GameXpert-Bench Splits Coding-Agent Game Development Into Generation, Repair, and Multi-Turn Optimization Tracks

arXiv 2608.21833·medium signalXBlueskyLinkedInCopy link

Game development stresses coding agents unusually hard because program logic, visual and audio content, interfaces, interaction, and playability must all work together in one executable artifact. Analyzing complete human-agent development trajectories, the authors identify three lifecycle stages and operationalize each as a benchmark track: GameGen for complete game creation from a single request in an empty workspace, GameFix for diagnosis and repair when defects are reported or left for the agent to find, and GameOpt for cumulative optimization through request chains seeded by real developer needs. It targets the gap where existing benchmarks score only the final artifact or one isolated stage.