This story was drafted with AI assistance and checked against retained source passages. Automated checking is not independent verification.

Coverage: arXiv: abstract.

Updates and corrections

No corrections so far.

First published 21 Nov 2026

New to the subject? AI from the Ground Up starts from the beginning.