It would be very useful to automatically minimize programs that trigger bugs.
The main obstacle here is that currently we don't attribute bugs to particular programs, so we cannot say that this program triggers a bug. Without such signal it is not possible to minimize. The signal is challenging, because console output is processed asynchronously and a VM can run several programs concurrently.
Probably we can separate minimization process from normal testing. That is, once we have a crash log, we can wind up a dedicated VM and run it in special slower mode when we attribute output/bugs to just executed program. This would simplify things significantly.