Constrained LLM-Guided Refactoring of JavaScript: A Smell-Targeted Transformation Framework with Human-in-the-Loop Validation
Emir Kuanyshev, Hashim Ali*
Department of Computer Science, Nazarbayev University, Astana, Kazakhstan
* Corresponding Author: Hashim Ali. Email: 
Computers, Materials & Continua https://doi.org/10.32604/cmc.2026.080737
Received 13 February 2026; Accepted 08 April 2026; Published online 23 April 2026
Abstract
Refactoring improves maintainability without altering externally observable behavior, yet it remains costly and error-prone when applied manually at scale. While large language models (LLMs) can generate plausible refactorings, practical adoption is limited by uncontrolled edit scope, inconsistent outputs under stochastic decoding, and weak traceability of why a change was produced. This paper proposes a smell-targeted, scope-bound refactoring framework for JavaScript that couples deterministic AST-based smell detection with constrained LLM transformation. The key design principle is to bind generation to explicitly detected smell instances, enforce a structured output contract (refactored code plus per-smell rationale), and log full refactoring artifacts for repeatable evaluation. We implement the framework as a microservice-based prototype (detector, prompt construction and routing, orchestrator, analytics, and UI) and evaluate it on LeetCode-style solutions and multiple GitHub repositories. Across the evaluated projects, the approach achieves an average smell reduction of 83.96% and an average maintainability index improvement of +5.366, while maintaining a mean developer acceptance rate of 91.66%. A targeted temperature study identifies an operating point around 0.4 that maximizes acceptance (approximately 95% in controlled trials), balancing determinism with sufficient flexibility for structure-improving edits. These results suggest that explicit scope control and structured traceability are central to making LLM-based refactoring reliable and reviewable, and motivate future integration with automated validation (tests, linting) and repository-conditioned policies.
Keywords
Large language models; automated refactoring; code smell detection; JavaScript; constrained generation; maintainability index; human-in-the-loop evaluation
Open Access
Login
Register
Submit a Paper
Propose a Special lssue
Download PDF
Downloads
Citation Tools
