Visual text similarity metrics mislead in multimodal AI models
The Alignment Illusion in Multimodal Large Language Models
Computer Vision and Pattern RecognitionMachine Learning
Summary
Multimodal large language models combine images and text to understand content. People often measure how well these models align visuals and words by looking at certain scores. The authors found that these scores can be misleading, because the model's inner structure can create false impressions of alignment even when the images are corrupted. They propose a new way to better detect true visual-text alignment by focusing on the geometry inside the model. This helps understand how these AI models process images and text together more accurately.
What this means in practice
- •For multimodal model developers: Improve evaluation of visual-text integration by using the principal-angle gap metric instead of standard scalar scores.
- •For ai model auditors: Detect when apparent visual-text alignment in models is superficial versus genuinely related to task performance using the principal-angle gap.
Authors
Hong-Han Wang, Yuntao Wang, Hu Ding
Abstract
Layer-wise visual-text similarity in Multimodal Large Language Models (MLLMs) is widely interpreted as evidence that the language model progressively integrates visual content into a shared representation space. This reading rests on the assumption that scalar alignment scores reflect content-level cross-modal interaction. To test this assumption, we apply controlled interventions to the visual stream. Across 13 MLLMs from five families spanning 0.5B to 72B parameters, replacing projector-output visual tokens with Gaussian noise sharply reduces task accuracy, yet four standard scalar measures (CKA, SVCCA, MIR, and the leading principal-angle cosine) fail to consistently separate the corrupted stream from the original. We call this failure the alignment illusion and trace it to the shared language-model pathway: anisotropic MLP down-projections pull visual and text tokens toward common output directions, producing weight-induced alignment. Because this component is essentially one-dimensional, we introduce the principal-angle gap (PA gap), defined as the difference between the top two principal-angle cosines, which separates weight-induced similarity from multi-directional visual structure. Under graded visual corruption, the PA gap tracks task accuracy more consistently than the scalar scores we consider; under a structured but irrelevant image, it further exposes regimes in which internal geometry and task accuracy come apart. Internal visual-text alignment in MLLMs is therefore best read as a geometric diagnostic of the visual stream inside the language model rather than a direct proxy for content-level cross-modal interaction, and is most informative when calibrated by controlled task evidence.