Selected papers and projects representing recent directions in the lab’s research.
Evaluating students’ natural-language explanations of code is challenging for real-time tutoring feedback. This paper fine-tunes open-source LLMs to automatically assess line-by-line Java code explanations, achieving strong correlations with human judgments and outperforming few-shot prompting baselines.
This work examines how well LLMs can identify and explain a program’s logical steps and corresponding code blocks in well-structured tasks. Similarity between LLM and expert annotations reaches up to 64.4% under multiple alignment strategies.
We study how intro-to-programming students prompt LLMs for help with code comprehension without training on advanced prompting. LLM explanations are generally accurate and complete; students use three main prompt types, and LLM access is associated with increased confidence and improved performance.
When are LLMs bad for learning? In our view, LLMs are bad for learning when they eliminate productive struggle. Unrestricted access to LLMs is fundamentally at odds with effective learning because they do too much for the learner.
Using pre-trained language models to automatically assess how well students explain code, providing insights into their understanding.
Investigating the capability of large language models to detect gaps and misconceptions in student explanations of code.
Examining how LLMs can generate detailed line-by-line code explanations for conversational tutoring applications.
Exploring the impact and applications of large language models in the context of intelligent tutoring systems.
Developing methods for automated essay scoring that leverage external discourse knowledge.
Using large language models to identify and analyze gaps in student code explanations.
A comparative study of reading materials versus interactive tutoring for improving novice programmers' code comprehension.
Leveraging large language models for automated assessment of student understanding of code.
Comparing the quality of code explanations generated by large language models versus human instructors.
Analyzing how large language models respond to prompts for generating code explanations.
Introduces SelfCode 2.0, a corpus of 3,019 student-expert explanation pairs for Java snippets, with expert labels for semantic similarity, correctness, and completeness to train models that assess learners' self-explanations. Includes baseline results using TF-IDF and Sentence-BERT toward stronger automated assessment in programming education.