Abstract:

Formative assessment in computer engineering education faces significant scalability challenges. Evaluating complex submissions that include both documentation and source code repositories is time-consuming, often leading to delays in feedback and inconsistencies in rubric application. While artificial intelligence presents opportunities for automation, concerns regarding accuracy, pedagogical alignment, and the preservation of teacher oversight remain prevalent.

This paper presents the design, development, and pilot implementation of "GradAssist," a web-based tool that generates draft formative assessments using Large Language Models (LLMs) to assist instructors with the grading process. The tool integrates with the Google Classroom API to extract assignment submissions and their associated rubrics. It processes text from attached documents and, for programming tasks, clones and analyzes source code from GitHub repositories. An LLM generates a preliminary evaluation and drafts feedback strictly aligned with the instructor's rubric criteria. This draft is then provided to the instructor, who conducts a final review and issues the definitive evaluation to students. The system is designed as an assistive tool with the instructor retaining full control and responsibility for all final assessments.

This paper details the tool's architecture and its application during the first semester of 2026 in two undergraduate courses: Software Engineering (n=20) and Object-Oriented Programming (n=20). Observational data were collected on instructor interaction with the tool, including the frequency and nature of modifications made to AI-generated draft assessments. Student perceptions of the feedback quality and timeliness were gathered through post-hoc surveys.

Preliminary observational findings indicate that GradAssist significantly reduced the time instructors spent on the mechanical aspects of grading, allowing them to focus on higher-level pedagogical interventions. Analysis of modification logs revealed specific rubric criteria where AI performance was weaker, providing valuable data for iterative system refinement. Student feedback highlighted the value of receiving rapid, rubric-aligned draft comments. The paper discusses the practical lessons learned from integrating LLMs into an authentic assessment workflow. It concludes by outlining the implications for the evolving role of the educator and proposes directions for future controlled experiments to measure the direct impact of such tools on student learning outcomes.

Keywords:

LLM, automated, grading, rubric.