Multimodal language models can simulate different observers, yet it remains unclear how assigning a perspective reshapes their internal representations and whether such changes drive the model's judgments. We examined the representational geometry of persona‑conditioned visual interestingness in Qwen3‑VL‑8B‑Instruct. Using 100 images and 53 prompting conditions, we systematically varied persona content and the linguistic framing that introduced it, while recording behavioural ratings and layer‑wise hidden states. We found a systematic reorganisation of geometry across processing depth: early layers mainly encode image identity, whereas persona‑specific structures emerge in later layers. Strikingly, prompt framing reshapes internal representations more strongly than persona identity, while persona content has a greater impact on behavioural judgments. Minimal identity statements can alter representations without changing responses, illustrating that representational distance does not necessarily reflect behavioural relevance. Broadly, these findings show that simulated perspectives form distinct structures within the representational space of artificial cognitive systems, highlighting a fundamental distinction between the geometry of internal representations and their functional consequences.
Review: The work reveals divergent effects of linguistic framing and persona information across model layers, offering valuable insight into how multimodal systems separate "seeing" from "evaluating" in their internal geometry.