GameHorizon Suite: Multi-Horizon Data and Evaluation in Gameplay
Modern video games provide a measurable testbed for AI models, combining abilities of visual understanding, instruction decomposition, goal planning, and precise action control over multiple temporal horizons. Existing datasets and benchmarks, however, either cover a narrow range of games, lack language instructions, or rely on high-variance online rollouts. To address these ch
Reading in Plain English mode — toggle below to switch
Modern video games provide a measurable testbed for AI models, combining abilities of visual understanding, instruction decomposition, goal planning, and precise action control over multiple temporal horizons. Existing datasets and benchmarks, however, either cover a narrow range of games, lack language instructions, or rely on high-variance online rollouts. To address these ch
This paper introduces a novel benchmark and dataset for evaluating AI agents in complex video game environments, addressing limitations of existing methods. For ML engineers, this provides a standardized way to test and improve models that require visual understanding, planning, and action control across various temporal horizons, which is crucial for developing robust AI agents.