@johnabs

dreamcoder
https://arxiv.org/abs/2006.08381

Its actually a cutout of this guy's thesis.
https://www.cs.cornell.edu/~ellisk/documents/kevin_ellis_thesis.pdf

But the data is self referential. They have enumeration as a baseline. The timeout of which can be set for whatever their program can beat. Aside from that, the data used as input has been tailored over 4 years to make their program able to learn it. In the code tasks that are too tough are just commented out. So of course they have a near 100% learning rate.

There is also nothing to compare it to besides an older version of itself.