arXiv subjects

Craig Zilles

Publications and source records attributed to Craig Zilles.

5 recordsLinked to original sources

arXiv2026

Consistently Good vs. Occasionally Great: A Rubric for Open-Ended Feedback Quality from Humans and Machines

Providing high-quality feedback on student work is essential for learning, yet delivering such feedback at scale remains challenging. In this paper, we focus on feedback for open-ended short answer questions in introductory programming, with the goal of nudging students toward success on reattempts without revealing the correct answer. We develop a five-criteria rubric grounded in educational literature for evaluating feedback quality: (1) acknowledging correct portions of the student answer, (2) identifying at least one flaw (if present), (3) providing actionable guidance for improvement, (4) maintaining appropriate concealment of the answer, and (5) using an appropriate conversational tone. Using this rubric, we compare feedback generated by a frontier LLM (OpenAI o1) to feedback from nine teaching assistants across 90 student responses, with three researchers and an LLM independently scoring all feedback. Our results show that while one TA often produced the best feedback, the LLM demonstrated consistently higher average performance than TAs, as evaluated by humans. However, we also uncover significant self-preference bias when using LLMs to evaluate feedback quality: the LLM systematically rated its own outputs higher than human experts did. This bias, which research suggests persists even in cross-model evaluation, raises important methodological concerns for researchers employing LLM-based evaluation. We provide detailed characterization of both TA and LLM performance, analyze sources of variance in TA feedback quality, and discuss implications for deploying LLM-generated feedback in educational settings.

cs.CY↗

arXiv2026

Automated Grading of Handwritten Mathematics Using Vision-Capable LLMs

Automated grading systems have enabled scalable assessment for many response types, but handwritten mathematics remains a barrier due to the complexity of multi-step solutions. Vision-capable large language models (LLMs) offer new opportunities here, yet their reliability in authentic instructional settings remains poorly understood. We present an empirical evaluation of an LLM-based grader for handwritten mathematical work using instructor-defined rubrics. Extending a prior pipeline for typed responses, we integrate transcription and rubric-based evaluation of photographic submissions within a single LLM call, evaluating on student work from two university STEM courses. Comparing AI grading decisions against human-assigned ground truth at the rubric-item level, we observe high overall accuracy, with most errors -- 87\% in the best model -- attributable to transcription failures rather than rubric misapplication. We categorize common error modes, including image quality issues, hallucinated content, and incorrect handling of equivalent expressions. These findings highlight both the promise and limitations of LLM-based grading for handwritten mathematics, providing guidance for system design, prompt refinement, and deployment in educational settings.

cs.CY↗

arXiv2025

ReDefining Code Comprehension: Function Naming as a Mechanism for Evaluating Code Comprehension

"Explain in Plain English" (EiPE) questions are widely used to assess code comprehension skills but are challenging to grade automatically. Recent approaches like Code Generation Based Grading (CGBG) leverage large language models (LLMs) to generate code from student explanations and validate its equivalence to the original code using unit tests. However, this approach does not differentiate between high-level, purpose-focused responses and low-level, implementation-focused ones, limiting its effectiveness in assessing comprehension level. We propose a modified approach where students generate function names, emphasizing the function's purpose over implementation details. We evaluate this method in an introductory programming course and analyze it using Item Response Theory (IRT) to understand its effectiveness as exam items and its alignment with traditional EiPE grading standards. We also publish this work as an open source Python package for autograding EiPE questions, providing a scalable solution for adoption.

cs.CY↗

arXiv2025

Counting the Trees in the Forest: Evaluating Prompt Segmentation for Classifying Code Comprehension Level

Reading and understanding code are fundamental skills for novice programmers, and especially important with the growing prevalence of AI-generated code and the need to evaluate its accuracy and reliability. ``Explain in Plain English'' questions are a widely used approach for assessing code comprehension, but providing automated feedback, particularly on comprehension levels, is a challenging task. This paper introduces a novel method for automatically assessing the comprehension level of responses to ``Explain in Plain English'' questions. Central to this is the ability to distinguish between two response types: multi-structural, where students describe the code line-by-line, and relational, where they explain the code's overall purpose. Using a Large Language Model (LLM) to segment both the student's description and the code, we aim to determine whether the student describes each line individually (many segments) or the code as a whole (fewer segments). We evaluate this approach's effectiveness by comparing segmentation results with human classifications, achieving substantial agreement. We conclude with how this approach, which we release as an open source Python package, could be used as a formative feedback mechanism.

cs.CY↗

arXiv2023

Code Generation Based Grading: Evaluating an Auto-grading Mechanism for "Explain-in-Plain-English" Questions

Comprehending and elucidating the purpose of code is often cited as being a key learning objective within introductory programming courses. To address this objective ``Explain-in-Plain-English'' questions, in which students are shown a segment of code and asked to provide an abstract description of the code's purpose, have been adopted. However, given EiPE questions require a natural language response, they often require manual grading which is time-consuming for course staff and delays feedback for students. With the advent of large language models (LLMs) capable of generating code, responses to EiPE questions can be used to generate code segments, the correctness of which can then be easily verified using test cases. We refer to this approach as "Code Generation Based Grading" (CGBG) and in this paper we explore its agreement with human graders using EiPE responses from past exams in an introductory programming course taught in Python. Overall, we find that CGBG achieves moderate agreement with human graders with the primary area of disagreement being its leniency with respect to low-level and line-by-line descriptions of code.

cs.CY↗