| Title |
Improving static code analysis using a consensus-based multi-LLM |
| Authors |
Labutis, Augustinas ; Čeponis, Jonas ; Grigaliūnas, Šarūnas ; Brūzgienė, Rasa |
| DOI |
10.83095/ceurws_04309_011 |
| Full Text |
|
| Is Part of |
CEUR workshop proceedings: TKTP 2026: proceedings of the annual doctoral symposium of computer science 2026, Turku, Finland, June 2026 / edited by E. Peltonen, O. Weerakoon, P. Puhtila, Y. Passeniouk.. Aachen : CEUR-WS. 2026, vol. 4309, p. 1-12.. ISSN 1613-0073 |
| Keywords [eng] |
static code analysis ; large language models ; consensus algorithms ; multi-agent systems ; automated program repair ; software quality ; code diagnostics |
| Abstract [eng] |
Static code analysis is a prevailing technique for enhancing software quality and security. It facilitates the identification of defects, vulnerabilities, and maintainability issues without requiring program execution, thereby ensuring the integrity and reliability of the software under scrutiny. However, its practical application is still constrained by false positives, limited semantic understanding, and reduced precision when applied to large and evolving codebases. This research proposes a consensus-based approach to improving static code analysis with the support of multiple large language models (LLMs). The proposed approach diverges from the conventional treatment of automated code repair and warning interpretation as the output of a single model. Instead, it frames these processes as a comparative process in which several independently generated repair candidates are assessed in relation to the same diagnostic issue. The approach commences with repository-level diagnostics extracted from the target codebase and transformed into structured repair tasks. Subsequently, multiple LLMs are engaged in the generation of alternative candidate fixes for the identified problem. Furthermore, the method incorporates a historical grounding mechanism in which previously faulty and corrected code pairs are compared to estimate the expected scale of code modification for a given diagnostic type. This grounding percentage functions as an empirical reference point for evaluating the proportionality and structural plausibility of generated fixes. Candidate repairs are further examined through validation criteria such as compilation outcome, diagnostic resolution, and the absence of newly introduced warnings or errors. The final result is determined through a consensus-based ranking that combines grounding consistency and convergence between repairs generated independently. This approach is intended to provide a more reliable basis for static analysis support. |
| Published |
Aachen : CEUR-WS |
| Type |
Conference paper |
| Language |
English |
| Publication date |
2026 |
| CC license |
|
