PEER-REVIEWEDPhysical Review Physics Education Research, June 2025

Kudu’s AI hints were associated with higher exam scores.

A peer-reviewed study from researchers at UCLA and UC San Diego examined two interventions deployed in Kudu: scaled extra credit on supplemental math, and AI-generated hints embedded in physics homework. Both interventions correlated with higher exam scores and reduced disparities in completion rates among students from groups historically underserved in introductory STEM.

Study
AIMS — Artificial Intelligence and Math Skills
Lead author
Yifan Lu (UCLA Physics & Astronomy)
Published
Phys. Rev. PER, June 2025
Funder
UCLA Teaching & Learning Center
  • 88%

    of students used the AI hints on at least one homework question, when hints were available.

  • p = 0.004

    AI hint use was significantly associated with higher exam 2 scores.

  • 188 and 194

    students, in the study’s two sections of calculus-based mechanics.

THE AIMS STUDY AT A GLANCE

One study. One question. Real classrooms.

Researchers at UCLA's Center for Education Innovation and Learning in the Sciences (CEILS), partnered with UC San Diego, set out to test whether two specific interventions — scaled extra credit and AI-generated hints — could close the gap between students who arrive in introductory physics with strong math preparation and those who don't.

Setting

Intro Physics

Undergraduate course at a major research university. AI hints embedded directly in homework via Kudu.

Question

Do these interventions help?

Specifically, do they help most the students least likely to ask a TA, peer, or instructor for support?

Design

Two interventions, controlled

Scaled extra credit on supplemental math, plus AI-generated hints in physics homework. Both optional, both available to every student equally.

Results

Both worked. The effects are concrete.

Higher exam performance for students using the interventions. Fewer completion gaps among students historically underserved in STEM.

THE HEADLINE FINDING

The students who needed help the most got the most benefit.

Controlling for incoming physics understanding (measured by Force Concept Inventory score) and other variables, students who completed the supplemental math assignment scored higher on subsequent exams. The effect was largest on the third exam — the exam most aligned with the supplemental material.

Crucially, the effect was strongest for students who entered the course with weaker math preparation. Black, Hispanic and Latine, Native American, and lower-income students disproportionately benefit, because they're statistically less likely to have had access to advanced math courses in high school.

"Using AI tools can remove social barriers that prevent students from asking instructors, teaching assistants, and/or peers for help."
- Yifan Lu, lead author, UCLA Physics & Astronomy
Read the full paper in Phys. Rev. PER ↗
Figure 6 from Lu et al. (2025): exam 2 score against Force Concept Inventory score. Students who used AI hints scored higher, most of all those with the lowest inventory scores; the two lines meet near the top.
Figure 6 from Lu et al. (2025): exam 2 score against Force Concept Inventory score, separated by use of AI hints. The colored bands show 95% confidence intervals. Lu et al. (2025), CC BY 4.0, cropped.

WHAT WAS TESTED

Two interventions, deployed together in one course.

Both interventions are grounded in Situated Expectancy-Value Theory of Achievement Motivation — students engage more in tasks they expect to succeed in and perceive as valuable.

Intervention 01

Scaled extra credit on supplemental math.

Optional supplemental math assignments were offered to all students. Extra credit was scaled so that students with lower exam scores could earn more from completing the assignments — making the effort proportionally more valuable to the students who most needed the boost.

This addresses the "gifts for the gifted" problem common to traditional extra credit, where high-performing students receive marginal benefit and lower-performing students receive a structural reward for catching up.

54% vs. 19%

Completion rates among students below the class FCI mean — with scaled incentives vs. without.

Intervention 02

AI-generated hints in homework.

Optional AI-generated hints were embedded directly in physics homework problems on Kudu. Hints were grounded in the chapter content the student was studying — so a student stuck on a problem could ask for guidance without leaving the course context, and without the social cost of asking a TA or peer.

The hints are explicitly designed not to provide solutions. They guide students through the reasoning rather than handing them an answer.

88%

Of students used AI-generated hints on at least one homework problem matched to an exam question.

More on AI hint usage

88%

Used hints on at least one matched problem

57%

Used hints on at least five problems

~20%

Average share of questions where hints were used

IN THEIR WORDS

What the researchers said.

By implementing scalable and low-cost interventions, we can create more equitable learning environments. Using AI tools can remove social barriers that prevent students from asking instructors, teaching assistants, and/or peers for help.
Yifan LuLead author — Doctoral student, UCLA Department of Physics & Astronomy
Our findings demonstrate that thoughtful integration of technology and targeted incentives can make a meaningful difference in student academic performance. These strategies are effective and feasible for widespread adoption in a wide range of undergraduate courses that include mathematical topics.
Elizabeth H. SimmonsCo-author — Executive Vice Chancellor and Distinguished Professor, UC San Diego

WHERE THE RESEARCH GOES NEXT

AIMS is expanding to more courses.

Following the publication, UCLA's Teaching and Learning Center funded an expansion of the AIMS program to additional Physical Sciences and Life Sciences courses over two academic years. Faculty in chemistry and life sciences are now adapting the supplemental math materials to their own course contexts and embedding AI-generated hints in their homework on Kudu.

The goal: increase use of these interventions specifically among students who could most benefit, and reduce DFW (D, F, Withdraw) rates and grade disparities across the science core.

This is the part that matters for institutions evaluating Kudu. The research isn't a one-off; it's the foundation of an ongoing program.

Courses adopting AIMS, 2025-2027
CourseDepartment
Physics 1 seriesPhysics & Astronomy
Chemistry 14 seriesChemistry & Biochemistry
Chemistry 20 seriesChemistry & Biochemistry
LS 30ALife Sciences Core

Expansion funded by UCLA's Teaching and Learning Center. Per the CEILS program description, instructors using the tools in subsequent iterations face minimal additional implementation labor.

KUDU'S ROLE, IN PLAIN LANGUAGE

Kudu is the platform. The research is independent.

Kudu provided the course platform on which AIMS was deployed. The AI-generated hints were delivered through Kudu. The supplemental math assignments were administered via Kudu. Student usage data came from Kudu.

The research itself is independent: designed, conducted, analyzed, and published by researchers at UCLA's Center for Education Innovation and Learning in the Sciences and UC San Diego. Funding came from UCLA's Teaching and Learning Center.

What that distinction matters for: when you read claims about Kudu's AI improving learning outcomes, those claims rest on peer-reviewed work that wasn't led, sponsored, or paid for by us. We helped build the tools, but we didn't grade the homework.

Problem 4.3

A block of mass 5 kg slides down a frictionless incline of angle 30°. Find the acceleration.

KAI Hint

Let's break this down step by step.

  1. Identify the forces acting on the block.
  2. Resolve gravity along and perpendicular to the incline.
  3. Use Newton's second law along the slope.
A Kudu AI hint, as it appeared to students in the AIMS study. Hints guide reasoning rather than providing solutions.

READ THE WORK

Citations and sources.

PRIMARY CITATION

Lu, Y., Supriya, K., Shaked, S., Simmons, E. H., & Kusenko, A. (2025). Incentivizing supplemental math assignments and using AI-generated hints is associated with improved exam performance. Physical Review Physics Education Research.DOI: 10.1103/v93w-ytyc ↗

WHERE TO NEXT

Different doors for different readers.

For institutions

Map this evidence to your campus.

A 30-minute mapping call with your teaching & learning office: how the AIMS interventions could apply to the courses you already run, and what the program would cost.

For institutions

For faculty

Try AI hints in your own course.

AI-generated hints are built into every Kudu course. Start a free Kudu course in your subject in minutes and see how they work in your homework.

For faculty

For press & researchers

Use this page as a citation.

This page is the canonical source for the AIMS study and Kudu's role in it. Stable URL. Updates tracked. Reach out if you'd like to discuss replication or expansion studies.

Contact us