179 questions
No questions match those filters.
Your team applied 5,000 sequential ROME edits to a 6B m...
This is one of the questions in the full AI/ML interview bank. Pro unlocks all 1789 questions; Premium includes the same bank plus the highest daily Practice limit.
See plansThe mechanism behind this failure is that each ROME edit is a rank-one update. The total accumulated update cannot exceed the rank limit of the model’s layers (min(d, dmlp) = 4,096). After roughly 4,000 edits, new writes are forced into directions already occupied by previous edits, causing interference. The high efficacy on the edit set is not evidence of success, but rather a reflection of the most recently installed facts. The structural fix is to batch edits and distribute them across a band of layers, as seen in the MEMIT approach.