AI for Learning — Daily Research

In plain terms

Students often face barriers like anxiety or fear when seeking academic help, leading to unequal access to support, especially in complex subjects like programming. While general AI tools exist, they can be inaccurate or not specific to a course. To address this, researchers developed "Beacon," a specialized AI system using Retrieval-Augmented Generation (RAG). RAG is an AI technique that grounds its responses in specific, approved teaching materials from the course, ensuring accuracy and relevance. Beacon was designed to lower the hurdles students face when seeking help and to encourage independent learning. Through surveys and interviews with students and staff, they found that students considered Beacon's responses closely aligned with module content and more trustworthy than unrestricted AI, valuing its scaffolded explanations. Participants viewed Beacon as a valuable first point of support before consulting lecturers or official resources.

In plain terms

AI tutors are becoming common, but we don't know how best to design them for student learning, especially regarding *pedagogical structure* (how the AI teaches, like guided questions) and *interaction modality* (how students talk to it, like voice or text). Researchers experimented with 86 graduate students, comparing a *structured AI tutor* designed to guide learning based on course materials against a control group with general AI access. Within the structured group, students alternated weekly between voice and text interaction with the tutor. They found that the structured AI tutor significantly boosted learning, with tutored students scoring 6.63 points higher and showing improved written reasoning. However, interaction modality (voice vs. text) had no impact on learning outcomes, despite voice being preferred and more costly to deliver. This suggests that the "how" an AI teaches (pedagogical structure) is vital for learning, not just making it seem more humanlike through voice.

In plain terms

Current AI systems can create educational videos, but they often focus only on making the video look good, neglecting how people actually learn. This results in videos that might lack a clear teaching structure, reliable quality, or ways to check if students understand the material. To address this, researchers developed PIVOT (Pedagogy-guided Instructional VideO Tutoring), a system designed to generate instructional video tutors specifically for STEM (Science, Technology, Engineering, Math) subjects. PIVOT embeds *pedagogy*, which is the art and science of teaching, into every stage of video creation. This includes guiding the initial storyboard, generating and verifying the video's content for both accuracy and teaching effectiveness, and connecting it with assessments to identify and correct student misunderstandings. Experiments and evaluations by experts across four STEM areas showed that PIVOT produced videos with content well-aligned with teaching goals, clear and engaging presentation, and a logical instructional flow, leading to perceived effectiveness for learning. This highlights that AI systems in education should be judged not just by what they create, but by how well they support teaching practices, learner understanding, and corrective feedback.

In plain terms

University programs teaching intelligent medical engineering are expanding, but their lab courses often don't keep up with the real-world complexities of AI in clinical settings, such as using multiple types of data (multimodal), handling rare cases (long-tailed), or adapting to changing data patterns (distributionally shifting). The researchers designed a new, four-level experimental teaching system. This system takes an ongoing research project on using advanced AI (multimodal deep learning) to diagnose endometrial cancer and turns it into a structured series of lab exercises for undergraduate students. They identified three key issues in current education (modality, authenticity, and deployment) and based their curriculum on established teaching frameworks like constructive alignment and experiential learning. The curriculum includes four progressive stages, plus an engineering layer, with 32 lab units spread over 64 hours. Students use a special virtual clinical workstation with real, but anonymized, multi-institutional patient data. Each stage focuses on a specific technical challenge, links to prior courses, and has clear assignments. They also outlined how data is handled, safety measures, and how students will be assessed. The plan is to collect data on student learning outcomes over two cycles to evaluate its effectiveness.

In plain terms

Spatial reasoning abilities are crucial for STEM fields, and games offer an engaging way to develop these skills in children. To make these games effective for learning, they need Artificial Intelligence (AI) that can provide human-like tutoring and guidance. This paper introduces a hybrid qualitative model designed for Camelot Jr., a block-puzzle game where players build multi-level bridges. Qualitative reasoning (QR) models use symbolic representations, allowing the AI to "think" in ways that can be easily translated into understandable feedback for players. To ensure the model accurately handles the game's physics, they integrated a mathematical center-of-mass stability logic into the qualitative solver. This work facilitates spatial skill training within the game and contributes to creating human-centric, explainable AI agents for educational purposes.

In plain terms

Writing clear 'how-to' instructions, known as procedural writing, is often challenging for learners. Traditional example-based learning, which provides feedback through examples, can be ineffective when everyone receives the same content, regardless of their individual needs. To tackle this, researchers developed RELEX, an adaptive learning system designed to personalize example-based feedback. The system first uses a fine-tuned Large Language Model (an advanced AI for text generation) to predict the quality of a learner's cooking recipe. Based on this quality score, RELEX then retrieves a higher-quality and contextually similar example recipe from a vast database, enriching it with personalized instructional explanations. A study with 200 participants showed that providing these tailored examples led to better writing performance and a more positive user experience.

In plain terms

Bharatanatyam is an ancient Indian classical dance form where students learn specific sequences of movements and postures, called Adavus. The challenge is to create an automated system that can accurately recognize these complex dance sequences, which are vital for proper learning and practice. The researchers developed a method that first uses a Convolutional Neural Network (CNN), a type of AI for image recognition, to identify individual "Key Postures" (specific body poses). They then used a Support Vector Machine (SVM), another machine learning technique, to recognize the motions occurring between these postures. Finally, an algorithm called Edit Distance was used to compare the recognized sequence of postures and motions with the correct, predefined dance sequences, achieving 98% accuracy. This system can contribute significantly to digital dance heritage and the development of intelligent dance tutoring systems, distinguishing itself by recognizing both postures and motions across all dance variations.

In plain terms

Teaching AI and robotics to K-12 students, particularly in rural areas, is difficult due to a lack of sustained expert mentorship. To address this, the researchers developed a framework called AI, Robotics, & Community (ARC). In this model, colleges act as central 'hubs' that train their undergraduate students to become mentors. These mentors then host workshops and support K-12 robotics teams in nearby schools, with mature school programs potentially becoming secondary hubs themselves. A trial deployment showed significant increases in K-12 students' programming knowledge and resource access, along with increased confidence in teaching for undergraduate mentors. Through simulations, the ARC framework is projected to reach a large percentage of schools and create numerous robotics programs over time, significantly outperforming natural growth.

In plain terms

Large Vision-Language Models (LVLMs) are advanced AI models that can understand both images and text. While these models have shown great progress, their abilities in educational settings, especially for tasks like AI-assisted language learning, haven't been fully evaluated. Existing evaluation tools largely ignore artistic educational content, which is crucial for interpreting visuals in a learning context. To address this, researchers developed MUSE, a new benchmark using diverse artistic images, including art from Singaporean, Southeast Asian, and Western traditions. MUSE comprises twelve tasks, from visual perception to understanding emotions, cultural meanings, and compositional reasoning within the art. Their findings showed current LVLMs struggle significantly with tasks involving affective interpretation (understanding emotions) and compositional reasoning, highlighting key challenges for developing reliable AI in education.

In plain terms

Beginners often struggle to identify their own technical errors in fast-paced activities like fencing, and coaches find it challenging to give specific feedback that truly connects to the student's exact movement. Standard video reviews also have limitations, often lacking different perspectives. To address this, researchers developed FenceXR, an augmented reality (AR) system that reconstructs 3D replays of movements from ordinary smartphone videos. It features a "Trainee module" designed to teach new fencers how to spot common errors in their lunges. Additionally, a "Reviewer module" allows coaches to attach text or voice notes to a specific joint and moment within the 3D replay, offering highly precise and shareable feedback. A study with novice fencers showed their error-detection accuracy significantly improved from 37.5% to 64.1% after one session, with learners focusing more on specific body parts and timing. Fencing experts also found the Reviewer module very useful for conveying feedback that is difficult to explain using traditional video.

In plain terms

Digital Educational Systems (DES) are platforms that help people learn new skills, using everything from simple textbooks to complex recommendation algorithms. This paper explores how these systems are structured by identifying two core design dimensions: "agency" and "legitimacy." **Agency** refers to the degree of decision-making power a learner has within the system, like choosing what to study or how fast. **Legitimacy** relates to how institutionally recognized and credible the platform and its learning outcomes are, such as granting official degrees. The authors mapped various existing DES onto a plane defined by these dimensions, revealing different design philosophies. They found that this mapping can highlight gaps in current DES designs and suggest how artificial intelligence (AI) could algorithmically adjust a learner's agency and a system's legitimacy to enhance the learning experience.

In plain terms

Training powerful AI models, especially those for science, technology, engineering, and mathematics (STEM) education, requires huge amounts of high-quality data. Currently, many large language models (LLMs) are too big for use on common devices like phones or tablets, and there's a shortage of good, openly available datasets specifically tailored for educational STEM content that can make smaller models learn effectively. To tackle this, researchers introduced QVAC Genesis III, a colossal dataset of 191.43 billion "tokens" (pieces of text) covering 19 STEM domains with various difficulty levels and educational styles. It was built using a smart "dual generation strategy" where a small AI model's mistakes are turned into corrective explanations, and its successful answers are expanded with detailed reasoning for all choices. When smaller 1.7-billion-parameter AI models were trained from scratch using QVAC Genesis III, they significantly outperformed models trained on other open-source datasets on key STEM knowledge tests. These improvements reached up to 28.57% on some benchmarks, with models achieving very high accuracy and valid answer rates.

In plain terms

Students often prefer using digital pens for active study, but leveraging AI tools like Large Language Models (LLMs) for learning can be cumbersome due to their reliance on typed input. This creates a gap between the natural, fluid interaction of a pen and the more rigid requirements of asking questions to an AI. The paper identifies two main challenges: a "Referential Barrier," where it's hard to clearly point to specific parts of a document with a text query, and an "Expressive Barrier," where non-textual things like diagrams are hard to convert into typed questions. To solve this, the researchers created "Penquiry," a system allowing learners to ask questions directly on their digital study materials using a pen. Penquiry uses "Content Snapping" to precisely link pen input to content and "Question Autocompletion" to expand sparse ink keywords into full, semantically rich questions for the LLM. Through user studies, they found that Penquiry significantly reduced the effort involved in asking questions compared to traditional typing, offering a more intuitive way to use AI for self-directed learning.

In plain terms

Traditional online learning systems often recommend the next practice exercise based on a student's past performance, which is called Exercise-Centric (EC) recommendation. However, these systems often struggle to provide truly personalized guidance toward specific learning goals because they don't deeply consider what a student already knows, their individual mastery, or how different learning topics are connected. To address this, the researchers propose a new approach called Knowledge-Centric (KC) personalized learning path planning, where an AI system reasons over a student's profile, their current understanding of topics, and a map of how knowledge concepts relate (a *knowledge graph*) to suggest specific textbooks, units, or concepts to study next. They created a new benchmark dataset named PersonaPath, which includes 2,000 detailed profiles of fictional learners (*learner personas*) and a comprehensive knowledge graph mapping out educational content across 77 subjects. When testing powerful AI models, known as *Large Language Models (LLMs)*, on PersonaPath, the best models only achieved a 29.5% success rate in guiding learners through basic education, showing a significant challenge in tailoring paths to individual learners.

In plain terms

Current construction safety reporting often relies on manual logs and static forms, which don't effectively connect daily activities to relevant safety regulations, making it challenging for students to learn. This paper introduces CARES (Conversational AI Reporting for Enhanced Safety), a conversational AI system designed to integrate regulatory guidance into daily reporting for construction safety education. CARES uses proactive multi-agent dialogue (where different AI components work together to guide the conversation) and Retrieval-Augmented Generation (RAG), a technique that combines an AI's ability to generate human-like text with its capacity to retrieve accurate information from a specific database of documents. The system guides users through reporting, retrieves relevant regulations, and converts conversations into structured daily reports, displaying sources and the evolving report alongside the dialogue for user review. A preliminary evaluation with 15 construction management students showed strong "faithfulness" (how accurately the AI's responses reflect the source material) and "answer relevance" (how on-topic the AI's answers were), though "retrieval ranking" (the system's ability to find the *most* relevant regulations out of many possible ones) still needs improvement. These results suggest that a regulation-grounded conversational reporting system is technically feasible.

In plain terms

Durable skills, such as teamwork and critical thinking, are vital for success in today's workforce but are challenging to measure, leading to them often being overlooked in educational curricula. Effective assessments for these skills need to feel like real-world interactions (ecological validity) while also being standardized and repeatable (psychometric rigor). This paper introduces a framework that uses Large Language Models (LLMs), which are AI systems capable of understanding and generating human-like text, to achieve both goals. In this setup, a person interacts with AI teammates in a conversation designed to mimic human interaction. An 'Executive LLM' strategically guides the discussion to elicit specific evidence of skill proficiency, and a separate 'AI evaluator' then measures these skills from the conversation. The research found that this method significantly increases the observable evidence of skills, and that AI-automated scoring largely aligns with assessments from human experts.

In plain terms

Creating engaging, interactive learning experiences, like educational games or simulations, is known to help students learn best through active engagement. However, developing these experiences is typically expensive, time-consuming, and requires significant effort from teachers and designers. To address this, the researchers present "learning interactives," a structured approach that uses advanced Artificial Intelligence (AI) called Generative User Interface (UI) capabilities, specifically from Google's Gemini model. Generative UI means the AI can automatically create user interfaces and interactive content. Their framework focuses on optimizing these AI-generated interactives for pedagogical precision, ensuring they align with good teaching practices rather than just general AI outputs. They evaluated their method through studies with educators and expert teacher ratings of various AI-created simulations, indicating success in generating effective, customized learning tools.

In plain terms

Generative AI tools like ChatGPT are now common in programming education, posing a challenge for educators who want to leverage them without fostering over-reliance. This study investigated whether a structured, "scaffolded" AI tool, which provided limited hints without generating code, would lead to better learning and less dependence compared to students using any AI tool they chose freely. Over seven weeks, 33 master's students were split into these two groups. While assignment performance was similar for both, students using the scaffolded AI reported higher confidence but managed their limited hints poorly. Those with unrestricted access were satisfied with their tools but worried about how much they relied on them. Crucially, students who had developed their own rules for AI use and understood how AI models worked (often self-taught) performed best in both conditions. The research suggests that the specific AI tool's design was less important than a student's capacity for self-governance and their understanding of AI.

In plain terms

Large Language Models (LLMs) are increasingly being used to evaluate students, but it's not well understood how student demographic information, like education level or background, affects these AI tools. This paper addresses that problem, investigating whether considering demographics might improve things (e.g., tailoring feedback) or lead to unfair outcomes (e.g., biased scoring). The researchers set up controlled experiments where they tested six advanced LLMs on three educational tasks: automated essay scoring, giving helpful feedback, and answering questions about language rules. They examined two scenarios: "explicit demographic effects," where details like education level were directly stated, and "implicit effects," where demographics were hinted at through conversation history. They found that LLMs are indeed sensitive to these demographic cues in both explicit and implicit cases, altering their scoring, feedback, and answers. For example, LLMs frequently adjusted the readability of feedback when a student's education level was explicitly mentioned. However, implicit cues sometimes led to unpredictable biases, such as responses from lower-education levels receiving lower sentiment scores in question-answering tasks. These results provide clear evidence that demographics influence how LLMs perform in educational assessment tasks.

In plain terms

Students often find it tough to work with mathematical formalisms, such as logical formulas or mathematical equations, which are crucial in STEM fields. Identifying the common errors they make is very important for providing helpful feedback and creating better educational tools, but doing this manually for many students is difficult. This paper presents a new method that uses Large Language Models (LLMs), which are advanced AI systems, to automatically find these common mistakes. The LLM suggests ways to correct incorrect student answers (called "bug fixing transformations"), and these suggestions are then verified by computer algorithms. This approach helps to identify, group, and visualize common student errors for instructors and education researchers. They showed this method works well by finding known mistakes in logic and can handle large amounts of student data, making it useful for various types of formalisms.

In plain terms

Teaching business analytics requires diverse practice datasets, but real-world data is often hard to get, inflexible, or so common that students (and AI tools) might find pre-existing answers instead of practicing original thought. Instructors also spend significant time creating new case studies, assignments, and solutions. This paper introduces DataCanvas-EDU, an AI-powered system called an "agent" (an AI program designed to perform specific tasks independently). Instructors tell this AI agent their teaching goals and the specific data patterns they want students to discover. The AI agent then automatically writes code to generate completely new, artificial (synthetic) datasets, checks the data, and even prepares assignments, reference analyses, and rubrics (grading guidelines). This approach simplifies course material preparation and ensures students work with unique data, encouraging genuine problem-solving with AI rather than finding pre-existing answers.

In plain terms

Most AI tools in education focus on helping individual students (called individualized student assistance), but teachers often work with groups in collective classroom settings. This paper addresses the gap by creating Edustories, a new dataset of 1,492 real-world descriptions written by elementary and high-school teachers. These case studies detail challenging student behaviors, the specific actions teachers took (called pedagogical interventions), and the resulting outcomes. Researchers used this dataset to test how well powerful AI systems, known as Large Language Models (LLMs), could predict the success of these teacher interventions. They found that while LLMs showed promise, the most advanced AI models achieved 58% accuracy in predicting outcomes, falling short of human experts who reached 64%. This highlights both the current limits and future potential of AI as assistants for practicing teachers.

In plain terms

The CRITICS project aims to make science education more accessible and improve students' critical thinking skills. It addresses the problem that much scientific knowledge is only available in high-resource languages, creating a barrier for many students. To overcome this, the project uses advanced Machine Translation (MT), an artificial intelligence (AI) technology that translates text, specifically powered by Large Language Models (LLMs), which are advanced AI models capable of understanding and generating human-like language. They are fine-tuning these AI translation systems for scientific content to provide accurate and culturally relevant materials in students' native languages. On top of these translations, CRITICS designs and evaluates new teaching approaches focused on developing scientific argumentation and critical thinking practices. This includes using AI to generate textual feedback for students, aligned with specific learning objectives, ultimately working to democratize scientific knowledge and foster deeper thinking in science education.

In plain terms

Teaching complex skills like stage lighting involves abstract concepts, technical details, and making them understandable for learners. Virtual Reality (VR) helps remove physical limits, but existing VR systems don't offer much support for live instruction from teachers. Researchers developed LumiNote, a VR system that uses a Large Language Model (LLM), which is a type of AI that understands and generates human language. Instructors can speak their teaching intentions, and the LLM translates these into helpful tools within VR, including visual notes, interactive demonstrations, and text-based explanations. An exploratory study with 3 instructors and 24 students showed that LLMs were most useful for general, "expressive" teaching goals, though expert intervention was still needed for very specific tasks. While instructors appreciated refining the AI's suggestions, externalized expert reasoning sometimes didn't align with novice comprehension.

In plain terms

Classroom observations are essential for improving teaching quality and guiding how teachers teach, but they are expensive and need highly trained human experts. This paper investigated whether a large language model (LLM), specifically GPT-5, could automatically score teacher-child interactions in early childhood classrooms. The researchers analyzed 87 video-recorded observations from kindergartens in Hong Kong, using only the observation transcripts. The AI was configured to apply the Classroom Assessment Scoring System (CLASS), a common framework for evaluating classroom quality. They found that AI scores aligned better with human scores for aspects of "Emotional Support," especially how teachers give feedback. However, there was less agreement for more routine or context-dependent interactions in "Classroom Organization" and "Instructional Support." This suggests that while AI can capture some differences in teacher-child interactions from text, it cannot yet consistently match the nuanced judgments of trained human observers. Therefore, AI-assisted observation might be more useful as a preliminary tool for teacher reflection rather than for high-stakes evaluations.

In plain terms

In hands-on design courses, like knit yarn design, students often struggle to turn their creative visual ideas into actual products due to complex material properties, technical details, and limited opportunities to make physical samples. To tackle this, researchers developed a new teaching method that combines "flipped learning" (where students learn basic concepts at home before class), reviewing examples, using Generative AI (GenAI) for visual prototyping, and getting feedback in the studio. This method was implemented via a digital platform offering pre-class videos, quizzes, an example gallery, and a GenAI tool for early ideation. An exploratory study comparing students using this new method to a control group showed higher scores in creativity, design skills, problem-solving, and overall course performance for the intervention group. Student and instructor feedback suggested that flipped learning effectively prepared students for studio work, while GenAI primarily aided early-stage visual exploration rather than providing precise technical instructions.

In plain terms

Reinforcement Learning (RL) is a complex type of artificial intelligence where systems learn by trying actions and getting rewards or penalties in an environment. This paper introduces RLLBC-Lib, a new software library designed to make it easier for students and beginners to understand RL, especially its use in controlling systems. The creators observed that the complex interactions in RL are best understood through accessible implementations. The library provides clear code examples, starting with simpler "tabular RL" methods to build foundational knowledge, then progressing to more advanced "deep RL" techniques while showing their connections. It also illustrates core RL principles and how RL differs from other learning-based control approaches. Ultimately, this library serves as a practical resource for educators to create hands-on programming assignments that can be automatically graded, helping students truly grasp these challenging AI concepts.

In plain terms

One significant challenge in the field of AI in education is that AI systems develop so quickly, traditional large-scale evaluations often become outdated before their results are even published. This paper proposes and demonstrates a faster evaluation method called "micro-randomised controlled trials" (micro-RCTs), where participants are randomly assigned to different conditions (like using the AI or not) multiple times over a short period. They tested an AI tutoring platform named Medly with nearly 1000 secondary school students revising for GCSE science exams (Biology, Chemistry, Physics). The study found that students using the Medly platform achieved higher post-test scores compared to those doing their usual self-directed revision. While promising, the authors acknowledge limitations like student drop-out, suggesting these rapid trials are best for generating quick, cumulative evidence rather than definitive, one-off conclusions.

In plain terms

This study addresses the challenge educators face in managing the increasing use of generative AI (GenAI), like ChatGPT, in classrooms while ensuring ethical learning. Researchers explored a guided activity where students actively helped create policies for using GenAI in a course. Students first brainstormed individual policy ideas, some even using GenAI for inspiration, then collaborated to refine these ideas by considering different viewpoints. The study found that students highly valued training for both themselves and their instructors on GenAI use, clear rules for when and how to disclose AI use in assignments, and greater support from their institutions. They also expressed a desire for more involvement in decisions about GenAI. This policy co-design process helped uncover students' perspectives, concerns, and understanding of how GenAI should be used in learning.

In plain terms

Mechanical engineering students often lack sufficient training in artificial intelligence (AI), electrical, and computer engineering, despite these skills becoming crucial for tackling complex problems like those in thermal systems. Their heavy course loads also limit opportunities for new learning. To address this, researchers at the University of Arkansas designed a new curriculum to integrate AI into their mechanical engineering program. This curriculum uses a project-based learning approach, with a special focus on thermal engineering problems, and is divided into introductory, application, and advanced levels. The main goals are to improve students' understanding of AI models, boost their ability to solve engineering challenges using AI, and develop their skills in communicating across different technical disciplines. The curriculum and its resources are openly available to help other educators implement similar programs.

In plain terms

When medical students use Artificial Intelligence (AI), especially advanced generative AI, there's a concern they might over-rely on it, use it improperly, or fail to develop crucial clinical skills. This paper suggests that the real issue isn't simply "misuse," but rather a "misclassification"—meaning learners choose the wrong way to interact with AI for a specific task. To address this, the authors introduce "SCAN," a human-centric framework for deciding how AI should be allocated across different tasks: Substitute, Complement, Aid, or Non-Negotiable. This framework is rooted in Vygotsky's Zone of Proximal Development (the gap between what a learner can do alone and what they can do with help) and metacognition (thinking about one's own thinking process). A key insight is "passive engagement," where learners might appear to use AI correctly but don't deeply learn, potentially leading to mis-skilling; in such cases, the paper suggests human experts serve as "epistemic auditors." The SCAN framework is designed to be practical, offering educators a tool for clinical curriculum design, student supervision, and assessment.

In plain terms

This study looked at how different groups at a large university use and feel about Artificial Intelligence (AI). Researchers surveyed over 2,000 students, faculty, and administrative staff to understand their frequency of AI use, how useful they found it, their trust levels, and their concerns about issues like academic honesty. They found a big difference, or "AI adaptation gap," where students use AI more frequently and see it as more useful than faculty and staff. Conversely, faculty and staff had greater worries about academic integrity and stressed the importance of using AI responsibly. The study also found that believing AI is useful strongly increases trust in it, and clear institutional policies also help build trust. This suggests universities need to consider these varying perspectives when implementing AI.

In plain terms

Current methods for assessing student knowledge, like those based on **psychometric models** such as **item response theory**, are good at measuring what students know but struggle to explain *why* they perform a certain way or predict the effect of giving them help. This paper proposes using **structural causal modeling (SCM)**, a framework that explicitly maps out cause-and-effect relationships between different factors in a learning scenario. The authors developed a clear protocol for building these SCMs, which relies on gathering logical rules from experts rather than complex statistical assumptions. They demonstrated how SCM can effectively model the impact of support interventions, like giving a hint, on student performance. Crucially, this approach allows for **interventional reasoning** (understanding what happens if we change something) and **counterfactual reasoning** (exploring "what if" scenarios), which are largely inaccessible with traditional assessment models. This means educators can better predict the outcome of different support strategies and even analyze what might have happened if a student had received different help.

In plain terms

Predicting how engaged a student is during online learning videos is incredibly difficult because engagement is a complex mix of behaviors, emotions, and thoughts. The challenge is amplified by high individual differences among students and the subjective nature of judging engagement. To tackle this, the researchers developed a comprehensive AI system that combines various types of information, including subtle spatial and temporal features extracted from video, audio, and images. It also incorporates structured behavioral data like head pose, gaze (where someone is looking), facial expressions (action units), and emotions, integrating these via an advanced neural network architecture. Additionally, the system models student and instructor personalities to better account for individual variations and uses statistical methods to not only predict engagement but also to indicate how confident it is in its predictions. While all methods struggled on a challenging dataset, their framework achieved competitive performance and uniquely offers reliable quantification of prediction uncertainty, which is crucial for practical applications.

In plain terms

Geometric proof is a core but challenging topic in mathematics that requires students to combine visual understanding with logical reasoning. While technology has significantly improved learning in other math areas, its impact on geometric proof has been limited. To understand this, the researchers interviewed 18 geometry teachers to find out what features they need in educational tools. They then reviewed 33 existing commercial and research tools. They found that teachers appreciate digital tools for brainstorming ideas but ultimately return to pen-and-paper for formal proofs. This is because physical paper allows them to easily annotate diagrams and explore multiple solution paths, a key feature that current digital tools lack. The paper proposes design guidelines for future tools, including integrating diagrams with proofs, automatically generating problems and feedback, supporting various proof formats, and creating user-friendly experiences.

In plain terms

In mathematics education, teaching complex proofs like structural induction is challenging because students need instant feedback and step-by-step guidance, which is hard to provide manually. While 'proof assistants'—software tools that help verify mathematical proofs—could offer this, they are often too complicated for students and don't always help them apply what they learn to pen-and-paper exams. Researchers developed "Hazel Prover," a new proof assistant specifically for classrooms, designed for teaching equational and inductive reasoning. They tested it in two classes, collecting detailed usage data, surveys, and exam responses. They found that students learned to use the tool effectively and improved their proof-writing skills. However, the first version of Hazel Prover, which gave too much assistance with simple algebraic steps, didn't help students transfer their skills to traditional pen-and-paper proofs. After redesigning the tool to require students to engage more actively with these basic steps, the second deployment showed much better transfer to traditional assessments.

In plain terms

Learning from step-by-step problem solutions, called worked examples, can be a passive activity. A more effective approach is self-explanation, where students actively explain problems and solutions to themselves, but providing feedback on these free-text explanations has been a challenge for human instructors and automated systems. This research investigates if a Large Language Model (LLM)—an advanced AI capable of understanding and generating human-like text—can offer timely, automated feedback. They developed ESSE, a self-explanation tutor for introductory programming (CS1) students, which uses an LLM to assess student explanations of code line-by-line. The study found that the LLM's feedback was sufficiently reliable when compared to human experts. Furthermore, students using ESSE refined their explanations, making them more thorough and conceptually deep, and demonstrated improved learning outcomes.

In plain terms

Traditional cybersecurity training in universities often struggles with low student engagement and effectiveness, especially as AI-driven cyber threats become more sophisticated. To tackle this, researchers developed several short, mobile-friendly games to teach cybersecurity topics, like password safety and scam recognition. These games used 'AI-powered gamification,' meaning they incorporated game-like elements such as quizzes, stories, and simulations, potentially using AI to personalize the learning experience. The goal was to make learning more interactive and user-friendly, aiming to improve both engagement and educational outcomes. They tested these games with 59 college students, a mix of technical experts and general users. The findings suggest that this AI-powered gamification approach has the potential to significantly improve student engagement and attention to crucial cybersecurity topics.

In plain terms

Generative AI (GenAI) is often viewed simply as a helpful tool in education, but this paper argues it's causing a major shift in who holds power and authority over knowledge in classrooms. Traditionally, teachers and students produce knowledge, but now AI is increasingly involved, raising questions about who validates claims and creates learning evidence. The author critically examined various established theories on agency (the ability to act and make choices) and technology use in education to understand how GenAI redistributes this ability among students, teachers, and machines. While relevant, existing theories didn't fully address inequalities in power, data ownership, or accountability in human-AI interactions. To tackle this, the paper introduces the Ecological Co-Agency Framework, which defines agency through relational, regulatory, and pedagogical processes. This framework emphasizes human accountability for the knowledge claims made when working with AI.

In plain terms

How can universities effectively teach Artificial Intelligence (AI) and engage students in practical applications, research, and community service? Researchers at the Federal University of Santa Catarina developed an organizational framework for their Academic League of Artificial Intelligence (LIA). This framework focuses on student-centered, project-based learning, combining democratic governance (where students have a say), collaborative learning, and flexible project organization. They aimed to develop both technical skills in AI and 'transversal competencies,' such as leadership and teamwork. The framework was successfully applied through various initiatives, including student teams for competitions, study groups, public lectures, and creating AI tools for social good. The authors suggest this flexible model can be replicated by other student organizations to integrate teaching, research, and community extension in engineering and computing fields.

In plain terms

This research looks at how modern technology, especially things like fake news and deepfakes (content often created or spread by artificial intelligence), affects education, specifically in teaching people to be critically aware of media. The researchers wanted to understand what students and experts think about these challenges and how well current education prepares students to critically evaluate information. They did this by interviewing leading experts in critical media literacy and surveying 141 university students. The study found that while new technologies offer ways to improve learning, they are not deeply integrated into school curricula. Furthermore, teachers often lack the necessary training to effectively use these tools to help students develop critical thinking skills. The authors conclude that we need a stronger focus on critical media literacy that teaches students not just how to use technology, but also how to question and ethically evaluate the information they encounter, fostering active and thoughtful citizenship.

In plain terms

Providing timely and personalized feedback is crucial for student learning, but it's hard to do at scale. While generative AI can help create feedback comments, students often don't use them effectively. This study investigated how different ways of delivering AI feedback—called 'workflows'—impact student engagement and learning. They compared three approaches across over 13,000 students: simply receiving AI comments, opting to get AI feedback, or being actively prompted to select, evaluate, and discuss AI suggestions. The researchers found that when students were guided to actively interact with and reflect on the AI feedback (the 'Enacted Feedback' workflow), they were significantly more likely to use it. This active engagement also led to higher student confidence in their self-assessments and better quality in their submitted work, showing that the design of the AI interaction is as important as the feedback quality itself.

In plain terms

Large Language Model (LLM) tutors, while capable, can hinder long-term learning if they too easily provide direct answers, as demonstrated by previous research where unguarded chatbots led to lower test scores later. This paper tackles the challenge of teaching an LLM tutor to act more like a Socratic teacher, guiding students without immediately revealing solutions. The authors designed a "supervisor architecture" with a core policy that controls help levels, a tool to remove solution code, and an LLM "judge" to check responses for over-helping. They also developed an automated testing method using simulated student scenarios, allowing them to systematically identify and fix how the tutor might inadvertently give away too much information, from direct solutions to overly specific hints. This iterative process helped them refine the tutor's behavior, leading to full compliance with its answer-withholding requirements.

In plain terms

Automated writing evaluation tools often mix up simple grammatical mistakes with phrases that don't sound natural to a native speaker, making it hard for language teachers to provide targeted feedback. This study tackled this issue by developing a system that uses Large Language Models (LLMs), which are advanced AI programs capable of understanding and generating human language. Their system analyzes English writing samples from Japanese junior high school students in two distinct layers: first, it identifies clear grammatical errors, and then it looks for unnatural phrasing or less common ways of expressing ideas. By doing this, they could precisely measure both 'accuracy gaps' (incorrectly used grammar) and 'idiomatic gaps' (underused or overused natural expressions). They found specific areas where students struggled with accuracy (like 'the' or '-s' for verbs) and others where they struggled with idiomatic use (e.g., underusing '-ing' forms but overusing 'can'). This allows teachers to pinpoint whether a student's difficulty comes from making mistakes, avoiding complex structures, or over-relying on patterns from their first language, leading to more effective, personalized teaching.

In plain terms

Large Language Models (LLMs) are powerful but risk making students just get answers instead of truly understanding math, especially when preparing for important exams. Researchers developed AITutor, an interactive system designed to translate theoretical teaching methods into practical user interface features. They tested it with 12 junior-high students preparing for high-stakes exams (Zhongkao) using various methods, including a generative study, usability study, and field deployment. The study revealed that under time pressure, students resisted traditional Socratic dialogue (guided questioning) and instead used "answer-first" shortcuts as crucial checkpoints to diagnose their mistakes. They demonstrated that features like layered worked examples, step-linked visual grounding (connecting steps to visuals), and metacognitive scaffolding (prompts to reflect on thinking) effectively reduced the effort needed for students to repair their reasoning. The paper introduces a "Reasoning-Centered Product Loop," offering practical implications for designing AI that structurally supports the inspection, local repair, verification, and later recall of mathematical reasoning in real-world learning.

In plain terms

Large AI models are being explored as tutors, but they come with concerns about privacy and cost, especially in K-12 schools. Smaller AI models (SLMs) offer a promising alternative, but it's hard to choose the right one for specific educational contexts, particularly for subjects like block-based programming (a visual method for teaching coding) that might not be well-represented in their training data. To address this, researchers developed "CSTutorBench," a benchmark to evaluate how effectively language models can tutor students in VEX VR, a block-based robotics environment. This benchmark includes 17 scenario-based questions, and the models' tutoring responses are scored using a pedagogical rubric, which judges their quality based on established teaching and feedback principles. They found that while models perform well on basic criteria like vocabulary and tone, they struggle with more complex teaching behaviors, such as avoiding giving away the answer directly (answer leakage) and effectively using a student's past debugging attempts. Interestingly, the model family and how it was specifically instruction-tuned seemed to predict tutoring quality better than just the model's size, and a small improvement to the prompt (the initial instructions given to the AI) boosted scores for most models.

In plain terms

Engineering students often struggle to connect spatial concepts with abstract ideas, a gap that traditional lectures might not fully address. This pilot study explored whether interactive learning environments, which allow students to engage with material through action and feedback, could help. Researchers compared two types of post-lecture interventions: an immersive Mixed Reality (MR) application, which blends real and virtual worlds, and a tangible "Engineering Toolkit" with physical models. These were compared against a control group that only received a lecture, with 24 undergraduate solid mechanics students participating. The study found that both the MR and tangible tool conditions significantly improved students' conceptual understanding compared to the lecture-only approach. While Mixed Reality showed the highest learning gains, the tangible toolkit was rated higher for usability and boosting student confidence.

In plain terms

The widespread availability of powerful AI tools like ChatGPT presents a significant challenge to education, as students can now easily complete complex tasks without genuine effort or learning. The author calls this the "effortless bypass dilemma," which threatens to undermine education at scale. This paper argues that the field of Artificial Intelligence in Education (AIED) needs to adapt its mission. Instead of solely focusing on building more effective intelligent tools, AIED must prioritize ensuring learners *choose* to engage authentically with their learning, even when easier AI alternatives are available. Drawing on established theories of motivation and learning, the paper outlines five key directions for AIED: supporting student autonomy and agency (their sense of control and choice), building learner resilience to metacognitive threats (like the urge to take shortcuts), designing for interest and relevance, amplifying process-based assessment (evaluating how students learn, not just what they produce), and empowering teachers. The author also shares four envisioned technologies that embody these future-focused principles.

In plain terms

When students use Generative AI (GenAI) for tasks like programming, how do they effectively think and learn? This paper addresses this by introducing a new concept called "Epistemic AI Literacy" (EAIL). EAIL reframes AI literacy as a process of how students acquire, evaluate, and justify knowledge while interacting with AI, rather than just knowing about AI. The researchers examined how students set learning goals (epistemic aims) and the strategies they use (epistemic processes) during human-AI co-programming. They analyzed a large dataset of student-AI dialogues, identifying processes like outsourcing, explanation seeking, and verification. The study found that most students (78.8%) exhibited low EAIL, relying on less effective strategies like simply letting the AI do the work or just checking its output. Only a small percentage (11.1%) demonstrated high EAIL, combining mastery-oriented goals with deeper engagement like seeking justifications.

In plain terms

Statistical graphs are typically visual, making them difficult for blind and low-vision students to understand. While 3D-printed tactile graphs allow these students to 'feel' and interpret data, their creation has been limited due to the specialized skills and manual effort required to design each graph. To solve this, researchers developed a three-layer software pipeline written in JavaScript that automates the entire process. The first two layers use tactile perception research and modular builders for common chart types (like bar and scatter plots) to design the graphs. The crucial third layer utilizes a large language model (LLM), an advanced AI, to extract chart specifications directly from an uploaded image of a graph, though a teacher must review it before printing. This pipeline can generate print-ready 3D files in less than 250 milliseconds, making accessible educational materials much faster and easier to produce.

In plain terms

Existing educational frameworks, like Bloom's Taxonomy, don't fully account for the real-world consequences of student actions, especially in practical computing fields like command-line environments where simple commands can have significant impacts. To address this, researchers developed CogTax, a new four-level cognitive taxonomy that combines the cognitive complexity of a task with its 'operational impact'—how much a command affects a system. This framework categorizes actions from safe observation to advanced system management. CogTax helps instructors structure course material and calibrate assessment difficulty, and provides students a clear reference for self-assessment. To make this framework scalable, they trained an AI classifier (a machine learning model) using syntactic information and semantic meanings of commands. This AI can automatically assign commands to the correct CogTax level with 89% accuracy, demonstrating a practical way to integrate AI into curriculum design.

In plain terms

Large Language Models (LLMs) enable a new way of programming where users describe their computational goals in natural language, known as "prompt-based programming." However, users, especially those new to computer science, often struggle to write clear and specific instructions, or "prompts," for these LLMs, leading to less effective results. This study investigated whether a brief lesson could help address this issue. The researchers developed a 45-minute lesson with guided practice focused on teaching effective prompt writing for LLMs. They then conducted a randomized controlled study with 55 engineering students, comparing the effectiveness of this lesson against a standard computer science activity of the same duration. The findings showed that students who received the prompt-based programming lesson improved more in their ability to specify computational goals to LLMs and significantly increased their confidence in using prompts, compared to the control group. This suggests that even a short intervention can start to improve learners' skills in interacting with LLMs, though developing deeper proficiency might require more extensive practice.

In plain terms

Intelligent tutoring systems aim to personalize learning, which requires predicting student performance and understanding how accurately students assess their own knowledge, known as **metacognitive calibration**. Traditionally, predicting performance and analyzing calibration were done separately, making it hard to interpret everything together. This paper introduces the UBP-CAP framework, an integrated system that uses **machine learning** to process student behavioral **telemetry**—data collected from their interactions within a learning system. First, it employs a **LightGBM classifier** (a type of machine learning model) to predict if a student will answer a question correctly based on their actions before attempting it. Second, it calculates formal metrics to measure how well students' self-assessments align with their actual performance, identifying errors in their metacognitive calibration. Finally, a **Generalized Linear Mixed-Effects Model (GLMM)** is used to break down the reasons for these calibration errors, distinguishing between factors specific to the student versus the task. The study found that students often systematically misjudge their own understanding, and this self-assessment ability is more dependent on the specific learning task than on a general student trait.

In plain terms

It's hard to understand how students work together in teams, especially their communication roles, and existing computer-based methods often miss insights from educational theories. To address this, researchers observed students in a university computer science course, analyzing over 6,000 Slack messages from 55 students in 18 teams. They applied an existing educational framework to identify eight distinct communication roles, like "initiator" or "supporter." To scale this, they tested if large language models (LLMs) – AI programs skilled at understanding human language – could accurately identify these roles, finding they could approximate human expert labels. They discovered that students play a more varied set of roles as a project progresses, and different roles become more active at various stages of team work. Using these AI-identified roles, they could successfully predict which students would be recognized by their peers and even predict improvements in team performance in a separate dataset.

In plain terms

This research explores whether the famous initial level of Super Mario Bros (World 1-1), known for subtly teaching players game mechanics, truly has an empirically measurable pedagogical structure. The authors implemented World 1-1 as a digital environment and trained different Artificial Intelligence (AI) agents, specifically using Reinforcement Learning (RL) algorithms like Q-Learning and Monte Carlo, to play it. Reinforcement Learning is a type of AI where an agent learns to make decisions by trying actions and receiving rewards or penalties. They found that the Monte Carlo agent was the most successful, especially because it learned to take paths that maximized smaller, intermediate rewards, not just the final win. Crucially, by shuffling the level's segments, they discovered that the original order of World 1-1 significantly accelerates the AI's learning, achieving the highest efficiency and preventing learning failures. This validates its exceptional design as a teaching tool, confirming that its structure measurably accelerates learning.

In plain terms

AI can now quickly create interactive learning materials, but it's difficult to properly assess how good their interactivity truly is—meaning how effectively they let learners explore and receive useful feedback. Current evaluation methods mostly focus on things like code working or visual appeal, not the actual quality of the learning experience. The researchers developed EE-Eval, an automated framework that formalizes the interaction logic of these AI-generated materials as a 'Finite State Machine' (FSM), which is like a map showing all possible learner actions and system responses. They extract these FSMs from AI-generated content and compare them to an 'ideal FSM' that represents the intended teaching strategy. EE-Eval consistently differentiated interaction quality beyond simple checks, and its assessments aligned much better with human judgments of how interactive and effective the learning materials were.

In plain terms

Many teachers want to use Artificial Intelligence (AI) to create teaching materials, but existing tools aren't always designed for their classroom needs. This paper introduces "scrutable interfaces," which are user-friendly designs that let people easily adjust an AI's output without understanding its complex inner workings. They developed an AI tool called Concept Catalyst with such an interface, specifically to help K-12 teachers reflect on their teaching as they use generative AI (AI that creates new content, like text or images) for curriculum development. Through an exploratory study using "Wizard-of-Oz testing" (where a human secretly simulates the AI), they found that these interfaces positively structure teacher interactions with AI. This approach helps teachers reflect on their practices, makes their AI use more effective and efficient, and increases their motivation.

In plain terms

This research investigates how Large Language Models (LLMs) can best help students learn programming by providing explanations for code. The problem is that while LLMs can generate explanations, it's unclear if a single generic explanation is the most effective approach compared to human instructors. Researchers hypothesized that multiple, diverse explanations, each highlighting a different aspect (like what a function does, a core concept, or the overall goal), might be more beneficial. They tested this with 971 first-year computing students, randomly assigning them to receive either 'diverse' or 'generic' LLM-generated explanations for programming exercises. The study found that students who received diverse explanations consistently scored about 7.7% higher on open-ended questions, showing improved understanding. Crucially, this improvement did not lead to an increase in their perceived mental effort, known as cognitive load.

In plain terms

When using AI to advise educators on student support, large language models (LLMs) often recommend too many interventions, even when students don't need them—a problem called "intervention bias." This study tested LLMs like GPT-4o and other AI systems, including retrieval-augmented generation (RAG) systems, on a dataset from the Open University, which tracks student performance and ideal intervention points. They found that zero-shot LLMs and RAG systems had a high false-positive rate, recommending actions for many students who didn't actually need help (e.g., 43% unnecessary recommendations at a certain point). However, traditional supervised machine learning models, like a Decision Transformer and an XGBoost classifier, trained specifically for this task, were much more accurate and didn't show this intervention bias. These supervised models were also very fast and could predict the correct action for nearly all students, even for rare situations, and the study also found that common methods for evaluating LLMs often fail to detect this over-prescription bias.

In plain terms

This study explored how Grade-9 students interact with general-purpose AI tools, like ChatGPT, for learning mathematics in open-ended settings, aiming to understand what kind of interaction leads to better understanding. The researchers observed 112 students using an AI tutor for math practice, analyzing their conversations with the AI. They coded student messages for self-regulated learning behaviors, types of help requested, and engagement in mathematical problem-solving, specifically looking at how these patterns changed over time during a session. They found that simply summarizing a student's overall AI use, like the total number of questions, did not predict learning outcomes. However, students performed better when their interactions showed a temporal shift: moving towards more conceptual or procedural help-seeking and independent mathematical work in later stages, instead of primarily asking for answers or verification. This suggests that actively and strategically guiding the AI to advance one's understanding, termed "epistemic proactivity," is key for productive learning.