Zhen Xu
Titles as the sources gave them. In 1 entry, with 0 checked quotations.
In the Base
Every entry they appear in.
DecoyCount: Evaluating Visual-Reasoning Misalignment in Counting
According to the abstract, the authors introduce a benchmark of 1,660 real images with designed decoys to test whether vision-language model counting errors come from perception or reasoning. They report frontier models below 50% accuracy and gaps of 11.8 to 22.1 percentage points between cued and uncued prompts.