Researchers Tackle Cross-Image Confusion in Multi-Image AI Models

Large vision-language models are increasingly asked to process multiple images at once, yet their ability to do so remains a weak point. Researchers have identified a key problem: when models handle several images together, visual details from different images bleed into one another, degrading their understanding of each scene.

The Problem: Information Leakage Across Images#

Current large vision-language models (LVLMs) perform well on single-image tasks, but their accuracy drops noticeably when they must reason about multiple images simultaneously. The underlying cause appears to be cross-image information leakage—a phenomenon where elements from different images become intertwined in the model's internal representations and responses. This confusion directly undermines the model's ability to focus on and reason correctly about individual images in a batch.

FOCUS: A Simple Solution Without Retraining#

Researchers have developed a method called FOCUS that addresses this problem without requiring model retraining or architectural changes. The approach is straightforward: the model processes a set of images one at a time, masking all others with random noise while keeping one target image clean. By focusing on each image individually in turn, the model generates predictions for each. These individual predictions are then combined and refined using a noise-only reference input, which filters out the leakage and produces more reliable outputs.

Broad Impact Across Tasks and Formats#

Testing on multiple multi-image benchmarks shows that FOCUS consistently improves performance. Beyond static multi-image scenarios, the method also generalizes to video understanding, suggesting it addresses a fundamental challenge in how these models process sequential or grouped visual information. By solving this leakage problem, FOCUS offers a practical enhancement for multi-image reasoning that practitioners can apply immediately to existing models.