Work

TRACE: Patch-Aware Retrieval-Augmented Reasoning for High-Recall Vulnerability Triage

Year: 2026
Type: preprint
Abstract: Security patches often preserve dangerous operations while changing the conditions under which they execute. LLM-based vulnerability detectors must therefore distinguish the presence of such operations from the persistence of their vulnerability causes. We present TRACE, a patch-aware retrieval-augmented framework for high-recall, function-level vulnerability triage. TRACE links structural evidence from historical vulnerable/fixed code pairs with explicit repair principles, retrieves relevant contrasts through dependency-aware re-ranking, and guides an LLM to check dangerous-path alignment, vulnerability-cause matching, and equivalent safeguards in the target function. Inference requires no target-side repair patch. On 586 PairVul repair pairs, TRACE achieves 93.52\% recall and 75.48\% F1, compared with 95.05\% and 73.58\% for VulInstruct. It correctly classifies both revisions in 40.27\% of pairs, versus 34.98\%, and produces 53 fewer false positives on repaired functions. On a fixed 301-function SecVulEval subset, TRACE achieves 92.50\% recall and 70.48\% F1, compared with 78.33\% and 62.88\% for VulInstruct. The F1 differences on both datasets and the PairVul pair-accuracy difference do not reach significance under Holm adjustment. Ablations show lower F1 without repair principles or fixed-side evidence, while error analysis identifies safeguard-equivalence failures as a recurring source of false positives in audited cases. These findings suggest that linked repair contrasts can support high-recall vulnerability triage, although reliably recognizing equivalent safeguards remains a limitation.
Language: English
Open Access status: green
OpenAlex ID: W7225993871
Cites: 3
Cited by:
Related to: 10