AI & ML
X-LocVul: An Evidence-Propagation Audit of Cascaded C/C++ Vulnerability Detection, Localization, and Explanation
arXiv cs.SE0 views
arXiv:2611.21626v1 Announce Type: new Abstract: Developer-facing vulnerability analysis should link a function-level warning to a source location and a supported explanation. Cascaded systems can propagate weak evidence despite plausible components. We present X-LocVul, a measurement-first audit of a modular C/C++ detector-localizer-explainer pipeline combining UniXCoder, CodeT5, and Qwen2.5-Coder. We test whether evidence passed across stages is useful, merely source-resident, or misleading. Under duplication and temporal-leakage controls, UniXCoder has higher point estimates than CodeBERT, but true-CWE supervision is statistically indistinguishable from a shuffled-CWE control on PR-AUC, MCC, and vulnerable-class F1. Raw generated statements fail exact source matching in 90.83% of cases. Fuzzy projection reaches 83.79% Top-1 accuracy and exceeds semantic projection by 1.59 percentage points, although its advantage over raw ranking is not established. In a randomized, condition-label-masked two-rater evaluation of 50 paired cases, a predicted line improves evidence support, whereas a predicted CWE improves no endpoint. Oracle evidence improves evidence support (Holm-adjusted p=0.00453) and CWE consistency (p=3.49 x 10^-7), but not root-cause quality after correction (p=0.250); unsupported claims remain common. A chronological stress test retains ranking signal (ROC-AUC 0.782; PR-AUC 0.110), while the transferred threshold yields zero recall and F1. The contribution is not a state-of-the-art claim, but an auditable account of evidence construction, transfer, and consumption. It separates source residency from localization correctness, auxiliary labels from semantic supervision, and predicted evidence from useful evidence. Public E4 scores regenerate all 15 tests; missing manifests, final prompts, and profiler logs bound stronger reproducibility claims.
Source:https://arxiv.org/abs/2611.21626PreviousEvaluating AST-Derived Knowledge Graph Retrieval for AIOps Incident Diagnosis in Apache KafkaNextUsage Patterns of Programming Languages and Tools: A Systematic Literature Review
Related posts
AI visibility: mentions, recommendations and citations need separate checks
0dev.to
Background AI Makes Inference a Scheduling Problem
0dev.to

DeepSeek vs Claude vs Kimi vs Codex vs Gemini: A 70-Call AI Ethics Test
0dev.to
Fluent English is not evidence of ASL understanding: lessons from building a video translator
0dev.to

Audit Trails and Explainability for Agentic AI in Code Events
0dev.to
By 3 p.m. the daily high was already final on 78% of days. In Madrid, 32% (49 airports, one year)
0dev.to