Using Technology to Enhance Math Learning

In 2012, a tiny YouTube channel took on the entire U.S. education system. And schools ignored it—until STUDENTS made it go viral. Here's how one man revolutionized modern learning: A hedge fund analyst, tutoring his cousin after work, started recording simple math videos from his closet. What started as a family favor revealed a fundamental truth about learning. But first, he faced massive resistance... The education establishment was completely dismissive: "Where's your teaching degree?" "This is too simplistic." Yet while the experts debated, students were voting with their clicks. That's when Sal Khan and Khan Academy proved what's possible in education. Students weren't just watching videos—they were mastering concepts they'd struggled with for years. The platform spread like wildfire through study groups and social media. Why? Because it solved a fundamental problem in education: Forcing 30 students to learn at the exact same pace. And Khan Academy shattered this model. For the first time, students could truly master concepts before moving forward. Then Los Altos School District dared to experiment in 2011. They flipped everything: video lectures at home, problem-solving in class. The data became undeniable. State math exam scores jumped 106%. Not for some students—for everyone. But this wasn't just about better test scores. Students were actually ENJOYING learning again. The platform aligned with standards while revolutionizing delivery. By 2017, 85% of U.S. school districts saw this vision and embraced it. Then came the next evolution: AI. In 2023, Khanmigo emerged—not just another AI tool, but a Socratic tutor powered by GPT-4. Instead of replacing teachers, it amplified their impact. And the results are exceeding everyone's expectations: The pilot data tells a compelling story: • 85% of students gained deeper understanding • 92% of teachers reclaimed precious time • Problem-solving skills soared by 30% We've proven learning doesn't have to be a source of stress and anxiety. When we meet students where they are—truly where they are—they soar. The old industrial model of education is crumbling. And it's being replaced by a system that celebrates critical thinking and individual growth. Where students develop real-world skills at their own pace. 150 million registered users 190 countries 36 languages and growing But these aren't just numbers—they're proof every student can excel when given the right tools. They can go farther than the current system allows. Quality learning is accessible to everyone. AI tutors support each student's unique journey. Engagement replaces enforcement. The future of education is happening now.

A Survey of Mathematical Reasoning in the Era of Multimodal Large Language Models: Benchmark, Method & Challenges ... Mathematical reasoning is a fundamental cognitive skill that underpins learning across disciplines. As we explore the integration of artificial intelligence in this area, a new research paper titled "A Survey of Mathematical Reasoning in the Era of Multimodal Large Language Models: Benchmark, Method & Challenges" offers valuable insights. This comprehensive survey reviews over 200 studies published since 2021, dissecting the current state of mathematical reasoning capabilities of Multimodal Large Language Models (MLLMs). Here are some key takeaways worth discussion: 👉 1. The Shift Towards Multimodal Reasoning The integration of diverse data types—text, diagrams, and visual elements—is critical for addressing complex mathematical problems. Traditional models often excel with textual input but struggle when visual information is introduced. This shift highlights the necessity for MLLMs to enhance their reasoning capabilities across modalities to accurately interpret and solve mathematical tasks. 👉 2. Current Landscape of Math-LLMs The paper sheds light on significant advancements in models such as GPT and Minerva, which have made strides in handling complex mathematical reasoning tasks. These models now operate within multimodal settings, enabling them to process and analyze various input forms, thus improving their efficacy in mathematical problem-solving. 👉 3. Key Challenges in MLLMs Despite the strides made, several challenges persist: - Insufficient capabilities in visual reasoning. - Limitations in generalizing across different mathematical domains. - Ineffective error detection and correction mechanisms. - Gaps in adapting to real-world educational contexts. Addressing these challenges is crucial for the ongoing development of robust MLLMs. 👉 4. Implications for Educational Technology The research suggests that advancements in MLLMs can revolutionize educational tools, offering personalized support tailored to individual learning needs. By leveraging enhanced reasoning capabilities, educational technologies can provide students with tools that adapt in real-time, fostering deeper understanding and engagement with mathematical concepts. 👉 5. Future Directions in Research Looking ahead, the paper outlines essential areas for research, including refining model architectures and expanding datasets to encompass a broader array of mathematical problems. Collaborations across disciplines will be vital in pushing the boundaries of MLLM capabilities, ultimately enhancing their performance in educational contexts. 👉 Conclusion As industry professionals, engaging with this research can catalyze innovations in how we approach teaching and learning mathematics.

The University of Chicago Education Lab recently released a fascinating study of Saga Education, a high-dosage tutoring model. The study found that substituting some tutor time with educational technology can reduce costs by one-third and halve the number of tutors needed without compromising effectiveness. This is an excellent read for anyone building new AI-powered tutoring products or school leaders exploring program design to better leverage strategic staffing options. 📌 The paper highlights the results of a 4K student RCT that tests the effects of a 4-to-1 tutoring model in which four 9th-grade students sit at a table with one in-person tutor, and the students alternate days working either with the tutor in student pairs (2:1) or working on computer-assisted learning for the entirety of a class period (50 minutes). 📈 Overall, the gain in students' math scores was equivalent to between three-quarters and one full year of additional learning over the program year. ⚖ Compared to everyday tutoring, incorporating technology reduces costs by 30%, reduces the number of tutors required to serve a given number of students by 50%, and has almost as large an effect on student learning. 💡 What I found interesting is how the paper explores why this hybrid approach works. The study estimates that 2/3 of the total learning gains could be attributed to high-quality computer-assisted learning (CAL). The study also showed that, on average, students used the program for about 30 hours of the total possible 45 hours in the program design, which signals a good amount of program adherence to me. 🔍 My interpretation of this is that the impact of reducing direct in-person instruction time is not linear. While the standard tutoring model and the hybrid model produced nearly identical results, most of the gains in the hybrid model could be associated with the support of technology. This indicates a likely compounding positive impact of daily in-person tutor time. That said, the program's structure likely produced higher adherence to recommended tech usage for students, and the combination of the two modes of learning offers a compelling lower-cost and scaleable model for instruction. "What tutoring can do that CAL cannot is provide human connection. This human connection between the tutor and the student might help to sustain student engagement and motivation. It follows that the absence of a human connection may be one reason why there seems to be diminishing marginal returns to student time spent on CAL (Bettinger et al., 2023)." https://lnkd.in/gMyScSWX #EducationInnovation #FutureOfLearning #EdTech #K12 #aiineducation #genai Overdeck Family Foundation

As Kunjan Narechania reminds us, tracking usage of edtech isn’t enough. We need to connect usage data to learning outcomes. It’s refreshing to see this conversation gaining traction as the field doesn’t talk enough about dose: the amount of use a program requires to drive learning, backed by data. The next step for the field: go beyond identifying a dose that works and study dose–response: how impact changes with more or less use. Districts and states need this insight to set edtech usage goals and support educators in using edtech for impact. At Zearn, we’ve spent the past decade studying how lesson completion relates to learning across states and districts. In 12 state and district studies, students completing 90+ grade-level lessons per year consistently show the strongest growth in math achievement, which equates to roughly 3 grade-level Zearn lessons per week. While this remains our recommended dosage, a recent quasi-experimental study from Johns Hopkins University (2023–24) found a statistically significant +0.20 SD impact on Louisiana’s LEAP math scores for students in districts averaging 60+ grade-level lessons per year — demonstrating that meaningful impact occurs within levels of usage in our dosage funnel. These quasi-experimental findings complement the causal evidence base behind Zearn’s ESSA “Strong” (Tier 1) rating, adding real-world evidence of how impact scales with use across diverse contexts. At Zearn, studying dose response isn’t just research. It’s part of our nonprofit mission to ensure all kids have access to high-quality math learning to  drive impact.

📉 Learning outcomes in Côte d'Ivoire remain low with only 17% of students reaching proficiency in mathematics, and nearly half of grade 4 students not able to read a simple sentence. Through the World Bank's Youth-RISE project, supported by the Mastercard Foundation, we piloted AI-powered adaptive learning platforms across 25 TVET institutions with approximately 2,000 students. 📊 Impact analysis shows active users gained 0.234 standard deviations in mathematics (about 11 months of learning) and 0.121 standard deviations in French (about 6 months). ⚡ The most striking finding: struggling learners in the bottom 15% progressed 6 to 15 times faster than average performers, showing how adaptive technology can meaningfully reduce educational inequalities when students actively engage with it. https://lnkd.in/dvemmuMd

Over the past two weeks, I have been focused on building AI powered tools that strengthen STEM learning, culminating in a project that brings multi agent systems directly into the hands of educators. As part of Harvard Graduate School of Education's intensive J Term program, 𝘉𝘦𝘤𝘰𝘮𝘪𝘯𝘨 𝘈𝘐 𝘓𝘪𝘵𝘦𝘳𝘢𝘵𝘦 𝘧𝘰𝘳 𝘵𝘩𝘦 𝘍𝘶𝘵𝘶𝘳𝘦 𝘰𝘧 𝘌𝘥𝘶𝘤𝘢𝘵𝘪𝘰𝘯, I partnered with teammate Dorrie Nord to develop SpatialMind, an AI driven toolkit designed to help educators teach spatial reasoning, a foundational predictor of success across STEM fields. Our work centered on applying RAG pipelines, fine tuning strategies, safety guardrails, red teaming, and multimodal, multi-agent orchestration using OpenAI Agent Builder and Google AI Studio to design a product aligned with real classroom needs, grounded in learning science. SpatialMind includes a Teacher Hub that uses a multi agent debate framework to generate richer, more creative lesson plans. Each agent contributes a distinct perspective, such as a teacher, teaching assistant, toy designer, play expert, and learning designer, as they critique and refine one another's ideas through structured rounds of discussion. The Student Zone features two interactive 3D learning experiences, Rotation Match and Shadow Sculptor, which help students build mental rotation skills through hands on practice. Here's a short video example of Shadow Sculptor: https://lnkd.in/eBUGWiy6 This project brought together my teammate's mathematics teaching expertise and my background in spatial research to create a tool that bridges pedagogy, cognitive science, and applied AI. The experience strengthened my ability to translate research insights into practical product features, collaborate across domains, and build AI driven workflows that support real users. Thank you to Professor Ying Xu and our teaching fellows Zhonghao Shi, Bharath Sriraam Ramasubbu Raja, Kelly Ding, Stacey Cho, and Srushti Jayaramu for fostering a collaborative environment that challenged us to think boldly about the future of AI supported learning. I am excited to build on this momentum and continue developing tools that meaningfully support educators and students. A link to the full presentation can be found in the comments. #GoogleAIStudio #AIinEducation #EdTech #HGSE #SpatialReasoning #STEMeducation #OpenAI #GoogleAI #Gemini #MultiAgentAI #AIAgents #LearningScience #EducationInnovation #MultimodalLearning #MathEducation

We are excited to announce the release of our "Guide to Integrating Generative AI for Deeper Math Learning" - the second guide from the collaboration between AI for Education and Student Achievement Partners. This guide is an initial exploration of how to harness GenAI and other AI applications to enhance, not replace, the cognitive lift and meaningful learning in math classrooms. Key Highlights: • Strategies for enhancing productive struggle and lessons counterproductive struggle in mathematics education • Recommendations for when to use Large Language Models like ChatGPT or math-specific AI tools like (Desmos, Snorkl, and Khanmigo) both in instruction and planning • Practical use cases for both students and teachers with detailed examples, instructional benefits, and areas of caution We learned an enormous amount through this process as commercial GenAI chatbots still struggle with computation, and cannot be relied upon for math instruction without significant educator oversight and AI literacy. We will continue to update this resource with emerging research, new technological advancements, and feedback from math educators. Link in the comments to learn more or download a PDF copy of this guide. You can also find the link to the guide on Integrating GenAI for Deeper Literacy Learning. #Aiforeducation #aieducation #Math #GenAI

It might be cherry picking of evidence but are worth exploring anyway: AI Tutors: Hype or Hope for Education? Is B. F. Skinner back? - Tutor CoPilot (AI System): Improved student mastery by 4 percentage points in a randomized controlled trial with 1,800 students. The greatest benefit was observed for lower-rated tutors (9 percentage points improvement), helping them achieve outcomes similar to more effective peers.    - Harvard AI Chatbot Tutor: Found that students using a custom-designed AI chatbot in a physics course showed double the learning gains and significantly higher engagement compared to traditional classrooms. Personalized feedback and self-pacing were key benefits. - Rori (AI Math Tutor in Ghana): Improved math growth scores in students using it for one hour per week, with an effect size equivalent to an extra year of learning. Its low cost ($5 per student) made it an effective intervention in resource-limited settings. - Bridge Method with GPT-4: AI responses to student math mistakes were rated 76% better when guided by expert teacher decision-making data, highlighting the importance of expert knowledge in AI tutoring. - AI Tutor vs. Active Learning Classroom: In a controlled study, students using an AI tutor showed significantly greater learning gains in less time (49 minutes vs. 60-minute lecture). They also reported higher engagement and motivation, with 83% rating AI explanations as good as or better than human instructors’. - AI as Education Expert: AI tutors successfully replicated teaching principles and created math worksheets that aligned with teacher judgments, suggesting potential to speed up lesson design while still requiring human expertise for real student testing. - GPT-4 Field Experiment (Turkey): AI tutoring improved math performance, but the effects were reversed when AI access was removed (17% reduction), showing that GPT-4 could be used as a crutch unless safeguards are in place to ensure continued human learning. Full details https://lnkd.in/eRUxP3SR

Some very happy AI research was published this week 🥰 : AI increases students higher-order thinking and overall learning performance! This week at AITIA Labs I prepared a presentation for education executives. I came across an eye-opening meta-analysis in Nature. The research reviewed over 50 studies about ChatGPT in classrooms, and here is what they found: Contrary to widespread concerns, the data shows that ChatGPT usage improves student learning outcomes. Especially interesting was the improvement in higher-order thinking. Higher-order thinking is typically what people are worried about when they talk about "AI is making students dumber". High-order thinking is the ability to analyze, create, and self-reflect. The study showed a substantial effect size (Hedges g of 0.46) for higher-order thinking improvements. At those effect sizes you will see a noticeable difference in classroom settings. It is important to get the setting right, as to get those positive results, you can't just give students access to ChatGPT, you have to design the learning experience correctly, here is what worked: 1. Using ChatGPT to stimulate discussion and critical thinking rather than as a simple answer machine 2. Explicitly teaching students how to question and critically interact with AI-generated responses 3. Incorporating reflective activities that prompt students to evaluate AI outputs 4. Combining AI-assisted learning with traditional instruction to reinforce understanding The question isn't whether AI belongs in education, but how we can optimally integrate it to enhance what already works! Have you tried AI for trainig or education? Anything you notice that works or doesn't? I post regular info on AI, follow to stay in the loop.

🎙️ Amplifying Learning Through Student Voice with Snorkl In Episode 268 of My EdTech Life, I had a great conversation with Jeff Plourd and Jon Laven, the founders of Snorkl, about their mission to transform education by harnessing the power of student's voice. Snorkl's innovative platform allows students to record verbal explanations of their problem-solving process, such as walking through how they determine the width of a rectangle, given its perimeter and length. By capturing students' spoken thoughts, Snorkl creates powerful tools for personalized learning and deeper engagement. Their AI analyzes each student's response and provides timestamped feedback, helping students solidify their understanding and catch their own mistakes. This technology supports students' individual learning journeys and empowers teachers with automatic scoring tools. Discover how Snorkl amplifies learning and transforms math education by giving students a voice.