179 questions
No questions match those filters.
← All questions
RAGVery Hard· Marginalization· API Constraints· Training Efficiency
A colleague proposes REPLUG-style marginalization over...
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 plansAcknowledge that the API does not block marginalization, as per-token log-probabilities are available. The constraint is compute: 32 calls per step means 32x fewer gradient steps compared to Gumbel top-1. The decision rests on whether the retriever’s failure is due to ranking within the candidate set (marginalization helps) or recall (neither helps). The nightly index rebuild limits how much the encoder can drift, favoring query encoder training.