Controlling for Stereotypes in Multimodal Language Model Evaluation
This work addresses the issue of stereotype bias in multimodal AI evaluation for researchers and developers, providing a controlled setting to assess model reliance on visual signals, though it is incremental as it builds on existing evaluation frameworks.
The authors tackled the problem of measuring how much multimodal language models rely on visual signals versus stereotypes by proposing a methodology and two benchmark sets for stereotypical colors and gender. Their results showed significant variation among models, with FLAVA being more sensitive to images and less affected by stereotypes compared to older models like VisualBERT and LXMERT.
We propose a methodology and design two benchmark sets for measuring to what extent language-and-vision language models use the visual signal in the presence or absence of stereotypes. The first benchmark is designed to test for stereotypical colors of common objects, while the second benchmark considers gender stereotypes. The key idea is to compare predictions when the image conforms to the stereotype to predictions when it does not. Our results show that there is significant variation among multimodal models: the recent Transformer-based FLAVA seems to be more sensitive to the choice of image and less affected by stereotypes than older CNN-based models such as VisualBERT and LXMERT. This effect is more discernible in this type of controlled setting than in traditional evaluations where we do not know whether the model relied on the stereotype or the visual signal.