Language similarity is a central concept in multilingual NLP. Prior work typically infers similarity from the static geometry of representation spaces, and the results often depend on model architecture, training data, and probing methods. In this work, we propose transition-based language similarity (TBLS), a new measure that characterizes language similarity through cross-lingual transition. The key intuition is that if two languages occupy nearby regions in a model’s representation space, model outputs are more likely to transition between them under small local variations. By aggregating such transition behavior across languages, TBLS can be derived, reflecting language proximity encoded by the model.

We operationalize this idea by perturbing model activations and measuring the resulting cross-lingual transition behavior. Across five multilingual language models and two corpora covering 44 languages, TBLS exhibits strong cross-model and cross-corpus consistency. Further analyses show that TBLS captures meaningful linguistic regularities and transfer-related structure, while activation perturbation provides behavioral evidence consistent with the hypothesis that semantically aligned multilingual representations occupy nearby, locally connected regions. These findings suggest that perturbation-induced transition dynamics reveal a surprisingly stable form of multilingual relationships across models and corpora.