Knowledge-Component Guided Generation of Personalized Worked Examples from Student Code

Overview

Field: Machine Learning / AI in Education Authors: Griffin Pitts, Muntasir Hoq, Peter Brusilovsky Published: 2025-04-29 arXiv: 2504.20651

Summary

This paper introduces a knowledge-component (KC) guided approach for generating personalized educational content in programming practice environments.

Problem

Adaptive programming practice typically depends on fixed libraries of worked examples and practice problems. These libraries require substantial authoring effort and often fail to correspond to the logical errors and partial solutions students actually produce while writing code. Consequently, students may receive content that does not directly address the concepts they are trying to understand, while instructors must either expand content libraries manually or accept coarse personalization.

Approach

• Extract pattern-based knowledge components (KCs) from student code submissions
• Use AST-based analysis to identify recurring structural KC patterns in students' code
• Use these extracted KCs to condition a generative model for producing worked examples tailored to individual learners

Evaluation and Findings

The method was applied to worked example generation, and expert evaluation compared baseline generation against KC-conditioned generation. Results indicate that KC-conditioned generation improves:

• Topic focus of generated examples
• Relevance to learners' underlying logical errors

This enables fine-grained personalization of programming learning content without heavy manual authoring overhead.

Links

• Paper: https://arxiv.org/abs/2504.20651