Author, here. The paper is about the Collatz sequence, how experiments with a transformer can point at interesting facts about a complex mathematical phenomenon, and how, in supervised math transformers, model predictions and errors can be explained (this part is a follow-up to a similar paper about GCD). From a ML research perspective, the interesting (but surprising) take away is the particular way the long Collatz function is learned: "one loop at a time".
To me, the base conversion is a side quest. We just wanted to rule out this explanation for the model behavior. It may be worth further investigation, but it won't be by us. Another (less important) reason is paper length, if you want to submit to peer reviewed outlets, you need to keep pages under a certain number.
For training, you need a generator because you want millions of solved examples for deep learning to work.
At test time, you usually want a test set from the same distribution as the training data (or at least related to it in some controllable way), or it becomes very difficult to interpret the results.
Suppose my test set come from a different and unknown distribution (real problems sampled in some way).
If I get good results, is it because the training worked, or because the test set was "comparatively easy"?
If I get bad results, is it because the model did not learn, or because the test set was too far away from the training examples?
Textbook problems are usually short, with short solutions,
and demonstrating one specific rule. They are better handled by classical (rule-based) tools. Deep learning tools would either memorize them or resort to a rule based sub-module.
For integrals, solvability depends on the function set you work with. Since we use elementary functions on the real domain, a lot of integrals have no solution. We could have gone for a larger set (adding erf, the Fresnels, up to Liouvillian functions). This would mean more solvable cases.
As for the engineering distribution, no one knows what it is. The best we can do is to generate a broader training set, knowing that it will generalize better (this is the key takeaway of our appendix). BWD+IBP is a step in this direction, but to progress further, we need a better understanding of the problem space, and issues related to simplification. We are working on this now.
Since the paper was presented, it was reviewed in ICLR and significantly expanded, and many questions raised in September were addressed.
The performance metric is the number of equations/integrals correctly solved over a held out test set, generated randomly. This is done using an external tool (SymPy), which some might consider "cheating" (but you need a way to test, don't you?). Anyway, keep in mind that checking a solution for correctness is much easier than finding one.
Another issue was about using timeout for Wolfram. This is now discussed in the appendix (it makes very little difference).
During review, an interesting point was discussed: out of distribution generalization, or how performance depends on the random problem generator we used. This is now measured and discussed at the end of the appendix.
To me, the base conversion is a side quest. We just wanted to rule out this explanation for the model behavior. It may be worth further investigation, but it won't be by us. Another (less important) reason is paper length, if you want to submit to peer reviewed outlets, you need to keep pages under a certain number.