MiHoYo's Gamer Agent Beats GPT-4o at Honkai Star Rail by 32 Points
miHoYo's Gamer model turns a Qwen3-VL backbone into an agent that plays Honkai: Star Rail with keyboard and mouse at human speed.
Takeaways
- miHoYo's Gamer agent plays Honkai: Star Rail side quests via raw keyboard and mouse. Paper
- Built on Qwen3-VL, trained on 10K+ hours of gameplay traces with aligned device events.
- Scores 49.7% on 7 held-out HSR minigame types vs 17.1% for GPT-6.1-Sol.
- TA-GRPO handles long-horizon RL by splitting rollouts into memory-linked segments with verifiable rewards.
- Streaming-vLLM cuts inference latency from 1,306 ms to 65 ms, a 20x speedup.
- Code and CPT scaling checkpoints released; HSR-tuned weights withheld.
MiHoYo’s Honkai AI R&D team has introduced Gamer, a vision-language-action agent that reads game frames and emits keyboard and mouse events. The team reports successful side quests and minigames in Honkai: Star Rail, along with tests in Genshin Impact, Zenless Zone Zero, Minecraft, and Doom Eternal.
On 40 held-out Honkai: Star Rail puzzles spanning seven categories, Gamer-RL reached a 49.7% macro-average success rate, giving each category equal weight. GPT-6.1-Sol, the strongest general-purpose baseline tested, scored 17.1%. The game continued running during inference, exposing the latency and control errors hidden by benchmarks that pause their environments.
Live games expose the latency bill
Many VLM agent benchmarks freeze the environment while a model generates its next action. A live 3D game continues processing movement, enemies, timers, and interface changes during that delay. Its agent must also preserve goals across long sessions and translate decisions into precise mouse deltas and key-state transitions.
In Gamer’s ablation study, switching from paused to real-time evaluation reduced general-purpose models’ average scores by 17 to 35 percentage points. Gamer-35B-A3B-CPT gained 3.4 points under the same change. The paper organizes the underlying engineering problem around three constraints:
- Finite context: quests can last longer than the model’s context window.
- Compounding errors: a policy may chain thousands of mouse and keyboard events before completing an objective.
- Sparse rewards: many environments reveal success only when a quest, route, or puzzle is complete.
Three stages turn traces into control
MiHoYo initializes Gamer from Alibaba’s Qwen3-VL, using an 8B dense model and a 35B-A3B mixture-of-experts variant. The unmodified base model completed none of the target quests. Continued pre-training, supervised fine-tuning, and reinforcement learning each add a distinct capability.
1. CPT learns the motor vocabulary
Continued pre-training uses more than 10,000 hours of gameplay aligned with keyboard and mouse traces. From screenshots sampled at 5 Hz, the model predicts a mouse displacement and six consecutive 33 ms key-state blocks for each 200 ms interval. This representation supplies 30 Hz controls through ordinary token generation, with no separate action head.