An Explainable AI Assistant for Introductory Programming Education: Improving Feedback Reliability with Instructor-AI Collaboration

A research team has developed an AI classroom assistant that delivers instructor-authored feedback on student programming assignments, aiming to address the shortage of timely, personalized support in introductory courses. The system, described in a paper submitted to arXiv on 12 May 2026, uses an explainable AI model to analyze student code and map logical errors to misconceptions that instructors have previously identified [1][2]. It then surfaces feedback written by the instructor, a design choice the authors argue grounds the tool's reliability in established pedagogical knowledge rather than in a language model's unaided generation [2]. Active learning is widely recognized as effective for introductory programming, but limited instructional staffing often means students wait days for feedback on foundational concepts [2]. Recent advances in large language models have created scalable opportunities for automated feedback, yet concerns about explainability and reliability have slowed classroom adoption [2]. Large language models are machine learning systems with many parameters, trained on vast text corpora for tasks such as language generation [8]. The new assistant attempts to sidestep those concerns by keeping the instructor in the loop: the AI diagnoses the error, but the response text is instructor-authored [2]. The researchers conducted an expert evaluation to measure alignment with instructor-verified feedback and deployed the system in a classroom to gauge student perceptions of usability [2]. Results indicate the assistant can provide accurate, instructor-verified feedback while fostering a positive experience, according to the paper [2]. No timeline for wider release was included in the preprint. The paper appears on arXiv, a repository that has integrated with the Hugging Face Hub to link machine learning papers with interactive demos, models, and datasets [4][5]. Since November 2022, arXiv abstract pages have included a Demos tab that surfaces community-built Spaces, allowing readers to try models without writing code [5]. Authors can link a Space to a paper by including the arXiv URL in a repository's README file or by associating a model on the Hub with the paper [6]. The Hub extracts the arXiv ID and adds it to the repository's tags, enabling filtering across other models and datasets that cite the same work [4]. The assistant's release comes amid broader industry attention to the cost and accessibility of AI systems. Chinese firm DeepSeek, founded in July 2023, reported training its V3 model for US$6 million, a figure far below the estimated US$100 million cost of OpenAI's GPT-4 [7]. DeepSeek's models are released under open-source licenses such as the MIT License, though training data is not openly licensed [7]. Alibaba Cloud's Qwen family of large language models is distributed under Apache 2.0 and other licenses [9].