Benyamin Tabarsi

The variety of help that large language models (LLMs) provide has made them popular among students across fields. Computing education has been particularly affected, as LLMs can handle coding tasks effectively and provide feedback, raising hopes for supporting students while creating concerns about learning and academic integrity. Researchers have responded by developing customized tools that leverage LLMs' potential while mitigating these risks. Despite a rapidly growing body of empirical studies, no prior review has focused specifically on these purpose-built tools, where educators' pedagogical and technical design decisions are explicitly represented. We address this gap by reviewing LLM-driven, student-facing tools in undergraduate computing education, examining their pedagogical roles, technical design, evaluation methods, and impacts on learners, and discussing what they imply for the broader educational system. Drawing on PRISMA guidelines, we searched three major databases (ACM, IEEE, and ScienceDirect), conducted rigorous screening, and analyzed 52 papers. Technically, prompt engineering and multi-stage pipelines are the dominant approaches, with guardrails and prompt chaining also widely adopted, while more demanding methods such as fine-tuning remain rare. Pedagogically, most tools position the LLM as a Mentor who gives feedback on student work or a Tutor who delivers instruction, whereas roles that support metacognition, learning-by-teaching, and especially teamwork are very rare. Evaluation relies heavily on surveys and interaction logs, capturing student perceptions and usage far more often than learning outcomes, with little attention to long-term effects. Across studies, impacts on engagement and student perception are generally positive, whereas evidence for gains in learning and performance is more mixed. We argue that future work should prioritize longitudinal and learning-focused evaluation, personalization for different learners, and collaborative sensemaking and teamwork. This review offers researchers, developers, and educators a roadmap for designing and studying the next generation of AI-enhanced learning tools.

Large language models (LLMs) are rapidly reshaping software development, but their impact across the full software development lifecycle is underexplored. Existing work tends to focus on isolated activities such as code generation or testing, leaving open questions about how LLMs affect developers, processes, products, and the broader software ecosystem. We address this gap through semi-structured interviews with sixteen early-adopter software professionals who integrated LLM-based tools into their day-to-day work in early to mid-2023. We treat these interviews as early empirical evidence and compare participants’ accounts with recent work on LLMs in software engineering, noting which early patterns persist or shift. Using thematic analysis, we organize our findings around four dimensions: people, process, product, and society. Developers reported substantial productivity gains from reducing mundane tasks, streamlining search, and accelerating debugging, but also described a productivity-quality paradox: they frequently discarded generated code and shifted effort from writing code to critically evaluating and integrating it. LLM use was highly phase-dependent, with strong uptake in implementation and debugging but limited influence on requirements gathering and collaborative work. Participants developed new competencies to use LLMs effectively, including prompt engineering strategies, multi-layered verification, and security-conscious integration to protect proprietary data. They also anticipated changes in hiring expectations, team practices, and computing education, while emphasizing that human judgment and foundational software engineering skills remain essential. Our findings, consistent with evidence from large-scale studies, offer actionable implications for developers, organizations, educators, and tool designers seeking to integrate LLMs responsibly into professional software practice.

The variety of help that large language models (LLMs) provide has made them popular among students across fields. Computing education has been particularly affected, as LLMs can handle coding tasks effectively and provide feedback. This has raised hopes for supporting students, while creating concerns about learning and academic integrity. Researchers have responded by developing tools that leverage LLMs' potential while mitigating risks. Despite growing empirical studies on LLM-driven tools, there is no comprehensive study examining these tools in computing education, critical for understanding the future of learning systems. We review LLM-driven, student-facing tools in undergraduate computing education, examining their pedagogical approaches, technical design, evaluation approaches, and educational impacts. We also discuss how these tools impact the broader educational system, including institutional and pedagogical structures. Following PRISMA guidelines, we systematically searched three major libraries, conducted rigorous screening, and analyzed 52 papers. Our findings reveal that prompt engineering and multi-stage pipelines are the dominant technical approaches, with guardrails and prompt chaining also widely adopted. Pedagogically, most systems provide scaffolding, problem-based learning, and direct instruction. For evaluation, researchers commonly relied on student surveys and interaction logs, with limited assessment of long-term learning outcomes. Evidence shows generally positive impacts on student performance, efficiency, and engagement, although significant challenges remain in delivering adequate scaffolding, preventing over-reliance, and supporting the development of meta-cognitive and problem-solving skills. Future work should focus on longitudinal outcomes, systematic evaluation of learning effectiveness, and the integration of established pedagogical frameworks. This review provides researchers, developers, and educators with a roadmap for designing and studying the next generation of AI-enhanced learning tools.

Large Language Models (LLMs) have shown promise in educational applications, but challenges such as hallucinations, lack of contextual relevance, and limited personalization impede their practical adoption. To address these issues, my research introduces MerryQuery, an LLM-powered educational agent that integrates Retrieval-Augmented Generation (RAG), rule-based content control, and Reinforcement Learning from Human Feedback (RLHF). The system features a dynamic learning profile module for adaptive personalization and a multi-step verification framework that cross-checks responses against external sources to enhance trustworthiness. A functional prototype of MerryQuery is being piloted in a real-world classroom. Preliminary results demonstrate improved response reliability and student understanding.

Background and Context. The importance of CS to 21st-century life and work has made it important to find ways to integrate learning CS and programming into the regular school day. However, learning CS is difficult, so teachers integrating programming need effective strategies to scaffold the learning. In this study, we analyze students’ log data and apply a novel technique to compare Parsons Problems with from-scratch programming in a middle school science class. Objectives. Our research questions aimed to investigate whether, how, and when Parsons Problems improve learning efficiency for a programming exercise within science, utilizing log data analysis and an automated progress detector (SPD). Method. We conducted a study on 199 students in a 6th-grade science course, divided into two groups: one engaged with Parsons problems, and the other, a control group, worked on the same programming task without scaffolding. Then, we analyzed differences in performance and coding characteristics between the groups. We also adopted an innovative application of SPD to gain a better understanding of how and when Parsons problems helped students make more progress on the coding task, with an objective measure of final student grades. Findings. The experimental group, with scaffolding through Parsons Problems, achieved significantly higher grades, spent significantly less time programming, and toggled less between block category tabs. Interestingly, they ran their code more frequently compared to the control group. The SPD analysis revealed that the experimental group made significantly higher progress in all four quartiles of their coding time. Implications. Our findings suggest that Parsons problems can improve learning efficiency by enhancing novices’ learning experience without negatively impacting their performance or grades, which is especially important when programming is integrated into K12 courses.

Affective state, referring to an individual's feeling, can impact students' confidence and retention in CS, particularly for novice programmers. However, little research has been conducted to examine how moments that occur during programming impact students' affective states in real-time. In this pilot study, seven undergraduate students in an introductory block-based programming course completed a programming assignment and were surveyed and interviewed about their experience and self-efficacy as programmers. While programming, students periodically recorded their affective states via a popup in the programming environment. We performed retrospective think-aloud interviews with students afterward, asking them to watch and reflect on recordings of their programming. We subsequently analyzed student interviews using thematic analysis to derive 206 codes. These codes were grouped into three areas that impacted affect: the environment, objective progress, and perceptions during programming. To explore why students responded as they did to moment occurrence, we further categorized students based on four dimensions: programming experience, assignment completion, confidence, and the impact of the programming session on self-efficacy. Our initial results suggest that while certain moments elicit similar affective states among students, the interaction of the aforementioned four dimensions may have a higher impact on novices' affective states during programming. We conclude with recommendations for educators to improve students' affective states during and after programming.