Calibrating Trustworthiness: Co-Designing Metrics and Visualizations for Evaluating LLMs in Education

Large language models (LLMs) are increasingly integrated into educational tools, yet evaluating their responses for pedagogical alignment remains a significant challenge. Learning engineers often struggle to identify subtle pedagogical disruptions—such as biased feedback or misinformation—within the vast outputs of these black-box models. This study addresses this gap by treating trustworthiness as a structured evaluation lens. Through a longitudinal co-design process with learning engineers developing an LLM-powered digital textbook, the researchers adapted machine learning trustworthiness criteria into 20 pedagogical measures, organized into five core metrics. They then developed visualization tools to map these violations directly onto LLM responses, allowing for side-by-side A/B comparisons.

Integrating trustworthiness metrics and visualizations significantly improved the evaluation process. By making potential risks visible, the tools increased inter-rater reliability among experts, who were previously prone to subjective disagreement. The visualizations were particularly effective at helping engineers reconcile conflicting objectives—such as balancing helpfulness with accuracy—by providing a clear, evidence-based view of where a model failed to meet pedagogical standards. The study demonstrates that trustworthiness cannot be reduced to a fully automated score; instead, the most effective approach is a hybrid, human-centered model where metrics provide structured guidance while experts make the final, nuanced judgments.

As LLMs become standard components of digital learning environments, the ability to ensure they are pedagogically sound is critical. This research provides a practical framework for moving beyond ad-hoc evaluation, offering design guidelines for future tools that prioritize transparency and expert reflection. By enabling learning engineers to trace failures to specific parts of an LLM response, these tools help mitigate risks before they reach students, ultimately fostering more reliable and effective AI-powered education.