The research
The unglamorous research behind why Upwise actually works.
The signals
These studies point in the same direction: guided tutoring, mastery progression, and efficient feedback loops outperform generic practice.
2x
A carefully designed AI tutor more than doubled median gains in a Harvard-led RCT.
What this means for families
For a child, this points to support that actually helps them think, not just more questions piled on top of them.
27%
Students moved faster with AI teaching support while pass rates held up.
What this means for families
For parents, this suggests better learning does not have to mean long, draining sessions after school.
107
The large ITS meta-analysis pooled evidence across a broad base of studies.
What this means for families
For families, this means adaptive tutoring is supported by a real research base, not just marketing language.
14.3k
This is one of the clearer scale signals behind intelligent tutoring systems.
What this means for families
For parents, this points to a model that can bring more personalised support to many children, not only the few who can access one-to-one tutoring.
The shape of progress
Source thread: Bloom (1984), Ma et al. (2014), Letourneau et al. (2025)
Mastery path
Support stays tight until the skill is actually stable.
In Upwise terms: the child gets another pass, another hint, or another check before the gap has time to spread.
Move-on path
The class moves along even when the base is still shaky.
This is the risk Upwise is built to avoid: content can look “covered” while the child still needs another clean repetition.
Why it works
Upwise is trying to do three things well at once: close gaps before they spread, keep the next step personal, and use guidance without taking over the learning.
TL;DR
The research keeps rewarding the same pattern: mastery first, useful feedback second, and calm consistency over frantic volume.
Weak foundations do less damage when a child moves on only after the current skill is genuinely steady.
Decades of evidence point the same way: tutoring works. The hard part has always been cost and scale.
The best systems adapt difficulty, revisit weak spots, and respond to the actual pattern of errors in front of the child.
Repeated signal
The literature keeps rewarding systems that wait for understanding before moving on.
Repeated signal
Students improve faster when support adapts to the exact error pattern in front of them.
Strong signal
Hints, corrections, and worked guidance help most when they recover a child’s thinking.
See it in action
The point of all this research is not to sound clever. It is to build a calmer, sharper alternative to endless worksheets and generic tutoring.
The mastery model
Mastery is not just “more practice.” It is a pacing decision: stabilise the idea, revisit the weak spot, and only then let the next concept matter.
Start with the skill the child can almost do, then make sure it becomes genuinely steady.
Upwise behaviour
Upwise keeps the concept active instead of rushing ahead.
Use targeted questions and feedback while the misunderstanding is still small and repairable.
Upwise behaviour
That is why hints and rechecks sit inside the flow.
Only move on when the skill still holds up after a little time and a little variation.
Upwise behaviour
That is why mastery in Upwise is earned, not assumed.
The comparison
Understanding over sessions
Early correction steepens the learning curve.
Sessions
Source framing: Bloom's mastery model, reinforced by later tutoring and ITS evidence. The point is not the exact curve. The point is that early correction changes the slope.
Pace
Moves on after understanding
Moves on when the timetable says so
Gaps
Weak spots get revisited early
Weak spots often stack quietly
Feedback
Tight loops and targeted correction
Broader, slower correction
Confidence
Challenge stays productive
Students can become bored or lost
Guardrailed AI
The stronger recent papers are very consistent on this. AI helps when it keeps the next teaching move close, fast, and supportive. It hurts when it turns into a shortcut machine.
Guardrailed AI
Start with diagnosis, not guesswork.
Keep difficulty close to the child’s current level.
Support thinking instead of short-circuiting it.
What good looks like
Prompt the child, keep the next move manageable, and let them do the real work once the hint lands.
What to avoid
Answer-giving can make practice feel easier while weakening later independent performance. That is exactly what the guardrails are there to prevent.
Evidence link: Bastani et al. (2025) on hints-before-answers, plus Wang et al. (2024) on AI support improving the quality of tutoring moves.
Evidence appendix
Start with the most recent papers, then skim the rest of the shelf if you want the broader research stack behind the product.
Bastani et al. (2025)
Unrestricted AI improved practice performance but hurt later exam performance. A guardrailed tutor using hints instead of answer-giving largely mitigated the damage.
What this means for families
This is the strongest case for Upwise's hints-before-answers philosophy.
Kestin et al. (2025)
A carefully designed AI tutor more than doubled median learning gains compared with in-class active learning, while also taking less time.
What this means for families
It shows that well-guardrailed AI tutoring can improve both learning quality and efficiency.
Letourneau et al. (2025)
In K-12 studies, AI-driven intelligent tutoring systems usually outperformed traditional teacher-led comparison conditions, especially when they used personalisation, adaptive feedback, and mastery progression.
What this means for families
It is one of the most direct matches to the product category Upwise sits in.
The rest of the shelf
More studies sitting behind the same thesis: mastery, adaptive tutoring, better feedback, and better guardrails.
Yamkovenko (2024)
Khan Academy efficacy at scale
Students meeting the recommended usage threshold achieved about 20% greater-than-expected learning gains, with stronger year-over-year gains as more skills were mastered.
Nickow et al. (2024)
Nickow et al. tutoring meta-analysis
Tutoring remains one of the largest-impact educational interventions, with stronger results when sessions are regular, personalised, and frequent.
Moller et al. (2024)
Time efficiency from AI tutoring
Students using an AI teaching assistant progressed faster through their courses, cutting effective study time by roughly 27% while preserving pass rates.
Wang et al. (2024)
Tutor CoPilot field RCT
Students whose tutors received real-time AI pedagogical support were more likely to master topics, largely because tutors shifted from answer-giving to guiding questions.
Ma et al. (2014)
Ma et al. meta-analysis on intelligent tutoring systems
Across 107 effect sizes and 14,321 participants, intelligent tutoring systems performed on par with human one-to-one tutoring and outperformed teacher-led large-group instruction.
Dunlosky et al. (2013)
Learning science for adaptive education
The strongest modern evidence supports retrieval practice, distributed review, informative feedback, reduced extraneous load, and carefully constrained adaptivity.
Bloom (1984)
Bloom's mastery learning foundation
The foundational argument is that most students can reach high levels of achievement when they get enough time, feedback, and corrective instruction before moving on.
