Verified, then logged.
Sun 22 Nov 2026 · 00:14 KST / 15:14 UTC한국어 | ENSearch /

Breaking01 / AI

Qwen3-VL 32B varies simulated gaze with a basic prompt alone

Researchers reported that asking the model to describe where a user would look before simulating gaze made the results sensitive to changes in profile traits unrelated to gaze.

A monitor in a naturally lit database workspace displays numerous data tables alongside five shortlisted tables arranged in ranked order. AI-generated image

Researchers reported on arXiv on 20 Nov 2026 that Qwen3-VL 32B produced different simulated gaze patterns for different user profiles with the base prompt alone.

The researchers used vision-language models, which process images and text, to simulate where users would look on web pages. They measured how much a model's simulated gaze changed between profiles that differed in a single trait.

The paper reported large differences between profiles when the model first wrote down where a user would look. This approach also responded to changes in traits unrelated to gaze. Marking areas of the page that matched a user's interests produced smaller differences between profiles and held steady when unrelated traits changed.

The researchers also trained Qwen3-VL 4B and Qwen3.5-9B using a reward based on how long the simulated gaze stayed in an area. On new pages, both models varied their simulated gaze according to interests left out of training, without an extra step.

The researchers did not quantify the size of the gaze differences by condition in the abstract.