miHoYo Unveils 8B-Parameter Gamer VLA Model That Plays Star Rail Autonomously

By 衡宇Original title: 米哈游发了个VLA论文
78AI Score

miHoYo's Honkai AI research team released Gamer, a general-purpose vision-language-action (VLA) foundation model built on Qwen3-VL-8B that outputs keyboard and mouse actions from screen pixels and task instructions alone, without reading game internals. In a roughly 50-hour long-horizon evaluation across five main story chapters of Honkai: Star Rail, it ran autonomously for over 95% of the time, unifying cross-game real-time control with long-horizon memory and planning in a single VLA framework.

Full Report

Contents3 sections

8B parameters, 50 hours of autonomous operation

henry from Aofei Temple QuantumBit | WeChat Official Account QbitAI

miHoYo, are you planning to personally step in and research "official scripts"? (doge)

Just in the latest research Gamer released by miHoYo's Honkai AI R&D team—

Would you believe it, the AI actually started playing Honkai: Star Rail on its own.

Hand the keyboard and mouse to the AI, and right off the bat it can fight Cocolia.

When it encounters scene mechanisms, it can handle them on its own too. Walking up to the "Bubble Pinball," it activates the device, launches itself into another area, and keeps running after landing.

Following NPCs around the map works as well—Firefly leads the way in front, and the AI follows behind, going up stairs, turning corners, and running all the way to the plaza.

Goodness gracious, fighting bosses, activating mechanisms, and following people around the map—the AI has learned them all.

In the future, will there be an "AI auto-grind" option on the top-up page?

(Note: Honkai: Star Rail is a turn-based RPG with a galactic adventure theme developed by miHoYo. Players ride the Astral Express to explore different worlds, advancing the story through map exploration, dialogue, and puzzles, and assembling character teams for turn-based combat.)

VLA auto-grind, no burnout

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 above playing Star Rail was further trained by the team based on the 8-billion-parameter Qwen3-VL-8B.

It can decide which keys to press, where to move the mouse, and when to click based on the game screen and task, and then actually execute these operations.

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

Compared with previous AI game-playing efforts, Gamer's biggest highlight this time is integrating cross-game real-time operation with long-horizon memory and planning into a single VLA framework.

In actual gameplay, the model makes judgments entirely based on screen images, task instructions, and existing memory, without needing to read the game's internal interfaces or underlying states.

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

In the paper, the research also presents a set of hardcore empirical data: in a long-horizon Star Rail evaluation covering five main story chapters totaling about 50 hours, Gamer's autonomous operation time accounted for over 95%.

To achieve this, the research designed a three-stage training pipeline:

First train operations with cross-game data, then supplement thinking and memory with task demonstrations, and finally let the model enter the game and continue improving through trial and error and feedback.

First is Continued Pre-Training (CPT), which enables the base model that already has image-text understanding capabilities to further learn to operate games.

Specifically, the team aligned screen recordings from different games with player operations, letting the model learn the relationship between visuals and actions: when a character walks forward, how does the scene change; when the mouse moves right, where does the camera turn.

These operations were organized into a unified keyboard-mouse format, allowing data from different games to enter the same training pipeline.

During execution, Gamer reads the screen every 200 milliseconds and plans the next six action chunks at once, each about 33 milliseconds, corresponding to a 30Hz control rhythm.

It's equivalent to taking a look, planning a short segment of actions, and then continuing to adjust based on the new screen.

Next, in the fine-tuning stage, the team curated about 1,000 hours of high-quality Star Rail demonstrations, dividing goals into specific operations, stage sub-goals, and complete tasks, teaching the model to organize actions.

Worth mentioning is that there are two extremely practical engineering designs here, both aimed at solving the inference latency and context explosion problems in long-horizon tasks:

– "Sparse thinking" mechanism: When advancing along a confirmed route, it can operate directly; only when the route fails, the goal switches, or a new interaction prompt appears does it expand reasoning. – Memory summary mechanism: When old records are about to be evicted from the context, the model first notes down task progress, failed attempts, important locations, and follow-up plans, so that subsequent operations have a basis to follow.

  • "Sparse thinking": When advancing along a confirmed route, it can operate directly; only when the route fails, the goal switches, or a new interaction prompt appears does it expand reasoning.

Finally, in the online reinforcement learning stage, the team proposed TA-GRPO, which splits long-horizon tasks into segments connected by memory, compares the performance of different attempts at the corresponding stages, and then updates the model using results verified by the game environment.

For example, whether the target location was reached and whether the current task was completed can both 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 emit generated and verified action chunks ahead of time, letting the game screen start moving first and greatly reducing operation latency.

Experimental testing

In cross-game experiments, the team tested Gamer's CPT version, covering Minecraft, ViZDoom based on Doom, and multiple game environments such as OmniGameArena and GameWorld.

For example, in Minecraft, it can rotate the view on its own to lock onto a slime, chase it down, and swing its sword to slash:

Switching to the first-person shooter ViZDoom, it can quickly flick its aim and fire consecutive shots at approaching enemies:

Furthermore, on the four public benchmarks above, Gamer-8B-CPT achieved an average score of 29.5, while among the listed general model comparisons, the highest was only 17.6.

Among the seven categories of mini-games, totaling 40 tasks not involved in training, the base model scored only 3.2%, which rose to 15.0% after CPT, increased to 39.2% after SFT, and further reached 49.7% with reinforcement learning added—

For comparison, the highest score among the general-purpose models competing in the same arena (GPT-6.1-Sol) was only 17.1%.

These results support the team's core judgment: By combining cross-game learning, long-term memory, and real-time operation through a unified keyboard-and-mouse interface, AI can continuously execute tasks in complex games and keep improving through interaction.

The team also noted at the end of the paper that such capabilities are expected to be applied to automated testing, accelerate game iteration, and support new gameplay mechanics, thereby reducing development costs.

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

Team Introduction

Finally, let us meet the team behind this research.

The paper's contributor list includes 8 co-first authors, 5 of whom participated in the project's R&D during internships at miHoYo:

Run Luo: A PhD student in Computer Science at the National University of Singapore, who has participated in projects such as GUI-R1 and OpenOmni, and was responsible for the three-stage training, data, infrastructure, and evaluation construction in Gamer.

Kuan Zhang: A PhD student at the Institute for AI, Tsinghua University, co-first author of GameScaling and GameVerse, mainly responsible for continual pre-training, data, and evaluation construction.

It is worth mentioning that he wrote on his personal homepage that he is a Master-tier player in League of Legends.

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

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 research RuCL, mainly responsible for supervised fine-tuning, data, and evaluation construction.

Zhihao Yang: A PhD student in Computer Science at the University of Chinese Academy of Sciences (undergraduate from Tongji University), focusing on games and agents, and responsible for reinforcement learning, data, and evaluation construction in the project.

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

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

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

(For other paper contributors, please refer to the paper homepage.)

Zooming out, this miHoYo Honkai AI R&D team's layout already covers three major directions: AI NPCs, game agents, and R&D automation, continuously exploring the implementation of large models in game experience innovation and industrialized pipelines.

One of the paper's advisors, Yinhe Zheng, is precisely the head of the AI NPC and Gameplay technology team for the Honkai series.

Previously, Yinhe Zheng disclosed at the Yunqi Conference the AI Pom-Pom in Honkai: Star Rail, as well as the EchoX agent platform that empowers internal R&D.

On the academic side, the team has also partnered with universities to launch GameScaling and Video-Arena, respectively investigating the Scaling Law of cross-game continual pre-training and the boundaries of model understanding of dynamic game visuals.

Reference links [1]https://arxiv.org/html/2611.18397v1 [2]https://mihoyo-gamer.github.io/

miHoYo Unveils 8B-Parameter Gamer VLA Model That Plays Star Rail Autonomously | Today for AI