We are pleased to announce the awardees selected through the Teaching & Learning (T&L) Benchmarks and Datasets Request for Proposals. Each awardee is addressing a specific use case including: Adaptive Learning Experiences & Feedback for Students, Lesson & Instructional Planning for Teachers, and Teacher Coaching.
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.
Adaptive Learning Experiences & Feedback for Students
Awardee: Portland State University
About the Project
Seeing Through Data will create an openly licensed Kâ12 mathematics corpus and benchmark that tests the full adaptive loop: what an AI system infers about a student’s understanding, which next instructional move it selects, and whether it updates that inference after the student responds. The corpus will contain about 1,200 task instances from Grades 3â12, focused on data reasoning, including linked trajectories for at least 300 students, and models will be scored separately on student-state tracking, adaptive action selection, and feedback quality. An ELA extension will apply the same framework to graphs and infographics in informational texts.
Why it matters
Existing resources capture pieces of this loop but not the whole: they evaluate answers about charts, provide student work without linked feedback and revision evidence, or map problems to skills without testing how a model updates its view of a student or chooses an intervention. Recent tutoring benchmarks also show substantial capability gaps, with frontier models barely beating a trivial baseline on one and no model exceeding 56% on another. This project fills the gap between inferring a student’s state and acting on it.
Public good outputs
- An openly licensed corpus of ~1,200 mathematics task instances with expert annotations
- Composable evaluators for State, Action, Feedback, Safety, and Efficiency
- Protected benchmark items with human-reference performance bands
- A public leaderboard
Although data collection is Oregon-based, the schema, task templates, APIs, evaluators, and reference integration will be curriculum- and platform-portable for national reuse.
Lesson & Instructional Planning for Teachers
Awardee: Teaching Lab
About the Project
A project to build a benchmark for evaluating how well AI models do real lesson and instructional planning, testing whether plans respond appropriately to class evidence, preserve mathematical rigor, prioritize content under time constraints, interpret student work accurately, and match expert-quality plans.
Why it matters
Lesson planning is teachers’ highest-volume use of generative AI, yet the field has no shared way to tell a good AI-generated plan from a convincing but generic one. Existing benchmarks test math-solving or standards-tagging rather than actual planning decisions, and independent testing shows most AI tools, including frontier models, misjudge the cognitive demand of math tasks and weaken the rigor of source materials.
Public good outputs
- A gold-standard corpus of authentic planning episodes paired with the class evidence teachers actually had
- A set of composable evaluators, one for each construct
- A public benchmark and leaderboard built from held-out episodes and run against all major models
The core materials are anchored in IM 360, which is openly licensed (CC BY 4.0), so they can be released publicly.
Teacher Coaching
Awardee: University of Maryland, College Park
About the Project
A project to build the first gold-standard dataset and benchmark measuring AI against the complete instructional-coaching loop in K-12 mathematics. Each session links the artifacts an expert coach reasons over (lesson materials, classroom video and transcript, student work, and prior coach-teacher conversations), and models are scored in stages on analyzing classroom evidence, generating lesson-grounded coaching, and producing coaching that is traceable to evidence, changes when the observed lesson changes, and takes an appropriate stance. An ELA extension will test whether the approach generalizes beyond math.
Why it matters
No existing resource captures the full coaching loop: coaching corpora contain conversations without the observed lesson, and classroom observation corpora contain lessons without coaching. This project fills two gaps: the multimodal classroom evidence that coaching depends on, and a benchmark scored against validated rubrics at both the utterance and conversation level.
Public good outputs
- A corpus of ~1,400 linked coaching sessions (~200 teachers), released as de-identified transcripts, whiteboard captures, student work, and coaching conversations
- Two-layer annotations on 600 conversations (utterance-level coaching moves and conversation-level coaching quality ratings), validated against improvements in classroom instruction and student end-of-year test scores
- A staged benchmark with composable automated evaluators and a public leaderboard
Raw audio and video are retained rather than released; benchmark items that require these recordings will be administered by the project team.