A systematic review posted on arXiv examines why large language models produce defective code. The review notes that LLMs frequently generate code with logical bugs or security vulnerabilities, making output quality a persistent concern.
The authors identify training-data quality as one contributor to these defects. They also point to model architecture and optimization as additional factors, meaning data quality is not the sole cause of poor code generation.
The review bridges generation and training, suggesting that improving code output requires attention to both how models are trained and how they generate code. The abstract does not provide further details on methodology or findings, but the framing points to a need for holistic quality improvements.