aifollow.news
Back 量子位 AI 报道
量子位 AI 报道· · Original publication time AI score54

miHoYo team publishes research on Gamer, a game-playing modelMachine translation

米哈游团队发布Gamer游戏操作模型研究

AI-assisted summary

miHoYo's Honkai AI research team developed Gamer, a vision-language-action model that uses game visuals, task instructions and memory to control a keyboard and mouse without reading internal game state. The report says the paper found that Gamer operated autonomously for more than 95% of an approximately 50-hour evaluation spanning five main story chapters of Honkai: Star Rail. In cross-game tests, its trained version also outscored the listed general-purpose model comparisons.

Article · AI translation

8B parameters, 50 hours of autonomous operation

By henry, reporting from Aofeisi

QbitAI | WeChat public account QbitAI

By henry, reporting from Aofeisi

QbitAI | WeChat public account QbitAI

miHoYo, are you planning to personally research an “official bot”? (doge)

In Gamer, the latest research released by miHoYo’s Honkai AI R&D team—

Can you believe it? AI is playing Honkai: Star Rail by itself.

Hand the keyboard and mouse to the AI, and it can jump straight into a fight with Cocolia.

It can also handle mechanisms in the environment on its own. It walks up to a “bubble pinball” device, activates it, launches itself into another area, lands, and keeps running.

It can follow an NPC across the map, too: Firefly leads the way, and the AI follows behind, climbing stairs, turning corners, and running all the way to the square.

Well, would you look at that: AI has learned to fight bosses, activate mechanisms, and follow people across the map.

Will the top-up page eventually have an “AI plays for you” option?

(Note: Honkai: Star Rail is a turn-based role-playing game about galactic adventure developed by miHoYo. Players ride the Astral Express to explore different worlds, advance the story by traversing maps, talking, and solving puzzles, and assemble teams of characters for turn-based combat.)

Let VLA handle the grind without the headache

Back to the point: miHoYo positions Gamer as a general-purpose vision-language-action (VLA) foundation model for long-horizon, real-time interaction.

The main version demonstrated playing Honkai: Star Rail above was further trained by the team from Qwen3-VL-8B, which has 8 billion parameters.

Based on the game screen and the task, it decides which keys to press, where to move the mouse, and when to click, then actually carries out those actions.

In other words, the screen is the visual signal it receives, the task instruction tells it what to do, and keyboard and mouse operations are the actions it outputs.

Compared with previous game-playing AI, Gamer’s biggest highlight is bringing real-time control across games together with long-horizon memory and planning in a single VLA framework.

During play, the model makes decisions entirely from the screen, task instructions, and existing memory, without reading the game’s internal interfaces or underlying state.

As the task progresses, it also dynamically records “what was just completed” and “what to do next.”

The paper also provides a set of hard performance figures: in a long-horizon Honkai: Star Rail evaluation spanning five main-story chapters and about 50 hours in total, Gamer operated autonomously for more than 95% of the time.

To achieve this, the researchers designed a three-stage training process:

First, it practices controls on data from multiple games; then task demonstrations add reasoning and memory; finally, the model enters the game and keeps improving through trial, error, and feedback.

First comes continued pre-training (CPT), which teaches a base model that already understands images and text how to operate games.

Specifically, the team aligns recordings from different games with players’ actions so the model learns the relationship between images and actions: how the scene changes when a character moves forward, and where the camera turns when the mouse moves right.

These operations are organized into a unified keyboard-and-mouse format so data from different games can enter the same training process.

During execution, Gamer reads the screen every 200 milliseconds and schedules the next six action chunks at a time, each lasting about 33 milliseconds, corresponding to a 30 Hz control rate.

It is like taking a look, planning a short sequence of actions, then adjusting based on the new image.

Next, during fine-tuning, the team assembled about 1,000 hours of high-quality Honkai: Star Rail demonstrations and divided goals into specific actions, interim subgoals, and complete tasks to teach the model how to organize its actions.

Notably, there are two highly practical engineering designs here, both aimed at addressing inference latency and exploding context length in long tasks:

– “Sparse thinking” mechanism: It can act directly while following an established route; when the route fails, the goal changes, or a new interaction prompt appears, it reasons things through.

– Memory summarization mechanism: Before old records leave the context, the model notes task progress, failed attempts, important locations, and subsequent plans, giving later actions a basis to follow.

– “Sparse thinking”: It can act directly while following an established route; when the route fails, the goal changes, or a new interaction prompt appears, it reasons things through.

Finally, in the online reinforcement learning stage, the team proposes TA-GRPO. It breaks long tasks into segments connected by memory, compares how different attempts perform at the corresponding stage, and updates the model using results verified by the game environment.

For example, whether it reached the target location or completed the current task can serve as training feedback, while the training backend handles initialization and verification.

In addition, the accompanying Streaming-vLLM inference engine can reuse already processed context and output generated and verified action chunks early, getting the game screen moving sooner and substantially reducing control latency.

Experimental tests

In cross-game experiments, the team tested Gamer’s CPT version across multiple game environments, including Minecraft, Doom-based ViZDoom, OmniGameArena, and GameWorld.

In Minecraft, for example, it can turn its view to lock onto a slime, chase it down, and slash it with a sword:

Switch to the first-person shooter ViZDoom, and it can quickly swing its aim onto approaching enemies and fire repeated bursts:

Across the four public benchmarks mentioned above, Gamer-8B-CPT scored an average of 29.5; the highest score among the listed general-purpose model comparisons was only 17.6.

Across seven types of minigames and 40 tasks excluded from training, the base model achieved just 3.2%. This rose to 15.0% after CPT, 39.2% after SFT, and 49.7% after reinforcement learning—

By comparison, the highest score by a general-purpose model in the same contest (GPT-6.1-Sol) was only 17.1%.

These results support the team’s central conclusion: combining cross-game learning, long-term memory, and real-time control through a unified keyboard-and-mouse interface can enable AI to carry out tasks continuously in complex games and keep improving through interaction.

The team also notes at the end of the paper that such capabilities could be used for automated testing, faster game iteration, and new gameplay mechanisms, thereby reducing development costs.

Games become a training ground for AI, and the AI trained there can in turn help develop games.

Meet the team

Finally, let’s meet the team behind this research.

The paper lists 8 co-first authors, 5 of whom worked on the project during internships at miHoYo:

Run Luo: A computer science PhD student at the National University of Singapore who previously participated in projects including GUI-R1 and OpenOmni. On Gamer, he was responsible for the three-stage training, data, infrastructure, and evaluation development.

Kuan Zhang: A PhD student at Tsinghua University’s School of Artificial Intelligence and co-first author of GameScaling and GameVerse. He was mainly responsible for continued pre-training, data, and evaluation development.

Notably, he writes on his personal homepage that he is a Master-rank League of Legends player.

Qi Deng: A member of miHoYo’s Honkai AI R&D team, responsible for reinforcement learning, training and inference infrastructure, and data and evaluation development.

Yukun Chen: A graduate student at the University of Chinese Academy of Sciences, co-first author of GameScaling and first author of the multimodal reasoning study RuCL. He was mainly responsible for supervised fine-tuning, data, and evaluation development.

Zhihao Yang: A computer science PhD student at the University of Chinese Academy of Sciences (and a Tongji University undergraduate alumnus) whose research focuses on games and agents. On the project, he was responsible for reinforcement learning, data, and evaluation development.

Yufei Zhang: A member of miHoYo’s Honkai AI R&D team, mainly responsible for building the training and inference infrastructure.

Renrong Hu: A member of miHoYo’s Honkai AI R&D team, mainly responsible for benchmark development and evaluation.

Yi Zhang: From the University of Science and Technology of China, he participated in the project during an internship at miHoYo and was mainly responsible for benchmark development and evaluation.

(For other paper contributors, see the paper’s homepage.)

Looking more broadly, this miHoYo Honkai AI R&D team now works across 3 areas—AI NPCs, game agents, and R&D automation—and continues to explore how large models can be applied to innovations in game experiences and industrial development pipelines.

One of the paper’s advisers, Zheng Yinhe, leads the AI NPC and Gameplay technology team for the Honkai series.

Zheng Yinhe previously disclosed Honkai: Star Rail’s AI Pom-Pom and EchoX, an agent platform supporting internal R&D, at the Apsara Conference.

On the academic side, the team has also partnered with universities to produce GameScaling and Video-Arena, respectively investigating the scaling laws of continued pre-training across games and the limits of models’ understanding of dynamic game visuals.

References

[1]https://arxiv.org/html/2611.18397v1

[2]https://mihoyo-gamer.github.io/

Found an error? Send a correction

Contents