Back to articles

miHoYo’s Gamer explores long-horizon game-playing agents

3 min read

miHoYo’s Honkai AI research team has introduced Gamer, a vision-language-action (VLA) system designed to complete game tasks without reading a game’s internal state or using dedicated gameplay APIs. It observes the screen, interprets a task instruction, and acts through keyboard and mouse controls. The showcased 8-billion-parameter version is built on Qwen3-VL-8B and was tested in Honkai: Star Rail on activities including combat, interacting with mechanisms, and following characters through the environment.

How the system is trained

  • A shared control interface: Gamer is trained on gameplay recordings paired with player actions from different games. The data is converted into a common keyboard-and-mouse format, teaching the model how actions change what appears on screen. During play, it reads a new frame roughly every 200 milliseconds, plans a short sequence of actions, and then adjusts to the next observation.
  • Memory for extended tasks: Training proceeds through continued pre-training, supervised fine-tuning, and online reinforcement learning. The fine-tuning stage uses about 1,000 hours of Honkai: Star Rail demonstrations. To keep long tasks manageable, Gamer summarizes progress, failed attempts, important locations, and future plans before older context is lost. Its “sparse thinking” design reserves more explicit reasoning for route failures, new prompts, or changes in the goal; routine movement can proceed directly.
  • Learning from outcomes: The team’s TA-GRPO method breaks long tasks into sections linked by memory and uses environment feedback—such as whether the agent reached a location or completed a stage—to compare attempts. A companion inference engine reuses processed context and prepares verified action chunks in advance to reduce interaction delays.

The paper reports that Gamer spent more than 95% of the time running autonomously in an approximately 50-hour evaluation spanning five main-story chapters of Honkai: Star Rail. On 40 unseen tasks across seven mini-game categories, reported performance rose from 3.2% for the base model to 15.0% after continued pre-training, 39.2% after supervised fine-tuning, and 49.7% after reinforcement learning. These figures describe the benchmark and setup used in the study; they do not establish that the model can reliably complete arbitrary games.

Why it matters—and what remains open

Gamer’s contribution is an attempt to bring cross-game control, planning, memory, and interactive learning into one pipeline, rather than simply teaching a model a single game. A shared input interface may reduce dependence on game-specific integration and make games useful, bounded environments for agent training. The researchers point to possible applications in automated testing, faster development iteration, and new gameplay mechanics. Still, benchmark performance is only one step toward those uses: reliability in unfamiliar scenarios and integration into real development workflows remain to be demonstrated.

Source: QbitAI

Comments

Checking sign-in status...

Loading comments...

Related articles