miHoYo AI Paper Revealed: How Reinforcement Learning Teaches NPCs to Act Proactively

The miHoYo AI Lab recently posted a paper on arXiv titled 'Learning to Act: Reinforcement Learning for Proactive NPC Behavior in Open Worlds,' detailing a new approach that gives non-player characters (NPCs) genuine agency. The research marks a key step for AI from simple scripted dialogue toward true autonomous behavior, drawing wide attention across the game industry.

Traditional NPC behavior usually depends on predefined scripts or limited decision trees, leaving character reactions mechanical and repetitive. miHoYo's framework instead uses reinforcement learning, letting NPCs learn how to interact with the environment and other characters through trial and error in simulated settings. In the paper, AI characters are given two reward mechanisms—'curiosity' and 'goal orientation': they proactively explore unknown areas while adjusting behavior around long-term tasks such as escorting the player or gathering resources. Experiments showed that trained NPCs could plan paths autonomously, avoid obstacles, and even react ahead of the player's actions—offering help or issuing a challenge as the player approaches.

The heart of the technology is efficient simulation-based training. miHoYo simplified Genshin Impact's open-world scenes into a test environment and trained the AI across thousands of parallel instances using the distributed reinforcement learning algorithm PPO (Proximal Policy Optimization). Each training round converges in just hours—nearly a 10x efficiency gain over traditional methods. The paper also notes the framework adapts easily to different game genres: adjust the reward function and it generates character behavior consistent with the world.

For players, this means a more immersive experience. Imagine that in Genshin Impact you no longer meet NPCs repeating the same lines, but characters who duck out of the rain when the weather turns and light a campfire on their own at night. The head of the miHoYo AI Lab said the technology has entered internal prototype testing and may be integrated into next-generation titles. They also acknowledge that large-scale deployment still faces challenges around compute consumption and narrative consistency.

Industry analysts believe miHoYo's research will push AI game development into a new phase. Unlike companies such as DeepSeek that focus on dialogue generation, miHoYo emphasizes behavioral autonomy, offering a new design paradigm for open-world games. As tools like Unity Muse also begin introducing reinforcement learning support, 2026 could become the breakout year for AI-driven NPCs. But developers must stay alert: overly free AI could wreck carefully designed story pacing, and balancing autonomy with narrative control will be the key task ahead.

The miHoYo AI Lab's paper is now public for academics and developers to reference, and the team plans to present more details at next month's GDC (Game Developers Conference). For game studios chasing innovation, this is unquestionably a direction worth following.