Projects & Research

Learning from Student Thinking: Building a Dataset of Diverse Mathematical Reasoning from Additive to Multiplicative Reasoning will produce a proof-of-concept public good for the Digital Promise K–12 AI Infrastructure. In grades 3–5, students’ transition from additive to multiplicative reasoning is a high-leverage learning progression that strongly predicts later success in proportional reasoning and algebra. Yet, it remains difficult for many students and challenging for educators to interpret in real time. AI-enabled formative assessment tools could expand teachers’ capacity to notice and interpret patterns in students’ strategies and representations, strengthening valid, equity-minded instructional decision-making. However, existing public datasets are scarce and rarely combine multimodal evidence of reasoning, learning-science-grounded annotations, and documentation for reuse. This is especially so for neurodivergent learners who may communicate mathematical thinking through varied combinations of speech, drawing, and partial verbal explanations. The Maine Mathematics and Science Alliance and TERC will partner with teachers to collect formative fluency interviews with students. Each interview consists of approximately 12 unique items, translating to 2,400 student reasoning episodes in the dataset. Data for each task will include redacted interview transcripts, reconstructed/redacted student work artifacts tied to interview tasks, structured teacher observation fields, and iPad stroke-log traces for interviews conducted on iPads. Researchers will produce segment-level labels aligned with Evidence-Centered Design, including additive/multiplicative strategy indicators, representation features, representation–talk coordination, instructional move tags, and explicit uncertainty/abstention conventions. The project team will also create and publish a single proof-of-concept benchmark task to demonstrate the viability of this data to create AI benchmarks. AMR Pilot will test the feasibility of scalable collection, conservative public de-identification, and reliable, rule-governed segment-level annotation as prerequisites for building a high-quality public dataset.