retrieval-harms-popular-entity-accuracy

Summary

Adding external document context to help an LLM answer a question can actually make it less accurate on well-known topics, because the fetched passages may contain outdated or contradictory information that overrides what the model already knows correctly. This means a system should not assume retrieval is always beneficial; for questions the model can already handle well, pulling in extra context risks doing more harm than good, so the decision to retrieve should depend on the type of question rather than being applied uniformly.