Correcting a deployed AI model's facts is already hard — keeping those corrections through further training may be harder.
Researchers tested 254 experimental configurations to measure what happens to knowledge edits when a large language model is subsequently fine-tuned. The short answer: the edits mostly disappear. The starkest example in the study involves AlphaEdit applied to GPT-J on the zsRE benchmark, where more than a quarter of previously successful edits failed after fine-tuning. The team also found a counterintuitive wrinkle — fine-tuning only the non-edited layers caused more edit decay than fine-tuning all layers at once.
This matters in two directions that pull against each other. On one hand, a company that carefully patches an AI's bad outputs before deploying it can lose those patches the moment any downstream fine-tuning touches the model. On the other, anyone trying to scrub a malicious or covert edit out of a model has a practical tool: targeted fine-tuning of the edited layers removed edits while only modestly hurting performance elsewhere.
Knowledge editing was pitched as a cheaper alternative to full retraining — a way to correct LLMs without starting over. What this study quietly implies is that any pipeline treating edits as durable configuration rather than fragile state is building on shaky ground.