Zero Shot Visual Question Answering With Instruction Following

via “zero-shot visual question answering with instruction-following”

Salesforce's efficient vision-language bridge model.

Unique: Achieves zero-shot VQA by leveraging frozen LLM's instruction-following and generalization rather than training task-specific VQA heads, enabling single model to handle diverse question types through prompt engineering

vs others: Outperforms CLIP-based VQA classifiers on open-ended questions because it generates free-form answers via LLM rather than ranking predefined options, and more efficient than fine-tuned ViLBERT because it doesn't require task-specific training