PLUS named a 2026 Tools Competition finalist in the Datasets for Education Innovation track.
See AnnouncementA strategic alliance of world-class universities and industry leaders committed to rigorous learning engineering at scale.
Explore the full archive of PLUS research.
Showing 20 of 20 in this preview (newest first)
Tags
A Festival of Learning workshop convening the National Tutoring Observatory, the SCALE Initiative and the LEVI HAT project around shared data and infrastructure for studying what makes tutoring work.
2026 · Festival of Learning 2026 · Danielle R. Thomas
Matching tutor style to student need — proactive support below the class median, on-demand above it — raised time on task 25%, skill proficiency 36% and standardized test growth 61% over AI-only tutoring.
2026 · EDM 2026 · Ashish Gurung, Ge Gao et al.
Forecasting models predict how much a student will practice and master each week 22–33% more accurately than percentile rules of thumb, giving tutors a basis for realistic goal setting.
2026 · EDM 2026 · Eric S. Qiu, Danielle R. Thomas et al.
Asking for a human tutor raises the odds of getting the next problem right by about 4 percentage points, with the largest gains going to students who knew the least beforehand.
2026 · EDM 2026 · Kirk Vanacore, Danielle R. Thomas et al.
Students spend only 40% of available classwork time actually practicing math. Most of the lost time comes from starting late or stopping early rather than idling mid-task.
2026 · L@S 2026 · Ashish Gurung, Jordan Gutterman et al.
A method for measuring how closely tutor and student talk stays anchored to the problem at hand, tested on 1,576 real math tutoring dialogues.
2026 · AIED 2026 · Conrad Borchers, Jiayi Zhang et al.
The widely used TalkMoves codebook transfers unevenly to one-on-one tutoring: raters agree more often, but it misses moves that a codebook built for AI-human tutoring catches, especially in video and chat.
2026 · AIED 2026 · Corina Luca Focsan, Marie Cynthia Abijuru Kamikazi et al.
Four practical shifts for building trustworthy labeled data in educational AI, arguing that a single inter-rater agreement score is a poor gatekeeper for what counts as ground truth.
2026 · AIED 2026 · Danielle R. Thomas, Conrad Borchers et al.
How long a student takes to start work predicts test performance in both math and English: chronic delayers (over 13 minutes) lose ground while early starters (under 5) gain it.
2026 · AIED 2026 · Jordan Gutterman, Ashish Gurung et al.
A doctoral thesis on adaptive goal support. Weekly practice targets and automated feedback, built into normal classroom routine, raised practice time and skill mastery across six middle schools.
2026 · CMU HCII Technical Report · Conrad Borchers
Twelve middle school math teachers describe why students stop persisting in adaptive software, pointing to silent disengagement that usage dashboards fail to surface.
2026 · ECTEL 2026 · Conrad Borchers
Tutors' scores in online training predict the quality of their real tutoring sessions, linking practice lessons to what actually happens with students.
2026 · ECTEL 2026 · Danielle R. Thomas, Marie Cynthia Abijuru Kamikazi et al.
Once you account for how much tutors learn during training, GPT-4's scoring holds up as reliable — and open-ended questions beat multiple choice.
2026 · Journal of Learning Analytics · Conrad Borchers, Danielle R. Thomas et al.
A full-day LAK 2026 workshop where participants annotate real tutoring dialogue with the Sandpiper tool to identify which teaching moves actually drive learning.
2026 · LAK 2026 Workshop · Danielle R. Thomas
Short human check-ins lift student engagement during and after the visit regardless of length, so several well-timed brief contacts beat one long one.
2026 · LAK 2026 · Conrad Borchers, Ashish Gurung et al.
A tool that maps the routes students take through algebra problems. Students who explored more dead ends scored higher on conceptual and procedural knowledge.
2025 · Educational Technology Research and Development · Siddhartha Pradhan, Erin Ottmar et al.
Over-reliance on rater agreement scores holds educational AI back, so this paper proposes five alternatives that test whether labels predict what students actually learn.
2025 · AIME 2025 · Danielle R. Thomas, Conrad Borchers et al.
Breaking lesson writing into sub-tasks produced better AI-generated tutor training lessons than a single prompt, though reviewers still found the feedback generic.
2025 · ECTEL 2025 · Jionghao Lin, Jiarui Rao et al.
A survey of 100 US teachers: newer teachers want analytics on classroom behavior, while veterans want sustained data on student learning — so reflection tools should adapt to experience level.
2025 · ECTEL 2025 · Ann-Christin Falhs, Conrad Borchers et al.
Adaptive time-management support helped 89 adult online learners study more regularly, and that regularity was linked to better learning outcomes.
2025 · ECTEL 2025 · Katharina Teich, Conrad Borchers et al.
Meet our team who are driving evidence-based breakthroughs in learning science, HCI, and artificial intelligence.
Dr. Ken Koedinger
Principal Investigator, Professor @HCII
Carnegie Mellon University
Dr. Danielle Thomas
Research Lead, Systems Scientist @HCII
Carnegie Mellon University
Want to get involved? Reach out if you are interested in conducting research with us.
