179 questions
No questions match those filters.
← All questions
LLM FoundationsMedium· Position Encodings· Context Length
What breaks when you deploy a model trained with absolu...
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 plansWhen a model encounters positions beyond its training length (L-1), it attempts to access embedding rows that were never trained. This causes the model to produce incoherent output, typically manifesting as NaN values or near-uniform probability distributions. Fixing this requires continued training with an extended position table, rather than just architectural changes.