VERIGATE: A VERIFICATION GATE-BASED NEUROSYMBOLIC METHOD FOR STATIC CODE ANALYSIS

Read full text Summarise with VIRAT
Authors: Illia Vokhranov, Bogdan Bulakh
Cited by: 0
References: 19

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

Static code analysis is essential for software quality and security, yet existing approaches face fundamental limitations.Traditional rule-based tools provide deterministic results but struggle with complex semantic defects, while Large Language Models (LLMs) offer powerful semantic reasoning but lack controllability and verifiable guarantees.The black-box nature of LLM outputs provides no mechanism to validate whether detected issues correspond to actual structural properties of the analyzed code, hindering integration into production CI/CD pipelines.We present VERIGATE, a neurosymbolic method that addresses these limitations through a structured Explain-Verify gate mechanism.The method's scientific novelty comprises three interrelated aspects: (1) an inverted neurosymbolic interaction direction (Neural → Symbolic), where the LLM generates defect hypotheses that symbolic analysis independently validates; (2) the Explain-Verify gate mechanism that enforces structured claim generation and subjects claims to independent symbolic validation; and (3) Evidence-Weighted Suppression for balanced final decision-making.In the Explain phase, the LLM generates structured claims specifying defect type, affected code spans, a minimal corrective fix, and supporting rationale.In the Verify phase, these claims are independently verified using symbolic program analysis techniques: AST pattern matching, taint analysis, and fix plausibility checking.This architecture grounds LLM-based detection in verifiable program properties, ensuring predictable and controllable behavior.Evaluation on 1,000 samples from the PySStuBs dataset with GPT-4 demonstrates substantial improvements: 76% higher F1-score (0.612 vs. 0.347), 187% higher recall (0.758 vs. 0.264), and 247 additional vulnerabilities detected (379 vs. 132) with comparable precision.Ablation analysis confirms that the verification gate provides controllability and interpretability with minimal performance cost (1.5% F1 № 2(56) 2026

Fields of Study & Subjects

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

Publication details

Journal / source
Наука і техніка сьогодні
Issue
2(56)
Publication date
21 March 2026
Language
EN