New Taxonomy and Detection Tool for LLM Code Smells
A new paper on arXiv introduces a refined taxonomy of nine LLM code smells and a static analysis tool, SpecDetect4LLM, for their detection. The study evaluates the tool's precision and recall, and analyzes 692 open-source projects (171,194 source files), finding that 73.5% of systems exhibit LLM code smells. The work consolidates earlier research to help developers mitigate poor LLM integration practices.
Key facts
- Paper available on arXiv with ID 2605.22976
- Taxonomy includes nine LLM code smells
- SpecDetect4LLM is a static source code analysis tool
- Empirical evaluation on 692 open-source projects
- 171,194 source files analyzed
- 73.5% of systems affected by LLM code smells
- Focus on poor integration of LLM inference in source code
- Aims to document inadequate LLM integration coding practices
Entities
Institutions
- arXiv