Mitigating Cross-Image Information Leakage in Multi-Image Understanding with Large Vision-Language Models
The episode discusses a paper titled "Mitigating Cross-Image Information Leakage in Multi-Image Understanding with Large Vision-Language Models." The hosts explain that large vision-language models struggle with multi-image inputs because they mix visual cues across pictures. The authors propose FOCUS, a training-free inference method using noise and contrastive subtraction to force the model to focus on one image at a time, leading to significant performance gains across multiple benchmarks without retraining.