The link to the paper seems to be broken but it is openly available at IEEE explore: https://ieeexplore.ieee.org/document/9152719

The paper explores manually guiding AFL-based fuzzers using annotations in the source code in order to more easily reach deep states.

The Super Mario Bros example highlighted in the git repo uses a modified version of a C++ variant converted from disassembly of the original game: https://github.com/MitchellSternke/SuperMarioBros-C/

In this example, the only guidance used is an annotation that maximizes the x-position while accounting for different y-positions, allowing it to faster complete the game levels: https://github.com/RUB-SysSec/ijon-data/blob/c75433e1ce52523...