CROSS-QUERY UTILITY MEMORY FOR RETRIEVAL-AUGMENTED GENERATION UNDER A FEEDBACK BUDGET

Abstract

Question-answering systems based on retrieval-augmented generation receive streams of queries that repeatedly require the same evidence. This paper studies a cross-query memory that stores which passages contributed to earlier answers and promotes them for similar new queries. The memory scores a passage based on its dense semantic similarity to the query and a kernel-weighted average of the credit assigned to it by similar past queries. We evaluate the method on replayed query streams using a frozen 4B-parameter language model. On MuSiQue, the memory yields a 0.016 F1 gain when it learns only from the passages it displays, while perfect labels for the same passages provide no further improvement. Allowing the model to inspect five to twenty graph neighbours of the displayed passages adds 0.019–0.024 F1, while the gain is approximately 0.04 F1 for each query that reuses earlier evidence. The novelty of the work lies in showing that the gain is constrained by the number of passages the system learns about at each step rather than by label quality, and that normalization of the query kernel is necessary.