We are pleased to announce the awardees selected through the Open Source AI Model for Tutoring (EDU AI) Request for Proposals.
Proposal review and award monitoring are managed directly by the Gates Foundation and Learning Commons teams. The Digital Promise K-12 AI Infrastructure Program will support communication, participation in the K-12 AI Infrastructure community, and dissemination of public goods. Learning Commons will also support funded projects as a public asset distribution partner, supporting benchmark hosting and dataset deployment.
Awardee: University of Massachusetts Amherst
About the Project
A coalition of researchers and four digital learning platforms (ALTER-Math, ASSISTments, Carnegie Learning, and Eedi) will build an open-source AI tutoring system made up of a core foundation model optimized for tutoring and a customizable agentic workflow that any platform can adopt. The model will be aligned with proven pedagogical principles through pre- and post-training, then trained with reinforcement learning against high-fidelity student simulators built from real student data to directly optimize learning outcomes. The system will be experimentally verified on the partner platforms, which together serve more than 1 million students.
Why it matters
Current frontier models have five major limitations for tutoring: they are not pedagogically aligned (they are verbose and give away answers rather than encouraging productive struggle), they are not trained on authentic student data, they are not optimized for learning outcomes, they may not benefit all students or learning settings equally, and they lack the support needed for wide adoption. This project addresses each of these limitations, while keeping training privacy-preserving so that large amounts of real student data can be used without being released.
Open-source outputs
- A foundation model optimized for tutoring
- High-fidelity student simulators trained on real student data from partner platforms
- Reward models trained on labels from expert tutors
- A customizable, plug-and-play agentic workflow that lets platforms tune the model to their own curricula and student populations
- A production-ready ecosystem with full API and documentation support
Awardee: Trillium Labs
About the Project
A partnership among Trillium Labs (frontier-grade model training), Teaching Lab (K-12 math pedagogy and harness development), and measurement scientists at Harvard University/MQI Coaching and the University of Maryland to create an open-source AI model that is as effective at math tutoring as human experts. The model will be trained to follow a Constitution drafted with Title I math and tutoring experts, replacing the standard “helpful assistant” with a culturally responsive “warm demander” persona that holds high expectations and gets students to do the thinking. Development will be guided by a new, validated evaluation pipeline and synthetic students, and once the model is shown to work and do no harm, it will be piloted with schools serving over 200,000 math students, with classroom learning and engagement data feeding back into training.
Why it matters
The core competencies of tutoring conflict with the commercial incentives of frontier model development: great tutors make students engage in productive struggle, while frontier labs are incentivized to build models that think for users, and research shows frontier models systematically over-scaffold under default prompting. Measuring tutoring quality is also difficult. Recent work has largely resolved the previous scarcity of training data, making the project feasible now.
Open-source outputs
- Open model weights
- Complete post-training and harness recipes
- Evaluation sets
- Annotation rubrics
- Training logs
- Datasets, where permissible