Language Confusion

A phenomenon observed in large language models whereby the model's output unexpectedly switches to an unintended language during inference, attributed to the mixing of multilingual training data. It is considered a characteristic defect of multilingual models.