Game2Player Improves PlayerBench Scores via Web-Game Training
November 19, 2026
Game2Player uses 568 verifiable web-game environments to scale visual-agent training. A 9B model trained on this data improved PlayerBench scores from 14.3% to 31.2%.
HOW THIS AFFECTS YOU
●
researcherYou can use these 3,436 tasks and 307 hours of gameplay to train visual agents in verifiable environments.