Can LLMs Refactor Their Own Smelly Code? Self-Refactoring Effectiveness and Behavioral Preservation Across Four Open-Source Models

(1) Background: LLM-generated code frequently contains code smells that degrade maintainability. Whether the generating model can effectively refactor its own output—self-refactoring—remains untested. (2) Methods: We construct a taxonomy of six LLM-specific Python 3.12 code smells (F1 = 0.87) and test self-refactoring across 1404 real-inference smelly snippets from four open-source LLMs (CodeLlama-13B, DeepSeekCoder-6.7B, StarCoder2-15B, Qwen2.5-Coder-14B) on BigCodeBench. (3) Results: Three self-refactoring profiles emerge: style entrenchment (CodeLlama, 0% median SRR), inherent capability (DeepSeekCoder, 100%), and prompt-dependent (Qwen, 100% with smell-aware prompting vs. 0% zero-shot). Because one smell (Narrative Comment) dominates, per-smell-type results are foregrounded and StarCoder2 (n=14) is preliminary. Unit tests confirm a 96.0% Behavioral Preservation Rate on test-passing snippets (95% CI 93.3–97.7%). A direct cross-refactoring experiment (356 runs, four independent refactorers) falsifies the Style Coherence Hypothesis: self-refactoring never beats the best independent model (0/89 snippets), and CodeLlama’s own code—which it cleans in 1/25 cases—is cleaned in 25/25 by other models (p=4.8×10−7). A generation–refactoring feedback loop, strong in simulation, is likewise null under real inference. (4) Conclusions: Self-refactoring effectiveness depends on the interaction of model, smell type, and prompting strategy. Smell-aware prompting enables lightweight, behavior-preserving post-generation quality pipelines without extra models or retraining; the closed-loop feedback variant, by contrast, does not transfer from simulation to real inference.