Counterfactual Twin-Margin

New method enhances vision-language model control over conflicting modalities

Researchers have developed a new method to improve modality control in vision-language models (VLMs). The technique, called Grounding on Request, focuses on teaching models to follow specific modalities (image or text) …