Researchers Catalog 9 LLM Code Smells, Tool Detects with 91% Precision

Large Language Models are increasingly embedded into software, but sloppy integration can wreck maintainability, security, and performance. A new paper from Université du Québec researchers tackles this with a refined taxonomy of nine LLM code smells—reusable anti-patterns like prompt leakage, hardcoded model dependencies, and hallucination-prone error handling. To make detection practical, they built SpecDetect4LLM, a static source code analyzer that scans for these smells without running the code. In a large-scale evaluation across 692 open-source projects (171,194 source files), the tool achieved 91.3% precision and 71.8% recall, proving its real-world utility.

The study also revealed that 73.5% of analyzed systems contained at least one LLM code smell, suggesting the problem is widespread across the ecosystem. The authors argue that developers lack structured guidance for integrating LLMs—unlike decades of best practices for traditional code. This taxonomy and detection tool aim to fill that gap, much like how “code smells” for object-oriented programming improved software quality. For teams using LLM APIs, this research offers a concrete checklist to audit their codebases and avoid common pitfalls that lead to brittle or insecure AI-powered features.