Researchers have identified a significant flaw in how the internal workings of multimodal large language models (MLLMs) are understood, a phenomenon they term the "alignment illusion." This illusion arises because common methods used to measure how well visual and text information are integrated within these models can produce misleadingly high scores, even when the visual data is deliberately degraded. The findings, published on arXiv, suggest that current evaluation metrics may not accurately capture the true extent of cross-modal understanding in these AI systems.
The study, conducted by Hong-Han Wang, Yuntao Wang, and Hu Ding, examined 13 different MLLMs from five distinct families, ranging in size from 0.5 billion to 72 billion parameters. The researchers applied controlled interventions by replacing visual tokens, which are the model's internal representations of images, with random Gaussian noise. This intervention sharply reduced the models' task accuracy, indicating a failure in processing visual information. However, four standard scalar measures, CKA, SVCCA, MIR, and the leading principal-angle cosine, failed to consistently detect this degradation. These metrics are typically interpreted as evidence that the language model is progressively integrating visual content into a shared representation space.
The researchers traced the "alignment illusion" to the shared pathway within the language model. Specifically, they found that anisotropic Multi-Layer Perceptron (MLP) down-projections pull both visual and text tokens towards common output directions. This process creates "weight-induced alignment," where similarity scores are inflated due to the model's architecture rather than genuine content-level interaction. Because this alignment mechanism is essentially one-dimensional, the study proposes a new metric called the principal-angle gap (PA gap). This metric, defined as the difference between the top two principal-angle cosines, aims to distinguish between weight-induced similarity and multi-directional visual structure.
The PA gap demonstrated a more consistent correlation with task accuracy under graded visual corruption compared to the scalar scores previously used. Furthermore, when presented with a structured but irrelevant image, the PA gap effectively exposed scenarios where the model's internal geometry and its task accuracy diverged. This suggests that internal visual-text alignment in MLLMs should be viewed as a geometric diagnostic of the visual stream's processing within the language model, rather than a direct proxy for content-level cross-modal interaction. The study emphasizes that these alignment measures are most informative when calibrated with controlled task evidence.
This research challenges the common practice of interpreting rising layer-wise visual-text similarity as definitive proof of cross-modal integration. The findings indicate that MLLMs might appear more aligned than they actually are, potentially leading to overestimation of their capabilities. The proposed PA gap offers a more nuanced approach to evaluating these models, potentially guiding future development towards more reliable cross-modal understanding.