Search for Coverage: Learning Coverage-Aware Retrieval with Augmented Sub-Question Answerability

Researchers introduced CoveR, a bi-encoder for coverage-aware dense retrieval in long-form RAG.

They also released SCOPE, a 90K-pair dataset built with synthetic coverage labels from LLM-generated sub-question answerability judgments.

CoveR improves nugget coverage by about 10% over strong dense retrieval baselines while staying competitive on relevance.

The work addresses a key long-form RAG challenge: retrieving passages that cover multiple needed facts, not just the closest match.