Tencent ARC Lab Ships GameHorizon Suite: 5,000 Hours of AAA Gameplay Across 21 Games, 47 Models Tested

Tencent's ARC Lab published GameHorizon Suite on Sept 21, comprising an automated multi-horizon instruction annotator, a 5,000-hour AAA gameplay dataset from 21 games captured by 100 expert players with temporally aligned actions and multi-horizon instructions, and a reproducible offline+online evaluation benchmark. Across 47 models and more than one million evaluation invocations, the paper reports substantial variance in gameplay capability across horizons. Dataset, annotator and benchmark are to be released publicly as a 'standardized yardstick' for evaluating gameplay agents.