A November 5 arXiv preprint examines how a multimodal model responds when asked to rate images from different perspectives. Mathis Immertreu and colleagues report that the framing used to introduce a persona can move hidden representations substantially, while the content of the persona can have a greater effect on visible ratings.
What the experiment varies
The team studies visual interestingness judgments in Qwen3-VL-8B-Instruct across 100 images and 53 prompting conditions. The prompts vary persona content and the linguistic framing used to introduce it. The authors record both the rating and the layer-wise hidden states for each condition.
They analyze representational geometry to see how similarly the model encodes conditions. This is a separate measurement from the rating the model produces. Comparing the two helps distinguish a shift in internal representations from a change in what a reader sees.
What changes across layers
The authors report that early layers primarily encode image identity. Persona-specific structure emerges in later layers. In the analyzed model, prompt framing reshapes representations more strongly than persona identity, while persona content has the greater effect on behavioral ratings.
The paper also examines minimal identity statements. It finds cases where these statements alter representations even when the response stays the same. The size of a representation shift therefore provides limited evidence about how a persona will affect the model’s judgment.
How to evaluate a simulated perspective
The result comes from one model and one visual rating task under controlled prompts. An evaluation of simulated observers can measure whether the ratings actually vary as intended, then use layer measurements to investigate where a difference emerges. Both measurements answer different questions about the prompt’s effects.
Comments