Search papers, labs, and topics across Lattice.
This study investigates whether monolingual language models can achieve cross-lingual alignment without the need for joint training, revealing that such models can indeed develop alignable representations through their inherent structure. The research demonstrates that alignment improves with increased data and model scale, and a Procrustes rotation can effectively map hidden states between models. Notably, the findings suggest that functional content can be transferred between languages, indicating that cross-lingual alignment is driven by the linguistic information rather than shared training processes.
Monolingual models can achieve cross-lingual alignment without joint training, revealing the power of linguistic structure over shared parameters.
Cross-lingual alignment in multilingual language models is typically attributed to joint training: shared parameters, mixed-language batches, or explicit alignment objectives. We ask whether monolingual models trained on non-parallel data learn alignable representations without joint training. By testing on strictly monolingual language models, such as the Goldfish model families and independently developed models from different research labs, we find three results. Correlation: these models develop alignable representational geometry across layers, with alignment strengthening as data scale, model scale, or linguistic proximity increases. Construction: a single Procrustes rotation fit on parallel sentences maps hidden states between models. Causation: the same rotation transfers functional content; patching a rotated English residual into a German model on a factual cloze flips the prediction to the donor's capital in most cases. We confirm that cross-lingual alignment can emerge from the structure of language and the information it carries rather than from joint training, and this points to practical future directions including model stitching, merging, and modular multilingual systems built from monolingual components.
