miHoYo's AI research team has developed Gamer, a foundational model for long-form, real-time interactive visual-language-action tasks. This model, trained on an 8-billion parameter Qwen3-VL-8B base, can autonomously play games like Honkai: Star Rail by interpreting visual input and outputting keyboard and mouse actions. Gamer demonstrated over 95% autonomous operation during a 50-hour gameplay session, showcasing its ability to learn complex sequences, manage context through memory summarization, and improve via reinforcement learning. AI
IMPACT This research could accelerate the development of AI agents capable of complex, long-term tasks in interactive environments, potentially impacting game development and automated testing.
RANK_REASON The item describes a research paper and a new foundational model released by a game company's AI research team, detailing its training methodology and performance on game-playing tasks. [lever_c_demoted from research: ic=1 ai=1.0]
- Continued pre-training
- Gamer
- Gameworlds: Virtual Media and Children’s Everyday Play
- Honkai: Star Rail
- Kuan Zhang
- miHoYo
- OmniGameArena
- Qi Deng
- Qwen3 VL 8B
- Run Luo
- Streaming-vLLM
- supervised fine-tuning
- TA-GRPO
- ViZDoom
- Yufei Zhang
- Yukun Chen
- Zhihao Yang
AI-generated summary · Google Gemini · from 1 sources. How we write summaries →