This is very interesting. Have they considered using multiple languages (functional, oo) as well as including 'garbage' languages (brainfuck) to track differences? I'd also be curious to see how it compares to a standard word problem (two trains leaving chicago etc).