Neuro Symbolic Automated Program Repair: A Systematic Review of LLM-Based and Symbolic Techniques
Abstract
Automated program repair (APR) involves creating patches for software defects with minimal human intervention. Classical template-based and search-based methods are not scalable, whereas large language models (LLMs) provide good generalization but are affected by hallucinations and have weak formal guarantees. Neuro-symbolic program repair (NSPR) integrates LLMs with static analysis, SMT solving, and symbolic techniques to trade off coverage, correctness, and interpretability. We performed a systematic review of APR and NSPR systems using the PRISMA 2020 guidelines. Predefined queries in IEEE Xplore, ACM Digital Library, Scopus, Web of Science, arXiv, and Google Scholar were used to search between January 2012 and January 2026. Architectures, benchmarks, outcomes, and deployments. Two reviewers screened the records against established predetermined eligibility criteria. A total of 70 pieces of academic primary empirical research on APR/NSPR and eight (8) pieces of industrial/production deployment reports were included, and 78 papers were ultimately included in the qualitative synthesis. Our results indicate that, compared to neural or symbolic systems, NSPR systems also tend to report higher correct repair rates and lower hallucination rates. It can be deployed in CI/CD pipelines, security patches, and large-scale open-source repositories. Nevertheless, standard benchmarks and transparent assessments may still be required to perform powerful comparisons and ensure reproducibility.
