Tasks

Real Unreal Bench · Task 11

Play an action RPG against its spec and fix what breaks

Task pass rate50% ±29%

An action RPG has to ship. Given a spec of its levels, world and menus, the agent plays it through the way a player would, finds the places where it breaks the spec (about nine: levels that never end, rooms and enemies out of reach, menus and controls among them) and fixes them without adding content.

Before

A village lane with no controls shown

After

The same lane with the controls on screen
The grader's frames of the same moment: the untouched start, and the reference solution.

Results