1 of 16

Explainable AI in the Loop: An Instructor-Transformer

Collaboration for Improving Explainability and Reliability

of Feedback in Introductory Programming Classrooms

Muntasir Hoq, Bradford Mott, Seung Lee, Jessica Vandenberg, Narges Norouzi, James Lester, Bita Akram

​

9th CSEDM Workshop at EDM 2025

2 of 16

Introduction

  • Adaptive and timely feedback in CS education is imperative for effective active learning [McConnell 1996, Denny et al. 2024, Tang et al. 2024].
  • Recent advancements in LLMs demonstrated their potential in feedback generation. [Phung et al. 2023, Jia et al. 2024, Xu et al. 2024]
  • Challenges related to reliable and explainable feedback generation still persist. [Phung et al. 2023, Jacobs et al. 2024, Jia et al. 2024]

3 of 16

Our Approach

​

  • Reliable and explainable feedback system.
  • Collaboration of Instructor-LLM-Explainable AI model.
  • Providing instructor-verified programming feedback based on key instructional points.

Instructor

Explainable AI

LLM

INSIGHT

4 of 16

INSIGHT Classroom Assistant System

An intelligent classroom assistant to support scalable, reliable and explainable feedback delivery system.

  • Instructor App
    • Instructor dashboard enables real-time distribution of exercises categorized by topics.
    • Authoring tool enables instructors to design new exercises collaborating with LLMs.
  • Student App
    • Enables real-time classroom engagement of students.
    • Allows students to submit assignments and get feedback.

5 of 16

Instructor Authoring Tool

6 of 16

Instructor Authoring Tool

7 of 16

Instructor Authoring Tool

8 of 16

Student App

9 of 16

Methodology

  • Instructor designs a new problem using the authoring tool and identifies important and common misconceptions in example solutions with associated feedback.
  • Pretrained SANN model fine-tuned on synthetic data and subtrees with logical errors extracted for incorrect code.
  • Instructor-verified feedback propagated when student code subtree with logical error matches with instructor provided code subtree, including an LLM verification layer.

10 of 16

Subtree-based Attention Neural Network (SANN) [Hoq. et al. 2023]

  • Trained on code correctness prediction.
  • Subtrees with higher attention contributes to model predicting a code as incorrect and contain logical errors.
  • [Hoq et al. 2025]

11 of 16

Feedback Propagation Pipeline

12 of 16

Code Correctness Prediction on Unseen Problem

Model

Accuracy

Precision

Recall

F1-score

Non-pretrained SANN

0.83

0.81

0.76

0.78

Mixed Fine-tuned SANN (1% pre training data+synthetic data)

0.84

0.83

0.77

0.80

Fine-tuned SANN (Synthetic Only)

0.89

0.89

0.84

0.86

13 of 16

Feedback Selection Performance

​

  • Correct Feedback: 56.8%
    • Correlates highly with instructor provided errors and feedback.
  • Incorrect Feedback: 28.4%
    • Includes mostly student misinterpretation or outside of instructor-authored example coverage.
  • No Feedback: 14.8%
    • Similar as incorrect feedback category.

14 of 16

Contribution

​

  • INSIGHT classroom assistant to promote active learning in CS classrooms.
  • Automated feedback system with instructor-LLM-explainable AI collaboration.
  • Step toward a reliable, explainable and instructor-verified feedback propagation system.
  • This approach allows instructors to select their preferred balance between reliability and coverage.
  • Feedback propagation focusing on key instructional points identified by instructors.

15 of 16

Limitations

Future Work

  • High incorrect feedback propagation.
  • Handling multiple errors in one code.
  • Coverage of instructor-generated solutions and synthetic data.
  • Handling the coverage issue.
  • Handling multiple error detection and composite feedback generation.
  • Investigating sequential, step-by-step, and immediate feedback supporting cognitive load management [Anderson et al. 1989].
  • Conducting classroom studies to measure student learning gains.

16 of 16

Thank You!