Large Language Models (LLMs) have sparked new possibilities in adaptive educational tools, as exemplified by iScore, a system designed to score written summaries with the help of LLMs. This tool not only enhances scoring accuracy but also provides insights into LLM interpretability.

• iScore showcases an interactive visual analytics platform that scales LLM interpretability methods.
• Allows learning engineers to track and visualize model weights at various abstraction levels.
• Users can iteratively revise language in summaries, monitoring the impact on LLM-based scores.
• A month-long deployment results in an improved LLM score accuracy by three percentage points.
• Qualitative feedback highlights iScore’s role in understanding and building trust in LLMs.

The practical application and validation of iScore implies that though LLMs might come with complexity, tools like iScore make them more accessible and trustworthy for critical applications in education.