Combining separately trained AI translation models breaks when languages diverge — here's why

A standard technique for merging independently trained AI models fails in multilingual translation, with performance dropping the more different the target languages are. The problem: fine-tuning on different languages reshapes how neurons inside the model organize themselves, making those independently trained models incompatible with simple merging techniques.

In practice, this means companies can't cheaply combine translation models trained separately on different language pairs by just averaging their internal weights — a shortcut that works fine for single-task models but breaks down as soon as you add language diversity. The finding matters because it forces a choice: either go back to the expensive, slow process of retraining everything from scratch, or invest in merging methods that account for how languages fundamentally reorganize a model's internal geometry. Right now, there's no clear winner.