The Alignment Illusion in Multimodal Large Language Models
Across 13 multimodal LLMs, standard visual-text alignment scores miss corrupted visuals, an effect called the alignment illusion.
The paper challenges the idea that layer-wise visual-text similarity in multimodal LLMs shows genuine content-level integration. Across 13 models from five families, from 0.5B to 72B parameters, replacing projector-output visual tokens with Gaussian noise sharply reduces task accuracy, yet CKA, SVCCA, MIR, and the leading principal-angle cosine often fail to separate the corrupted stream. The authors call this the alignment illusion and trace it to anisotropic MLP down-projections that pull tokens toward shared directions. Their principal-angle gap, the difference between the top two principal-angle cosines, tracks accuracy more consistently under graded visual corruption.
- Gaussian noise in visual tokens cuts accuracy but scalar scores often stay high.
- Tests cover 13 MLLMs from five families, spanning 0.5B to 72B parameters.
- Anisotropic MLP down-projections create mostly one-dimensional weight-induced alignment.
- The principal-angle gap tracks accuracy better than CKA, SVCCA, and MIR.
Full article236 words · extracted from arxiv.org · click to collapse
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.
Text extracted automatically; images, tables and formatting may be missing. Original: https://arxiv.org/abs/2609.30210