Code-view ensemble lifts LLM vulnerability detection to 78% F1
Original title: Enhancing LLM-Based Vulnerability Detection via Representation-Aware Logical Ensemble
Source, AST, and CFG models are complementary; logical rules plus majority voting beat the best single view by 5 F1 points.
Automated detection of vulnerabilities in C source code usually relies on one code representation. The authors compare LLMs on raw source, abstract syntax trees, and control-flow graphs and find each view stronger on different vulnerability classes. Their Representation-Aware Logical Ensemble combines those models with explicit logical rules and majority voting, raising F1 from up to 73% to 78%, with similar gains in accuracy, precision, and recall.
- Single-representation models reach at most 73% F1.
- RALE reaches 78% F1 and also improves accuracy, precision, and recall.
- Representations tested are raw source, AST, and CFG.
- The ensemble uses explicit logical rules and majority voting.
From the original article
1 min read at arxiv.orgAutomated vulnerability detection in C source code is a long-standing challenge, as traditional static analysis tools often trade precision for scalability and struggle with complex vulnerability patterns. Recently, large language models (LLMs) have shown promising results; yet, most approaches rely on a single code representation, implicitly assuming uniform suitability across vulnerability types.
By Francesco Crescenzo Grasso · arxiv.org
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.