Synthetic game practice improves Minecraft task success

Gameplay demonstrations cover only some situations a game-playing model may encounter, while scalable interactive environments are difficult to obtain.

The research asks whether practice in synthetic games can improve control beyond learning from examples, and whether the gains carry over to games outside those training environments.

Player-1 trains computer-use models on gameplay examples, then uses checkable rewards in synthetic games to improve game control.

Player-1 addresses whether synthetic game practice can improve control beyond gameplay demonstrations. It trains computer-use models on curated examples, then lets them learn through interaction in synthetic environments with rewards calculated from game score changes.

On 300 Minecraft tasks, the 9B model's average success rate was 38.33% after interactive training, versus 10.54% after demonstrations alone. Success was averaged across eight trials within each task, then equally across tasks, supporting a gain on this panel.

The results cover specific evaluation panels and trial budgets: 300 Minecraft tasks, 10 first-person game maps, and 170 browser-game tasks across 34 games. They show gains for the tested models on those panels, not performance across all games or tasks.