Practical, Learning-Based Tools for Finding and Fixing Bugs

Informazioni relative al progetto

BugGPT

ID dell’accordo di sovvenzione: 101155832

DOI

10.3030/101155832

Progetto chiuso

Data della firma CE 26 Gennaio 2024

Data di avvio 1 Giugno 2024

Data di completamento 30 Novembre 2025

Finanziato da

• European Research Council (ERC)

Costo totale

Nessun dato

Contributo UE

€ 150 000,00

Coordinato da

UNIVERSITY OF STUTTGART
Germany

Periodic Reporting for period 1 - BugGPT (Practical, Learning-Based Tools for Finding and Fixing Bugs)

Periodo di rendicontazione: 2024-06-01 al 2025-11-30

Sintesi del contesto e degli obiettivi generali del progetto

The proposal's main objective is to convert existing AI-based research prototypes for bug detection and repair into practical, deployable tools integrated into real software development workflows. It focuses on building a continuous-integration–based system that analyzes code changes and test failures to automatically propose bug fixes at moments when developers are already engaged with the affected code. The project aims to leverage large language models and learning-based program repair techniques to generate accurate fixes and accompany them with concise, human-readable explanations that clarify the underlying bug and the benefit of the proposed solution. By combining automated bug detection, repair generation, and explanation synthesis into a unified pipeline, the project seeks to create robust, scalable prototypes that can operate on real-world software systems and be refined through empirical evaluation and developer feedback.

Lavoro eseguito dall’inizio del progetto fino alla fine del periodo coperto dalla relazione e principali risultati finora ottenuti

We developed several new techniques to detect bugs and suggest strategies for fixing them. Several of these techniques apply directly at the point of code change, helping developers avoid introducing bugs in the first place. Most notably, Testora and ChaCo (both to be presented at ICSE’26) analyze code changes provided in pull requests. Testora cross-checks the intent of a code change, as expressed in natural language, against the actual behavior of the changed code. ChaCo fills gaps in coverage by generating new tests that target the changed code and then executes those tests to validate the change. To help developers fix bugs, we developed RepairAgent, an autonomous agent that uses large language models to suggest repairs for buggy code. At the time of its initial publication (March 2024), RepairAgent was the first LLM agent for this task. Since then, many other approaches have been proposed, and our work has already been cited over 260 times. We also identified a key challenge that, if not addressed, may prevent software developers from adopting AI agents in their workflow: trust. Our CACM Communications article on this topic describes the problem and suggests strategies for building trust in AI agents.

Progressi oltre lo stato dell’arte e potenziale impatto previsto (incluso l’impatto socioeconomico e le implicazioni sociali più ampie del progetto fino ad ora)

As a next step, this line of work can be extended toward more practical, end-to-end tooling that seamlessly fits into industrial development workflows. In particular, future efforts should focus on fully integrating automated bug detection and repair into continuous integration pipelines, enabling repairs to be suggested, validated, and, when appropriate, applied automatically as part of routine code review and testing processes. This includes improving scalability, reliability, and configurability to support large and complex codebases, as well as strengthening validation mechanisms to ensure that suggested fixes preserve intended behavior. Further development of user interfaces, feedback mechanisms, and customization options will also be essential to increase adoption and trust, ultimately enabling AI-based repair systems to function as dependable, long-term collaborators in software development.