Gamer: Generalist Vision-Language-Action Model for Long-Horizon Game Agents
·2026.11.16 19:58
Key point
The model combines perception, memory, and multi-step reasoning to handle long-horizon tasks in both 2D and 3D environments.
Details
- Introduces Gamer, a generalist agent framework for long-horizon tasks in complex interactive environments, featuring an end-to-end architecture for perception, memory management, and multi-step reasoning.
- Trained on tens of thousands of hours of device-level keyboard and mouse interaction data spanning simulated games and embodied environments to learn dynamics across physical and virtual domains.
- Utilizes sparse-thinking instruction tuning to reduce redundant computation and a stable multi-turn reinforcement learning method to address finite context windows, accumulated rollout errors, and sparse rewards.
- Demonstrates solid generalization and operational efficiency comparable to skilled players in challenging 3D open-world games, while also achieving competitive results on conventional 2D game benchmarks.
- Employs a cloud-edge inference engine that serves numerous simulated game environments concurrently, supports unbounded-context execution, and maintains low online inference latency.
This summary was generated automatically by AI. Check the original for the author's claims and context. Copyright belongs to the original author.
Our guide explains how the AI works. Report summary errors, attribution issues, or removal requests via Contact.