Multilingual fact transfer is limited during continued pretraining

The SIFT framework shows that factual knowledge learned in English fails to transfer to Persian during continued pretraining when target-language data is systematically stripped. Most multilingual models rely on direct target-language data rather than cross-lingual knowledge transfer for specific facts.