Math Adapt AI

What you do in Math Adapt AI

Math Adapt AI is our patent-pending adaptive engine. It watches how you answer, not just whether you are right. It is not a workout of its own: it works inside Add & Subtract, Multiply & Divide and I'm Feeling Clever.

• In a round it brings in more of the kinds of sums that slow you down, at the same level.
• Between rounds it offers a step up when you are fast and right. You can say no, or pick any level yourself.
• Your Progress screen shows the facts you know and the ones to practise.

Doc explains

The science idea

A practice program can only be adaptive if it keeps a running picture of the learner. One well known example, the Dutch Math Garden practice site, updates its estimate of each child, and of each sum, after every single answer, then picks sums the child has a good chance of getting right.¹

How hard should practice be? Scientists who modelled how learning systems learn fastest found a sweet spot near 85 right out of every 100. It comes from mathematics and computer models of learning, not from tests with children, but it makes a neat target.²

What the research found

In computer models, a broad family of learning algorithms making two way choices learned fastest when training was set so that about 85% of answers were right. The authors showed this in artificial neural networks and in a model of human and animal perceptual learning, and they say that testing it directly with real learners is future work.²

Math Garden aims for a 75% chance of a right answer. Over ten months, 3,648 children answered more than 3.5 million sums, about a third of them outside school hours.¹ When researchers set the same program to different success rates for six weeks, 207 children in grades 3 to 6 who got the easier, higher success settings attempted more problems and improved their math more, and math anxiety fell about equally in every group. The authors suggest the extra practice that comes with more success, rather than one exact percentage, drove the gains.³

Adaptive is not a magic word. A meta-analysis of intelligent tutoring systems in school math found an average effect from about zero to small (g = 0.01 to 0.09), with a smaller benefit for low achievers than for students in general.⁴ Another review of 50 evaluations found a much larger median gain, 0.66 standard deviations, but the size depended a lot on whether the test was written to match the lessons or was a standardised test.⁵

How Math Adapt AI uses it

• Inside a round, the engine watches only your first try at each sum. If you get a kind of sum wrong, or answer it slowly compared with your own usual speed, it quietly swaps a few of the sums still to come for more of that kind, at the same difficulty. That is adaptive practice in small steps.¹
• Guard rails keep it gentle: it only acts in the middle of a round (from 35% to 80% of the way through), at most twice a round, never rewrites more than 35% of the round and never puts more than 2 of one kind in a row. These are the app's design values, not numbers from a study.
• "Slow" is personal: your speed is compared with your own history, not with other players or a fixed time.
• The difficulty autopilot: after 3 rounds in a row of a mode with no wrong answers, at about your usual speed or faster, Doc offers one level up, at most once every 20 hours for that mode. You can accept, say not yet, or set the level yourself.
• The same idea runs Math Run's Practice runs, a separate system: there the next sum gets a little harder after a right answer and easier after a miss, so it settles near 84% right first time, close to the 85% from the computer models.² The classic workouts do not aim at a fixed percentage.

What this does and doesn’t mean