Can LLM Code Explanations Adapt to Diverse Problem-Solvers’ Needs?

This study investigates whether Large Language Models (LLMs) can tailor their code explanations to match the diverse problem-solving styles of developers. Drawing on the Gender Inclusiveness Magnifier (GenderMag) method, the researchers identified five key dimensions of problem-solving: learning style, self-efficacy, risk attitude, motivation, and information processing style. Each dimension has two extreme endpoints (e.g., process-oriented vs. tinkering-oriented learners). The authors developed 10 specific prompts based on these styles and generated 1,072 code explanations for COBOL programs across six open-weight LLMs.

Using natural language processing techniques, the researchers uncovered a taxonomy of 13 distinct linguistic adaptations that LLMs use to align with these styles. These strategies include behaviors such as explaining steps sequentially, inviting follow-up questions, normalizing knowledge gaps, and adjusting the level of conciseness. The study found that models do not treat all users the same; they systematically increase or decrease the frequency of these linguistic markers based on the specific problem-solving style requested in the prompt.

Software engineering is a highly diverse field, yet most AI coding assistants provide a "one-size-fits-all" explanation. By demonstrating that LLMs can be prompted to adapt to different cognitive styles, this research provides a roadmap for building more inclusive and effective developer tools. If an LLM can recognize that a user prefers a structured, process-oriented approach versus a hands-on, tinkering approach, it can adjust its output to reduce frustration, improve productivity, and lower the barrier to entry for developers with different backgrounds.