How do children learn two languages from the same input stream? A bilingual household doesn't label each sentence with its language — the child must simultaneously learn what each word means and which language it belongs to. These two tasks are usually modeled separately: language identification first, then word learning within each identified language. But children don't wait to identify languages before learning words.
Rastle et al. (arXiv:2603.12906) train language models on mixed French-English child-directed speech and find that the model learns both languages' statistical structures simultaneously, without explicit language labels. The word distributions, syntactic patterns, and phonotactic regularities of each language emerge from the mixed input — the model learns to separate what was never labeled as separate.
The mechanism: French and English have sufficiently different statistical signatures (word frequency distributions, bigram patterns, morphological regularities) that the model implicitly clusters the input into two overlapping but distinguishable distributions. The languages separate themselves in the learned representation because they occupy different regions of the statistical space.
The implication for acquisition theory: explicit language identification may not be a prerequisite for bilingual learning. If the statistical signatures are distinct enough, the languages self-segregate in the learner's representational space without top-down labeling. The child doesn't need to know “this is French” before learning French words — learning the words IS learning the language boundary.
The limitation is also instructive. When the two languages are very similar (e.g., closely related languages with overlapping vocabulary and syntax), the statistical signatures converge and the model struggles to separate them. Bilingual acquisition should be harder for typologically similar language pairs — a prediction that matches developmental data.