Modality distraction identified and reduced in vision language models
MoGround: Measuring and Mitigating Modality Distraction in Vision-Language Models
Machine Learning
Summary
Vision-language models sometimes get confused when they see information from both pictures and text together, even if they got the right answer looking at just one of them. The authors created a new dataset called MoGround where each question can only be answered by either the picture or the text, never both. This helps them find when a model is distracted by irrelevant information from the other source. They tested seven models and found distraction varies depending on how strong each model understands each type of input. They also designed a way to reduce this distraction that works across all tested models without much loss in overall accuracy.
What this means in practice
- •For machine learning engineers: Improve robustness of multimodal AI by detecting and reducing errors caused by irrelevant visual or textual information.
- •For computer vision teams: Design evaluation tests that isolate how well models rely on one modality without distraction from the other.
Authors
Luca Zhou, Bo Zhao, Rose Yu, Emanuele Rodolà, Roberto Dessì
Abstract
We release MoGround, a vision-language dataset spanning four visual domains in which the answer to every question is guaranteed to be available from exactly one modality. This guarantee enables us to measure modality distraction, the failure in which a model answers a question correctly from one modality alone and then flips to a wrong answer once irrelevant content from the other modality is added. Existing probes rarely establish single-modality answerability this way, making it hard to isolate distraction in the first place. Across seven open-source VLMs, we find that modality distraction is not universal but model-dependent. The weaker-grounded modality is the more distracted one (r = +0.86), and distraction scales inversely with grounding strength (r = -0.90). The single-modality guarantee also enables a mitigation method that needs to distinguish between relevant and irrelevant context. Trained on one split of MoGround alone, a weight-space robustness vector reduces distraction on all seven models by 9% to 51%, at a cost of only 0.1 average points of accuracy on standard multimodal tasks.