Brief

Researchers have identified specific attention heads within Large Multimodal Models (LMMs) that are crucial for processing visual relations. By extracting and manipulating these "function vectors," they can improve the models' zero-shot accuracy on relational tasks. This approach allows for fine-tuning these vectors without altering the main LMM parameters, outperforming standard in-context learning methods and demonstrating strong generalization capabilities for visual analogy problems. AI