Strengthen retrieval scoring under bounded cost

Develop stronger retrieval scores for selecting evidence windows within LEAP’s bounded-cost, block-wise audio-visual retrieval structure.

Background

LEAP partitions long recordings into fixed-duration blocks, scores candidate windows within each block, retains a bounded number of high-scoring windows, and re-encodes only those windows in the answer pass. The framework’s context and peak working memory therefore remain independent of recording duration, but its effectiveness depends on the quality of the localization and ranking scores used to identify relevant evidence.

The conclusion explicitly identifies stronger retrieval scores as an unresolved direction for improving evidence selection without abandoning the bounded-cost structure that enables hour-scale audio-visual question answering.