Reproduce “Asking What a Director Sees Can Change What Vision-Language Models Choose” on OpenResearch

Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.

Universal app for Apple silicon and Intel. Requires macOS 11 or later.

About this paper

Jiaxuan Guo, Xingyu Ren, Kejia Zhang, Jingxin Yang, Youran Sun, Haizhao Yang

The authors test whether asking a model what a director can see changes its later choice in a rendered block-moving task. They compare otherwise matched prompts, with and without a preliminary visibility question. For gpt-5.6-terra, the question increased hidden-block choices by 88.1 percentage points; increases occurred on many tested models, but not all, in this task.

Summary by alphaXiv.

Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.