Is LLM-Generated Code More Maintainable & Reliable Than Human-Written Code?
Bibliographic record
Abstract
Background: The rise of Large Language Models (LLMs) in software development has opened new possibilities for code generation. Despite the widespread use of this technology, it remains unclear how well LLMs generate code solutions in terms of software quality and how they compare to humanwritten code. Aims: This study compares the internal quality attributes of LLM-generated and human-written code. Method: Our empirical study integrates datasets of coding tasks, three LLM configurations (zero-shot, few-shot, and fine-tuning), and SonarQube to assess software quality. The dataset comprises Python code solutions across three difficulty levels: introductory, interview, and competition. We analyzed key code quality metrics, including maintainability and reliability, and the estimated effort required to resolve code issues. Results: Our analysis shows that LLM-generated code has fewer bugs and requires less effort to fix them overall. Interestingly, fine-tuned models reduced the prevalence of high-severity issues, such as blocker and critical bugs, and shifted them to lower-severity categories, but decreased the model's performance. In competition-level problems, the LLM solutions sometimes introduce structural issues that are not present in human-written code. Conclusion: Our findings provide valuable insights into the quality of LLM-generated code; however, the introduction of critical issues in more complex scenarios highlights the need for a systematic evaluation and validation of LLM solutions. Our work deepens the understanding of the strengths and limitations of LLMs for code generation.
Fetched live from OpenAlex and de-inverted. Abstracts are not stored in this database: the inverted indexes are 8.6 GB of the frame’s 9.3 GB of text, and the host has 13 GB free.
How this classification was reachedexpand
Full frame machine prediction
Teacher imitationNot calibrated prevalence, not ground truth. Human validation pending. The Gemma side is a direct model label for every work in the frame, read from the title-only record. The Codex side is a classifier learned from the 10,348 direct Codex labels and calibrated to design-weighted sample rates; fields without enough sample support carry no Codex call. Candidate is the union of the two sides; consensus is their intersection. These outputs are machine_predicted_unvalidated and are not human labels.
Distilled classifier scores by category (both heads)
| Category | Codex | Gemma |
|---|---|---|
| Metaresearch | 0.012 | 0.129 |
| Meta-epidemiology (narrow) | 0.000 | 0.000 |
| Meta-epidemiology (broad) | 0.000 | 0.000 |
| Bibliometrics | 0.002 | 0.001 |
| Science and technology studies | 0.000 | 0.002 |
| Scholarly communication | 0.003 | 0.002 |
| Open science | 0.001 | 0.002 |
| Research integrity | 0.001 | 0.001 |
| Insufficient payload (model declined to judge) | 0.002 | 0.001 |
Machine scores (provisional)
The two teacher heads of the student model, read on this work. A score orders the frame for review; it never asserts a category, and the validation status ships verbatim with every row.
Baseline scores from an immature model (maturity gate not passed, 7 training rounds). Scores rank; they never assert a category.
score_only:v0-immature-baseline · verbatim from the scoring run: score_only means the number may rank works, and no category label ships from itClassification
machine, unvalidatedMachine predicted; a candidate call from one source (direct Gemma or distilled Codex), not a consensus.
How this classification was reached, model by model and score by score, is at the end of the page under "How this classification was reached".