# Study finds asking vision-language models about a director’s view changes their actions

> When prompted to report what a director can see, six GPT models and Qwen3.8‑27B choose hidden blocks far more often, showing the question itself alters behavior.

Oossa · 2026-10-22 · https://oossa.com/en/study-finds-asking-vision-language-models-about-a-director-s-view-changes-their-

Researchers from arXiv report that vision‑language models (VLMs) behave differently when first asked what a director can see. In a “Director Task” where a large block is hidden from the director, asking the model which blocks the director sees raised the chance of picking the hidden block by 12 to 88 percentage points for six GPT models and by 17 points for Qwen‑3.8‑27B. The models still reported the director’s view correctly, but the prior question shifted their subsequent action toward their own perspective.

## The facts

- Paper submitted 21 Oct 2026; published 22 Oct 2026
- Hidden‑block choice increased up to 88 % points for GPT models after a perspective question

## Why it matters

Evaluators must include a baseline without a perspective query, or risk mis‑judging VLMs’ true ability to reason about others.

## Sources & references

1. [Asking What a Director Sees Can Change What Vision-Language Models Choose](https://arxiv.org/abs/2610.29318) – arXiv, 2026-10-22

Last updated: 2026-10-22