AI for Learning — Daily Research

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

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

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

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

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

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

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

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

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

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

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

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

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.

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

This paper reviews Transformer-based language models, which are the powerful AI systems behind tools like ChatGPT. It first categorizes different types of these models, explaining how they are built and what makes them unique. The authors then discuss recent breakthroughs since 2023, such as methods for training models to follow instructions better (instruction tuning) and using human feedback to improve their responses (reinforcement learning from human feedback). A key part of the paper surveys how these AI models are used in various fields like healthcare, finance, and importantly, education, linking specific model capabilities to their real-world uses. Finally, the paper critically assesses these models, comparing their architectures, energy costs, and how we measure their 'state-of-the-art' performance, concluding with open research questions.

In plain terms

Expertise has traditionally been explained by things like hidden knowledge or repetitive practice. This paper proposes a new idea called "Workflow Cognition" to describe how expertise really develops and changes. It suggests that expertise is not just stored knowledge, but a dynamic system where "Thinking Flow" (our ongoing thoughts, judgments, and reflections) constantly interacts with "Workflow Evolution" (how we adapt our actions and strategies in practice). Through this continuous interaction, expertise emerges and grows over time, meaning it's always evolving, not a static skill. The paper introduces a model for this and aims to help build AI systems that can better understand and represent human expertise.

In plain terms

Predicting how difficult a test question, or "item," will be for humans is crucial for creating fair and effective educational assessments. Existing methods often rely on expensive human trials or simply analyzing the question's text, which doesn't explain *why* an item is hard. The researchers developed Epi2Diff, a framework that uses the step-by-step thinking processes, called "reasoning traces," generated by advanced AI models (Large Reasoning Models or LRMs) when they attempt to solve a problem. Epi2Diff breaks these traces into "cognitive episodes," which are distinct stages of problem-solving like understanding or planning. By analyzing the flow, complexity, and transitions between these episodes, Epi2Diff creates unique features that predict human item difficulty more accurately. They found that Epi2Diff consistently outperformed other methods, showing that harder items lead to more iterative and complex AI reasoning, not just longer answers.

In plain terms

This study looks at whether learning benefits from using "adaptive AI-assisted pretesting"—where an AI tailors introductory questions before instruction—last over a long period. Pretesting is known to help students learn by activating what they already know, but it's unclear if the advantages of *adaptive* pretesting, especially when powered by Generative AI (AI that can create new content), persist over time. Researchers had college students complete an adaptive AI pretesting session, followed by instruction, and then assigned them to one of three different seven-week practice methods: adaptive spaced retrieval, fixed spaced retrieval, or learner-directed AI study. "Spaced retrieval practice" involves reviewing information at increasing intervals to improve memory, and in this study, it was either tailored by AI or fixed. They found that adaptive pretesting did improve initial understanding, but the sustained learning gains depended heavily on the subsequent practice method. Specifically, both retrieval practice groups outperformed the learner-directed AI study group, indicating that structured AI-supported practice is key for long-term retention.

In plain terms

This paper explores how Large Language Models (LLMs), which are advanced AI systems that can generate human-like text, can be used to make learning more engaging. Instead of just giving answers, the goal was to see if LLMs could encourage students to explore topics more deeply on their own. Researchers created a framework called CURIOBOT that uses specific conversational tactics, known as "linguistic interventions," designed to spark curiosity, like introducing novelty or uncertainty into the dialogue. They tested CURIOBOT in many tutoring conversations, using different LLMs and across various subjects. They found that these curiosity-focused interventions consistently increased "exploratory learner behaviors"; students engaged in up to 2.4 times more conversational turns within a fixed time, showing greater interaction. This suggests that fostering curiosity is a powerful way for AI tutors to improve how students interact with and learn from material, independent of just delivering information.

In plain terms

Collaborative programming projects are a key part of learning software engineering, but the widespread use of generative AI tools (AI that can create text, code, etc.) by students introduces a new challenge: knowing if and how much your partner is using AI. This research looked at whether students accurately perceive their partners' AI use and how that perception impacts their teamwork. The researchers conducted a study with 103 pairs of students in an introductory software engineering course, tracking their beliefs and project outcomes over time. They found that a greater "misalignment" – or difference – between partners' beliefs about each other's AI use early in the project was associated with lower final project scores. This problem was even more pronounced for students who had less prior programming experience, suggesting they pay a higher cost for these misaligned perceptions. The study suggests that simply working together doesn't always resolve these misperceptions, highlighting a need for better ways to make AI use transparent in student collaborations.

In plain terms

Artificial Intelligence (AI) education is becoming increasingly important, but adults who are not in higher education often receive less attention in this area. This paper presents a case study describing an AI education session with 54 adults in a predominantly African American community in a major Midwestern city. The researchers worked closely with local community partners to jointly develop, or co-design, this educational session. They discovered that while adults initially had general concerns about AI, these worries evolved into more specific, local questions about how AI is designed and put into use within their own community. The study also discusses AI literacy, which is the ability to understand and use AI, from the viewpoint of a community's overall capacity. They argue for creating AI literacy frameworks that are rooted in local community contexts to better empower communities.

In plain terms

Educational games, especially for complex topics like financial literacy, often struggle to assess what students truly learn without interrupting the game with tests. This challenge is known as "stealth assessment." Researchers developed an AI system called "Agentic BKT pipeline," which employs multiple large language models (LLMs) to observe students playing a financial literacy game. First, an LLM categorizes every player action; then, four specialized LLM "agents," each an expert in areas like investing or credit, analyze the player's behavior in their specific domain. These agents feed their observations into a statistical model called Bayesian Knowledge Tracing (BKT), which estimates a student's mastery level in each area, and an expert judge LLM combines these for an overall score. They found that this multi-agent system produced mastery scores that accurately reflected how much students learned and their scores on post-game tests, proving significantly more effective at predicting learning than a simpler single-LLM approach.

In plain terms

This paper explores how modern digital and AI environments are reshaping human thinking and learning. It suggests that the constant flow of information and algorithmic filtering in these environments can lead to "shallow cognitive patterns," meaning we might become less skilled at deep, complex, or "recursive cognition" – the ability to build on previous thoughts or knowledge iteratively – and rely more on external tools. This creates an "ecological mismatch" because academic settings still require integrated, deep reasoning. The authors introduce the "Digital Pirahã Condition" as a model to explain these cognitive adaptations as an outcome of our digital surroundings, not a personal deficit. To address this, they propose that educators should act as "cultural entrepreneurs" by actively redesigning learning environments. This redesign involves creating non-digital spaces, using AI-supported tools to help students reflect on their own thinking (metacognitive scaffolds), and structuring curricula to encourage deeper, iterative thought processes.

In plain terms

Generative AI tools, like ChatGPT, offer instant support to students learning to program, but researchers question if this truly helps students learn deeply or just gives them quick answers. This paper addresses this by looking at how students ask for help from AI, defining specific ways students prompt AI (called 'help-seeking trajectories'). They used a framework that linked these prompts to different types of support, like understanding concepts or debugging code. The researchers analyzed over a thousand prompts and thousands of code submissions from 71 students in introductory Python courses. They found that many students primarily used AI for 'reactive troubleshooting' – fixing problems as they arose – instead of engaging in more planned, 'self-regulated problem-solving'. Although different help-seeking patterns didn't significantly change final task scores, they did lead to substantial differences in how many times students submitted their code, suggesting varying levels of efficiency or struggle.

In plain terms

Educators are grappling with how to use AI, as it can both help and hinder student learning, creating confusion. This paper argues that the critical factor isn't whether to allow or ban AI, but *how* it's placed within the learning process. The authors introduce a 'six-move' model for learning—Prime, Probe, Point, Attach, Strengthen, and Test—to guide this placement. They found that poorly used AI can lead to an 'illusion of learning,' where students feel confident but struggle on unaided tasks, while well-designed AI tutors can significantly improve outcomes. Their framework suggests securing initial attempts and final assessments without AI, using 'guarded AI' for scaffolding in between, and using the rule: if AI makes a task feel effortless, it's in the wrong place.

In plain terms

Designing Intelligent Tutoring Systems (ITS) that effectively help students usually requires either costly real-world experiments or complex computer simulations, which might not always reflect how students actually learn. This paper introduces a method to learn improved personalized teaching strategies directly from existing student data, without needing new experiments. They treat the tutoring process as a "continuous stochastic bandit problem," which means picking the best next task for a student based on their progress, and use a statistical model called the Rasch model to understand student skill and task difficulty. They also developed a new way to measure success, called a "reward function," designed to keep students in a "flow" state by balancing challenge and success. The researchers tested their method on four large real-world datasets and consistently found that their newly learned teaching strategies were better than the original ones used to collect the data. This means improved adaptive learning policies can be found and visualized quickly using only existing information, making it easier to enhance tutoring systems.

In plain terms

Large Language Models (LLMs), while powerful, often struggle to act like good human tutors, sometimes revealing answers instead of guiding students. This paper addresses this by developing a method to train LLMs to follow effective teaching strategies, specifically for helping students correct math mistakes. The researchers used a two-step process: first, "supervised fine-tuning" involved training the LLM on existing tutoring conversations. Second, "Direct Preference Optimization" was used, where the LLM learned by being shown pairs of responses and instructed which one was pedagogically better (e.g., providing hints, known as scaffolding, versus giving direct answers). They also created a new dataset combining real and synthetic tutoring examples focused on pedagogical quality. Their approach significantly improved both the factual accuracy of the math help and the teaching quality of the LLM, with human evaluations showing it to be comparable to leading proprietary AI tutors.

In plain terms

AI is increasingly used in education, but we don't fully understand how different ways students interact with it affect their learning and cognitive engagement. This study addresses this by moving beyond simply comparing AI to no-AI learning. Researchers observed 48 high school students (ages 14-18) completing quizzes in three distinct AI interaction modes: as a 'Tutor' (guiding), a 'Collaborator' (working together), or a 'Solver' (providing answers). They measured student behavior, tracking initiation, processing, and stress, and also collected electroencephalography (EEG) data, which captures brain activity via electrical signals on the scalp. The study found significant differences in how students behaved across the three AI interaction modes, indicating that the way AI is presented significantly impacts learning approaches. While brain activity patterns were observed, they did not reach statistical significance, suggesting more research is needed, but the behavioral findings are robust.

In plain terms

Generative AI (GenAI), like ChatGPT, is making universities rethink how they assess students, creating pressure to maintain fairness and prevent academic misconduct. The Artificial Intelligence Assessment Scale (AIAS) is one framework designed to help universities adapt their assessment strategies for GenAI. This study examined how academic staff experienced implementing the AIAS at two universities, one in Vietnam and one in the UK, using focus groups with 30 staff members. Staff valued the AIAS for establishing a shared understanding of GenAI use, clarifying boundaries, and prompting reflection on assessment design. However, its implementation was influenced by factors such as university governance, access to AI tools, staff confidence, workload concerns, and the need to align with specific learning outcomes and disciplinary contexts. The findings suggest the AIAS can encourage more authentic assessment and student engagement, but it risks becoming a mere compliance measure if disconnected from teaching goals and staff capacity.

In plain terms

This research tackles the challenge of understanding student motivation and self-regulation using their digital learning behaviors. The authors investigated "delayed start behavior," which measures the time a student waits before beginning an online learning session. They analyzed data from 711 seventh-grade students using the iReady platform, examining if these delays predicted performance on standardized tests. They found that students who frequently delayed starting their math practice sessions tended to perform worse on both math and English standardized tests. Conversely, they identified "early starters" (20% of students) who showed greater academic growth, while "chronic delayers" (another 20%) experienced the opposite trend. This suggests that simply observing start times can provide valuable insights into student engagement and learning outcomes.

In plain terms

Many adaptive math programs used in schools, like i-Ready Math, face a challenge called the "5% problem," where most students don't consistently use them. This paper explores why this happens from the perspective of teachers, focusing on how they understand students' struggles to stick with these learning tools. The researchers conducted a 90-minute workshop with 12 middle school math teachers, asking them to identify problems and vote on key issues related to student persistence. They found four main reasons for low persistence: students' motivation and buy-in, cognitive roadblocks (where they get stuck on content), their ability to bounce back from challenges, and various contextual issues (like classroom environment). Teachers especially emphasized needing help to understand *why* students get stuck and to spot "silent disengagement," prioritizing tools that help diagnose these issues over just tracking how much time students spend on the platform.

In plain terms

Online tutoring platforms in the gig economy often struggle to provide useful feedback to tutors. Traditional feedback from learners can be limited and makes it hard to monitor tutor quality consistently. To address this, researchers explored using AI-powered automated feedback. They created a special tool that analyzed tutors' lessons on Ringle, a popular online English tutoring platform, and then provided feedback. They surveyed 36 tutors, finding that while tutors initially perceived automated feedback more negatively than learner feedback, they still found it useful for checking their own performance and understanding platform expectations. However, discrepancies between the AI's feedback and their own views sometimes caused confusion. Based on these insights, the paper proposes design ideas for future AI feedback systems in education.

In plain terms

Generative AI (GenAI), like ChatGPT, is significantly changing how students learn, especially in subjects like humanities and social sciences (HSS) where writing and interpretation are common. However, there's limited research on its specific effects on these students. To address this, researchers surveyed a large number of HSS students in China. They investigated how students use GenAI, its perceived effects on their learning processes and academic performance, the challenges they encounter, and their preferred ways for GenAI to be included in their courses. Over half the students felt GenAI improved their motivation, independent thinking, and creativity, and a notable majority reported gains in academic performance, though this might partly reflect assessment limitations. Students also raised concerns about ethical use, privacy, accuracy limitations, and becoming too reliant on GenAI, preferring optional course integration with practical training.

In plain terms

Students sometimes get the right answer in math but use flawed reasoning. Traditional automated feedback systems often miss these 'hidden misconceptions' because they only check for answer correctness. The researchers investigated how well AI could spot these issues using over 20,000 real student responses. They found that standard machine learning classifiers were only moderately effective, while more advanced 'open-weight reasoning models' (similar to large language models) were better but often produced many false alarms. To address this, they propose a 'detect-verify-escalate pipeline': the system first flags potential misconceptions, then uses follow-up questions to confirm, and only involves a teacher if necessary. This approach could improve both autonomous AI tutors and teacher dashboards by providing more pedagogically sound feedback.

In plain terms

This paper investigates how student motivation, specifically their belief in their own ability (self-efficacy) and how much they like a learning method (favorability), influences learning from different types of practice. Researchers compared learning outcomes from an Intelligent Tutoring System (ITS) – a computer program offering personalized instruction and feedback – with traditional paper-based practice, using matched problems and a counterbalanced design to isolate effects. They found that students who started with lower baseline self-efficacy achieved greater learning gains regardless of the practice format. Among these students with lower self-efficacy, greater favorability toward the tutor was linked to better learning during tutor practice. Interestingly, ITS-based practice did not significantly improve students' post-training self-efficacy compared to paper methods. These findings suggest that tailoring practice formats to a student's motivational profile could significantly enhance learning.

In plain terms

In science education, students often create drawings to show their understanding, but having expert teachers score these complex visual responses is time-consuming and expensive, making it hard to use on a large scale. This research addresses this challenge by developing an automated system to score student-generated scientific drawings using a type of artificial intelligence called a Vision Transformer (ViT). A Vision Transformer is an advanced machine learning model specifically designed to process and understand images. The researchers also introduced a "confidence-aware" framework, meaning the AI not only predicts a score but also indicates how certain it is about its prediction. This allows high-confidence scores to be automated, while uncertain cases are flagged for human review. They tested their approach on six middle school science assessment tasks and found it improved scoring reliability, providing a practical way to balance automation with the need for accurate and trustworthy educational assessment.

In plain terms

Many digital tools are available for teaching mathematics, but their true effectiveness depends on how teachers select and use them, often without clear educational criteria. This paper highlights a gap between the design of these resources and their actual use in the classroom. To bridge this, the authors propose a framework that breaks down learning theories into observable teaching variables, called metadata. This metadata allows for the detailed characterization of digital resources, specifically focusing on how they address students' prior knowledge, the way concepts are shown, and how new understanding is developed. The ultimate goal is to enable "teacher-support systems" to help educators choose and integrate digital tools more effectively based on strong educational foundations.

In plain terms

Generative AI tutors offer personalized learning support, but a problem arises when premium (expensive) services provide better help than free or low-cost ones, creating an "education inequity." To address this, researchers developed FairTutor, a system that manages how AI tutors respond to students. FairTutor first analyzes a student's question and plans a teaching approach, then attempts to generate an answer using a cheaper AI model. An "evaluator" AI checks the quality of this initial answer, guiding the cheaper model to revise it if needed. Only when necessary for complex problems or insufficient answers does FairTutor selectively use a more expensive, premium AI model, a process called "multi-agent orchestration." This approach achieved 97.1% of the teaching quality of using only premium AI, while drastically cutting serving costs by 71.6%.

In plain terms

This paper investigates how engineering students use and perceive Large Language Models (LLMs), which are advanced AI programs that can understand and generate human-like text, in their higher education studies. The author conducted a survey of 100 university students, mostly in engineering fields, and also reviewed existing literature to understand current trends and challenges. Students reported valuing LLMs mainly for help with writing, clarifying difficult concepts, assisting with coding tasks, and brainstorming new ideas. However, they also expressed significant concerns, including the risk of incorrect information (inaccuracies), potential biases, becoming too dependent on the AI (overreliance), issues related to academic honesty (academic integrity), and the extra effort needed to check the LLM's outputs (burden of verification). The study highlights how students often view LLMs as an all-knowing "oracle" or a perfect "tutor," leading to expectations of expertise and personalized support that these tools frequently cannot fulfill. Ultimately, the paper advocates for a careful, purpose-driven integration of AI, stressing the importance of teaching students how to critically evaluate AI (critical AI literacy) and designing assessments that promote thoughtful use rather than just efficiency.

In plain terms

Understanding how students learn together through "Collaborative Problem Solving (CPS)" is important in education, but there aren't many publicly available datasets of *student-to-student* conversations. Most existing data focuses on teacher-student interactions, making it hard to study how students interact and coordinate on their own. To address this, the researchers created a new dataset called PeerMathDial, which collects real conversations from middle school students working together on math problems in classrooms. This dataset contains 55 dialogues and over 6,000 conversational turns. To help analyze these conversations, they developed a "dialogue act taxonomy," which categorizes different types of speech acts (like asking a question or giving an explanation), assisted by Large Language Models (LLMs). They showed the dataset's usefulness by tracking how dialogues evolve, connecting student traits to their behaviors, and evaluating LLMs' potential for simulating student interactions in educational applications.

In plain terms

High-quality lesson plans often don't fit different student groups or regions due to educational inequality. Teachers usually modify existing plans, but current AI tools focus on creating new content, which adds work. There's a critical need to quickly adapt lessons for diverse student learning profiles. The researchers developed AdaPT, a system that uses Large Language Models (LLMs) to transform existing lesson plans. LLMs are advanced AI programs that can understand and generate human-like text. AdaPT provides an interactive interface where teachers input student profiles, see structured lesson representations, get explanations for why changes are made, and can iteratively refine the adapted content. They evaluated AdaPT with teachers and experts, and results show that the system effectively supports teachers' workflows and helps customize lessons. This could help address educational inequality by making high-quality instruction more accessible.

In plain terms

Large language models (LLMs) are increasingly used as educational tutors, but their ability to solve tasks doesn't automatically mean they are good at teaching. This paper addresses the problem of evaluating LLM tutors by distinguishing between their task-solving ability and their ability to support student learning (pedagogy). The researchers developed a simple diagnostic method that looks at the difference in performance on solving-oriented versus pedagogy-oriented benchmarks. They found that for eight public LLMs, these two aspects are only partially aligned, with a low correlation of 0.421, meaning models can shift ranks when evaluated for teaching quality. They also analyzed existing benchmarks, showing that good teaching behaviors like guiding questions and hints are already valued in rubrics. These findings suggest that evaluating LLM tutors should separately measure task success and learning support to better assess their true educational impact.