SupersedeRank: Query-Time-Aware Reranking of Superseded CVSS Evidence in Retrieval-Augmented Generation
Year: 2026
Type: preprint
Abstract: Retrieval-augmented generation systems ground their answers in evidence retrieved from a document corpus. Over time, older records may remain semantically relevant even after being superseded by newer records. This phenomenon can be systematically examined in vulnerability management because the National Vulnerability Database (NVD) records the old and new values of revised CVSS assessments and the corresponding change dates. In this study, we develop SupersedeRank, a query-time-aware, explainable temporal reranking method with no learned parameters. The method preserves first-stage retrieval candidates. Evidence in the query-targeted source chain is moved behind protected candidates only if it is inferred to have been superseded by a newer and different assessment visible at query time, without coefficients or thresholds. Reference labels were programmatically derived from explicit old-to-new value transitions in official change records without additional human annotation. In a future-entity temporal test comprising 1,506 queries for 1,356 vulnerabilities unseen during training, SupersedeRank-Rule-Target ranked current evidence first for 99.8% of queries. It preserved current evidence for every query and left no superseded target evidence among the top-ten candidates. The strongest baseline, BM25 with a recency bonus, promoted current evidence at a comparable rate but retained superseded evidence among the top ten. The experimental scope is limited to atomic CVSS replacement events within the same source chain and change event.
Source: Research Square
Author: Doygun Demirol
Institution: Bingöl University
Language: English
Subfield: Information Systems
Field: Computer Science
Domain: Physical Sciences
Open Access status: green
OpenAlex ID: W7236201804
