Gamer: A Generalist Vision-Language-Action Foundation Model for Long-Horizon Real-Time Game Agent
Gamer learns from tens of thousands of hours of keyboard and mouse interaction, with reported efficiency comparable to skilled players in several 3D open-world games.
The framework combines perception, memory management and multi-step reasoning across 2D interfaces and 3D environments. Sparse-thinking instruction tuning reduces redundant computation while retaining reasoning for complex decisions. Multi-turn reinforcement learning addresses limited context, accumulated errors and sparse rewards during extended play. A cloud-edge inference engine supports concurrent game environments with low latency; the authors also report competitive results on conventional 2D game benchmarks.
HF Daily Papers' note
The framework combines perception, memory management and multi-step reasoning across 2D interfaces and 3D environments. Sparse-thinking instruction tuning reduces redundant computation while retaining reasoning for complex decisions. Multi-turn reinforcement learning addresses limited context, accumulated errors and sparse rewards during extended play. A cloud-edge inference engine supports concurrent game environments with low latency; the authors also report competitive results on conventional 2D game benchmarks.
HF Daily Papers' note
score 5