Historical brief · 2026-11-13

Gamer: A Generalist Vision-Language-Action Foundation Model for Long-Horizon Real-Time Game Agents

Researchers from miHoYo introduce Gamer, a vision-language-action foundation model designed for long-horizon, real-time game interaction via standard screen pixels, keyboard, and mouse inputs.

Source image · Hugging Face Daily Papers

The miHoYo Honkai AI R&D team presents Gamer, a generalist Vision-Language-Action (VLA) foundation model designed to interact with complex 3D games through standard human interfaces—screenshots, keyboard, and mouse—without relying on privileged game states or custom APIs. The authors report a complete pipeline encompassing continual pre-training on over 10,000 hours of gameplay trajectories, supervised fine-tuning using an automated annotation pipeline yielding approximately 1.2 million reasoning and 223,000 memory annotations, and an optimized inference engine called Streaming-vLLM that achieves over 20 times lower inference latency. To handle long-horizon tasks, the team implements Temporal-Aware GRPO (TA-GRPO) for reinforcement learning over hour-long trajectories. In automated evaluations across main storylines in Honkai: Star Rail, the authors report that Gamer handles over 95% of challenges. The work provides insights into scaling laws for game intelligence and the interplay between inference speed and online reinforcement learning effectiveness. As this is a preprint, independent replication, code availability details, and comprehensive generalization baselines beyond the developers' proprietary test environments remain to be fully established.

View selection rationale
Media reportStage not established

Importance 77/100

Signal78Structure78Consequence75Confidence71Momentum72
About this brief
Content type
Research summary
Source type
First-party / official source
Source published
Process
AI-assisted discovery, translation, summarisation, classification and scoring; standards and corrections are maintained by the editorial desk.
Verification
The original URL and publication metadata are retained. This brief is not independently fact-checked unless explicitly stated.
Published by
Polyemergence Editorial Desk · PolyEmergence Pte Ltd.
Conflict disclosure
No financial or commercial conflict declared.

Editorial standards → · Corrections →

POLYEMERGENCE / EXPLORE

Keep exploring

Follow the topic, not just the headline.

Explore the Artificial Intelligence archive →