Where Does Alignment Emerge and What Does It Mean? A Closer Look at the Platonic Representation Hypothesis
This study finds that language and vision model representations can align more as language-model capability rises, provided captions are clean and informative. Its baselines show that cross-modal similarity remains limited.
The paper examines the Platonic Representation Hypothesis, which proposes that neural networks develop similar representations. The authors report that capability-related alignment between language and vision models persists for newer models when captions are **clean and informative**. They also find that cross-modal alignment can exceed that of explicitly aligned vision-language models, while remaining well below same-modality alignment.
For builders comparing model representations, account for model depth and width and use suitable unimodal and multimodal baselines. Caption quality also matters when interpreting cross-modal similarity.
The paper examines the Platonic Representation Hypothesis, which proposes that neural networks develop similar representations. The authors report that capability-related alignment between language and vision models persists for newer models when captions are **clean and informative**. They also find that cross-modal alignment can exceed that of explicitly aligned vision-language models, while remaining well below same-modality alignment. For builders comparing model representations, account for model depth and width and use suitable unimodal and multimodal baselines. Caption quality also matters when interpreting cross-modal similarity. The findings concern representation similarity, not whether models behave interchangeably in an application. The abstract does not provide numerical effect sizes, and the reported alignment depends on the data and metrics used.
This supports cross-modal representation alignment under clean, informative captions, while showing that it remains weaker than same-modality alignment. The result is about similarity in learned representations; it does not establish equivalent behavior or interchangeable models. Caption quality and suitable modality-specific baselines therefore shape what the measured alignment means.