The Alignment Illusion in Multimodal Large Language Models

TL;DR - A study of 13 multimodal LLMs finds that common visual-text alignment scores can remain high even when visual tokens are corrupted and task accuracy collapses. It attributes this “alignment illusion” to shared language-model weights and proposes the principal-angle gap as a more reliable geometric diagnostic.

• CKA, SVCCA, MIR, and leading principal-angle cosine often failed to distinguish genuine visual tokens from Gaussian noise.
• Anisotropic MLP down-projections can pull visual and text representations toward shared output directions, creating largely one-dimensional, weight-induced similarity.
• The proposed principal-angle gap measures the difference between the top two principal-angle cosines to separate this effect from multidirectional visual structure.
• Under graded corruption, the gap tracked task accuracy more consistently, though irrelevant-image tests showed that internal geometry still need not correspond directly to task performance.