Researchers at MIT CSAIL asked a large language model to edit another, smaller open model’s weights — the numbers that make up a neural network — directly.
Without classic retraining, by changing only 0.016% of those numbers, its math score rose from 7.9% to 48.1%. The AI mainly removed a direction tied to reasoning loops: the small model kept repeating the same text instead of reaching an answer.
It is not magic. The authors also show that this does not create knowledge the model never had. It mainly opens a new way to repair a targeted flaw.
Behind the numbers, a human still asks the question, and decides what is worth fixing.