A new paper on arXiv introduces IMPLICIT-Bench, a benchmark designed to test whether text-to-image (T2I) models can follow explicit prompts when those prompts conflict with stereotyped defaults. The authors observe that T2I models often produce stereotypical images when stereotype-relevant attributes are left unspecified, and they ask whether an explicit request for a different attribute is enough to override that tendency.

The study's findings suggest that explicit requests are not always sufficient. Even when users clearly specify an attribute that runs against a stereotype, the model may still fall back on its default behavior. This points to a limitation in current prompt-following capabilities, separate from the well-known issue of unspecified attributes leading to biased outputs.

Because the source is a single arXiv abstract, the article cannot compare independent findings or verify the benchmark's methodology beyond what is stated. The key contribution is the benchmark itself, which offers a way to measure how often models prioritize stereotyped defaults over explicit user instructions.