Original title: Probing Factual Knowledge Transfer with Training Data Interventions
Romina Oji, Marc Braun, Marcel Bollmann, Marco Kuhlmann +1 · September 2026
If an AI learns a fact in English, can it use that fact in another language? Researchers tested this to see if models truly share knowledge across languages, or if they just memorize facts directly from each language's texts. They took an English-trained model and kept training it on Persian text. Before training, they carefully erased specific facts from the Persian data. Then, they tested if the model still knew those facts in Persian. It turns out that knowledge transfer is very limited. When facts were strictly removed from Persian data, the model failed to transfer a large majority of facts learned in English. How often an entity appeared in English mattered greatly. Facts about Persian subjects were much rarer in English and hardly transferred at all.
Imagine you learned to cook a dish by reading instructions in English. Now, you're asked to cook the same dish in a kitchen where all labels are in a language you didn't use for learning. If you didn't explicitly memorize the ingredient names and steps in that language, you might fumble or guess. In this paper, AI models learn facts in English and are tested in Persian; without explicit learning in Persian, the facts don't transfer, much like forgetting a recipe when the labels change.
If AI models can't use facts across languages, then non-English users get less accurate information from tools like search or chatbots. This affects anyone relying on multilingual AI for knowledge in their own language.
The study is limited to Persian and a specific set of 500 facts; it may not generalize to all languages or types of knowledge. The removal methods might not perfectly eliminate all traces of facts, and easier test sets can inflate apparent transfer.
Ask anything about this paper. Answers come from its abstract, and say so when the abstract doesn't cover your question.
Probing Factual Knowledge Transfer with Training Data Interventions — Romina Oji, Marc Braun, Marcel Bollmann et al.
Abstract licence: CC0
Explanation generated by Synthesis from the abstract — it is not the original paper. This is a preprint, so it hasn't necessarily been peer reviewed.