GPT-5 scores 89.1% on Scrum certification questions with citation prompting

A new empirical study by Mirko Perkusich and colleagues evaluated GPT-5's accuracy on 993 validated Professional Scrum Master (PSM) certification questions. Using three prompting techniques—zero-shot, chain-of-thought, and with-source citation—the researchers found that all methods achieved certification-level accuracy above 85%, with the citation-based variant performing best at 89.1% and yielding the lowest error rate. The model excelled on well-defined topics like Definition of Done, Events, and Product Backlog Management, particularly on single-answer multiple-choice items. However, multi-select questions and more interpretive areas such as Scrum Team and Product Value proved less stable.

Among the 16.2% of questions where at least one prompt failed, errors clustered into three categories: misalignment with the Scrum Guide (28%), content falling outside its scope (34%), and outdated or biased interpretations (38%). The study underscores that prompt engineering—especially citation-based prompting—can modestly but consistently improve factual accuracy, reduce misinterpretation, and minimize version drift in LLM responses. This suggests that with proper technique, LLMs like GPT-5 can serve as reliable tools for Agile learning and certification preparation, though practitioners should remain aware of edge cases and context-dependent failures.