New training-free method helps AI models better identify relevant visual and textual evidence when answering questions about images

Look Twice (LoT) is an inference-time framework that improves Multimodal Large Language Models' ability to utilize both visual and textual evidence without requiring retraining

The method leverages model attention patterns to identify which image regions and retrieved text elements are most relevant to answering knowledge-intensive queries

LoT addresses a key challenge where MLLMs struggle with noisy or partially relevant retrieved text while needing to localize fine-grained visual information

The framework uses lightweight prompt-level modifications to highlight selected visual and textual cues before generating answers