miHoYo Published a VLA Paper

8B parameters, 50 hours of autonomous operation
henry reporting from Waifei Temple, Quantum位 | QbitAI public account
miHoYo, are you getting ready to research "official scripts" yourself? (doge)
In Gamer, the latest research 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 away it can get into a fight with Cocolia.

When it encounters scene mechanisms, it can also handle them itself. It walks up to the "bubble pinball" and activates the device, launching itself into another area, then keeps running after landing.

Following NPCs across the map also works. Firefly leads the way in front, and the AI follows behind, going up stairs, turning corners, and running all the way to the plaza.
My goodness, fighting bosses, activating mechanisms, and following people across 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 role-playing game with a galactic adventure theme developed by miHoYo. Players ride the Astral Express to explore different worlds, advance the story through map traversal, dialogue, and puzzles, and build 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 Qwen3-VL-8B with 8 billion parameters.
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 and mouse operations are the actions it outputs.
Compared with previous AI game-playing, Gamer's biggest highlight this time is connecting cross-game real-time operation with long-horizon memory and planning within the same VLA framework.

In actual play, 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 state.
As the task progresses, it also dynamically records "what was just completed" and "what to do next."
In the paper, the research also provides a set of hardcore measured data: in a long-horizon Star Rail evaluation covering five main story chapters and totaling about 50 hours, Gamer's autonomous operation time accounted for more than 95%.

To achieve this result, the research designed a three-stage training pipeline:
First train operation 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 images and actions: how the scene changes when the character moves forward; where the camera turns when the mouse moves right.
These operations are organized into a unified keyboard-and-mouse format, allowing data from different games to enter the same training pipeline.
During execution, Gamer reads the screen every 200 milliseconds and arranges the next six action chunks at a time, each about 33 milliseconds, corresponding to a 30Hz control rhythm.
It is equivalent to taking a look, arranging a short segment of actions, and then continuing to adjust based on the new screen.
Next, in the fine-tuning stage, the team organized about 1,000 hours of high-quality Star Rail demonstrations, dividing goals into specific operations, stage subgoals, and complete tasks, teaching the model to organize actions.
It is worth mentioning that there are two extremely practical engineering designs here, both aimed at solving reasoning latency and context explosion in long-horizon processes:
– Sparse thinking mechanism: when advancing along a determined route, it can operate directly; when the route fails, the goal switches, or a new interaction prompt appears, it then expands reasoning. – Memory summary mechanism: when old records are about to move out of context, the model first notes task progress, failed attempts, important locations, and follow-up plans, so that subsequent operations have a basis.
- "Sparse thinking": when advancing along a determined route, it can operate directly; when the route fails, the goal switches, or a new interaction prompt appears, it then expands reasoning.
Finally, in the online reinforcement learning stage, the team proposed TA-GRPO, splitting long-horizon processes into segments connected by memory, comparing the performance of different attempts at the corresponding stages, and then updating 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 become training feedback, while the training backend is responsible for initialization and verification.
In addition, the accompanying Streaming-vLLM inference engine can reuse already processed context and emit generated and verified action chunks in advance, 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 multiple game environments including Minecraft, ViZDoom based on Doom, as well as OmniGameArena and GameWorld.
For example, in Minecraft, it can turn the view on its own to lock onto a slime, chase it, and swing its sword:

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

In addition, 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.
And in 40 tasks across seven types of mini-games that were not involved in training, the base model scored only 3.2%, rising to 15.0% after CPT, to 39.2% after SFT, and reaching 49.7% after adding reinforcement learning—
By comparison, the highest score among general models competing in the same setting (GPT-6.1-Sol) was only 17.1%.

These results support the team's core judgment: through a unified keyboard-and-mouse interface, combining cross-game learning, long-term memory, and real-time operation can enable AI to continuously execute tasks in complex games and continue improving through interaction.
The team also points out at the end of the paper that this type of capability is expected to be used for automated testing, accelerating game iteration, and supporting new gameplay mechanisms, thereby reducing development costs.
Games become AI's training ground, and the trained AI can in turn help develop games.
Team introduction
Finally, let's 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: PhD student in computer science at the National University of Singapore, 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: PhD student at the Institute for AI, Tsinghua University, co-first author of GameScaling and GameVerse, mainly responsible for continued pre-training, data, and evaluation construction.
It is worth mentioning that on his personal homepage, he writes that he is a Master-tier player in League of Legends.

Qi Deng: 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: 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: 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: member of miHoYo's Honkai AI R&D team, mainly responsible for training and inference infrastructure construction.
Renrong Hu: 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 has covered three major directions: AI NPCs, game agents, and R&D automation, continuously exploring the application of large models in game experience innovation and industrialized pipelines.
One of the paper's supervisors, Zheng Yinhe, is precisely the head of the AI NPC and Gameplay technology team for the Honkai series.
Article title: miHoYo Released a VLA Paper
Body: Previously, Zheng Yinhe disclosed at the Yunqi Conference the AI Paimon of "Honkai: Star Rail," as well as the Agent platform EchoX 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/
- GPT-6.2 Astra Strikes Embodied Models Again: Success Rate on 16 Tasks Doubles Compared with GPT-6 Astra 2026-11-21
- Yao Shunyu and Hunyuan Team Answer Netizens on Reddit, Hy Small Model Already on Roadmap 2026-11-21
- Embodied Intelligence Going Global, Logistics Leads the Way 2026-11-09
- The Riemann Hypothesis Proven?! With Claude's Assistance, 34-Year-Old Mathematician Releases Lean Proof 2026-11-21
Related Reading

Ant Group's Embodied Intelligence Shows Its Hand: Building Brains, Competing in a Differentiated Way with Unitree and Others
First Real-World Validation of the VLA Scaling Law

In One Go, miHoYo Sues Bilibili Seven Times in a Row
May Be Related to Bilibili "Leaker" Videos

miHoYo Cai Haoyu's New Work iPhone Actual Gameplay Demo: Blushed by AI Xiaomei in 10 Minutes, Her Fate Is in My Hands

Behind the Commonality Between miHoYo and Burger King, a Technological Prophecy Has Already Become Reality
The 15 Years of Cloud Computing from Infancy to "Water, Electricity, and Coal"

VLA Is Dead, Teleoperation Is Dead Too! Said by NVIDIA's No. 1 Robotics Figure
Jim Fan's Latest Bombshell

miHoYo Releases Its First AIGC Tool: Everything Can Be "March 7th"
An Image-to-Image Tool Based on AI Algorithms