They see if reusing JVM bytecode analysis frameworks for languages that are not Java is useful. They find that it is, but not for dynamic languages such as Python and Ruby.

They determine the precision of the generated static CFG by also running the
compiled code with instrumentation to acquire the dynamic CFG. Using this you can determine which nodes and edges of the static CFG are not necessary (assuming you have a complete set of inputs to trigger all dynamically reachable edges and nodes).

They say that WALA is a bit better at reasoning around reflection, especially when it concerns string literals, but otherwise not much is said comparing WALA and SOOT, or why they even picked WALA in the first place. I think it's likely a combination of institutional affiliation, and because WALA supports both static and dynamic CFG construction.

Scala seems to be the winner in terms of runtime and memory usage :)