My experience has been that they excel at problems with solutions that are cheap to verify. Things one can unit test for example. Easier to RL. Architectural issues are harder to train for because the training data is dominated by tutorial/toy projects. There's 10,000 toy compilers for every LLVM