Retrieve Few, Verify Many: Segment-Level Evidence Reuse for Long-Form Factuality Assessment
Abstract
AbstractExisting methods for evaluating the factuality of long-form LLM responses decompose text into atomic claims and verify each independently, retrieving evidence separately for every claim. This atom-independence assumption causes substantial retrieval redundancy, because adjacent claims in coherent text often draw from overlapping sources, a property we term semantic locality. We propose SegmentFact, a deterministic factuality assessor that partitions responses into semantic segments, extracts claims within each segment’s local context, and propagates retrieved evidence forward through a shared pool. Experiments on multiple benchmarks show that SegmentFact significantly reduces retrieval calls while maintaining accuracy comparable to existing detections.