Researchers introduce Gamer, a generalist vision-language-action foundation model designed for long-horizon tasks in complex interactive environments, featuring an end-to-end architecture integrating perception, memory management, and multi-step reasoning. The model undergoes continued pre-training on a dataset comprising tens of thousands of hours of device-level keyboard and mouse trajectories spanning simulated games and embodied environments, enabling it to learn environmental dynamics across physical and virtual domains. Sparse-thinking instruction tuning is applied to substantially reduce redundant computation while preserving complex reasoning capabilities. For policy optimization, a stable multi-turn reinforcement learning method addresses finite context windows, accumulated rollout errors, and sparse rewards. Deployment utilizes a cloud-edge inference engine that serves numerous simulated game environments concurrently, supports unbounded-context execution, and maintains low online inference latency for real-time interaction. Across multiple challenging 3D open-world games, Gamer demonstrates solid generalization and operational efficiency comparable to skilled human players, while also achieving competitive results on conventional 2D game benchmarks. The findings suggest that scaling data and learning algorithms on repetitive keyboard and mouse trajectories yields substantial gains, allowing the model to acquire complex skills steadily in environments possessing both physical and digital properties, thereby offering a path toward generalist agents capable of interacting across digital and physical worlds.

Read original