Where Does Alignment Emerge and What Does It Mean? A Closer Look at the Platonic Representation Hypothesis
More capable language models still show greater alignment with vision models—provided the captions are clean and informative.
The authors examine representation similarity across model layers and dimensions, testing whether the relationship holds for newer models. They compare results against unimodal and multimodal baselines, since model depth and width can confound similarity metrics. Separately trained vision and language models can show greater cross-modal alignment than explicitly aligned vision-language models. Even so, their alignment remains substantially below that of models working in the same modality.
ArXiv · AI/CL/LG's note
The authors examine representation similarity across model layers and dimensions, testing whether the relationship holds for newer models. They compare results against unimodal and multimodal baselines, since model depth and width can confound similarity metrics. Separately trained vision and language models can show greater cross-modal alignment than explicitly aligned vision-language models. Even so, their alignment remains substantially below that of models working in the same modality.
ArXiv · AI/CL/LG's note
score 4