Semantic Patch Learning for Vulnerability Repair in LLM-Generated Code

Read full text Summarise with VIRAT
Authors: Elina Korhonen, Jukka Lehtola, Mikael Saarinen
Cited by: 0
References: 22

Summary

No summary yet. VIRAT can read this paper and write a plain-language summary of its aim, methods, key findings and limitations.

Summarise with VIRAT Free to try with your Science Catalogs account.

Abstract

We introduce a semantic patch learning model that generalizes repair templates from historical CVE patches. Extracted semantic patterns guide automatic correction of insecure LLM outputs. Using 28,000 real vulnerability patches, the model learns transformations for memory bugs, crypto misuse, and authorization bypasses. On a benchmark with 900 insecure LLM-generated samples, the repair tool fixes 68% of vulnerabilities automatically and partially mitigates another 21%. This work demonstrates that semantic patch abstraction is a powerful mechanism for improving LLM code security.

Fields of Study & Subjects

Software Engineering Research Software Testing and Debugging Techniques Security and Verification in Computing

Publication details

Journal / source
Global Media and Social Sciences Research Journal
Volume
7
Issue
1
Pages
161–167
Publication date
15 February 2026
Language
EN