Deep down in the report it is revealed that that 5.2% consists of a third of the class improving by 10% or more... and the rest remaining essentially flat. So it appears to me that it's particularly effective for a subset of learners.

Also note: the teacher doesn't have to really lead the class. The report says that the teacher spent most of their time 1 on 1 helping students who were stuck. Teachers may see value in simply having to do less work to achieve the same (or slightly better) result.

Also, this class was only a couple months. There is a question whether the small improvement would have continued and would have become wider when applied over a whole school year.

The class only took 5 weeks (24X2 hours sessions).

6.4/5.2 = 1.23 .

Is 23% improvement considered that small(assuming it's statistically significant) ?

I think you're looking at the wrong derivative. This is 23% improvement of improvement. It's still just 6.4% improvement.

Might be easier to see if you think of it as 106.6/105.2, in order to include the baseline.

I'm not sure 106.6/105.2 is the correct way to look at it either.

Intuitively, every point you got in a test corresponds to a micro-subject.

So before the class , you knew 100 micro-subjects(which probably took a few years to teach). And you gained 6.4 or 5.2 micro-subjects during the 5 week class.

Since one class taught 6.4 micro-subjects and the other only 5.2 micro-subjects ,you can say the difference is 32%.

The Khan Academy group spent 50% of their time on basic mathematical skills while the control group spend all their time on algebra. That might be a reason as to why we don't see better improvement in their score.