SWE-Game: Can Coding Agents Build the Games We Want?
The benchmark says leading coding agents still miss core gameplay requirements more often than they should.
SWE-Game tests agents on 247 tasks drawn from 41 executable Godot games across 2D and 3D categories. Its task set covers building from briefs and design docs, completing skeletons, repairing injected faults, and porting Godot games to Unity. Across six models, Opus5 led every task type, but the best construction scores stayed under 60/100. The paper says the main failures were omitted requirements and gameplay logic errors. Source: HF Daily Papers' note.
SWE-Game tests agents on 247 tasks drawn from 41 executable Godot games across 2D and 3D categories. Its task set covers building from briefs and design docs, completing skeletons, repairing injected faults, and porting Godot games to Unity. Across six models, Opus5 led every task type, but the best construction scores stayed under 60/100. The paper says the main failures were omitted requirements and gameplay logic errors. Source: HF Daily Papers' note.
score 5