Prompt-Oriented Code Understanding: Towards Natural Language-Driven Debugging in IDEs

The increasing complexity of modern software sys- tems has rendered traditional debugging methods, such as manual code inspection and breakpoint analysis, progressively inefficient. This paper introduces Prompt-Oriented Code Under- standing (POCU), a novel paradigm that leverages Generative Artificial Intelligence (GenAI) to enable natural language-driven debugging directly within Integrated Development Environments (IDEs). We propose a system architecture that translates devel- oper queries in natural language into actionable code analysis and debugging operations. To facilitate continuous model improve- ment without compromising intellectual property, we introduce a Trust-Metric Federated Learning (TMFL) framework. TMFL allows the underlying GenAI model to be fine-tuned across disparate, private codebases (silos) while ensuring the integrity and accountability of contributions through a novel trust metric. Furthermore, acknowledging the critical need for transparency in AI-assisted tools, we present a framework to quantify and optimize the inherent trade-off between the explainability of the AI’s suggestions and its raw performance. We define metrics for both dimensions and formulate an optimization strategy to achieve a Pareto-optimal balance. Our conceptual framework and proposed methodologies lay the groundwork for a new generation of intelligent, intuitive, and trustworthy developer tools designed to significantly reduce debugging time and cognitive load.