researchMihoyo

miHoYo's Gamer VLA model, built on Qwen3-VL-8B, runs autonomously for over 95% of a 50-hour Honkai: Star Rail evaluation

TL;DR

miHoYo's Honkai AI R&D team published Gamer, a vision-language-action foundation model for long-horizon, real-time game play. The main 8B-parameter version is trained from Qwen3-VL-8B and, per the paper, ran autonomously for over 95% of a roughly 50-hour Honkai: Star Rail evaluation.

3 min read
0

miHoYo's Honkai AI R&D team has published Gamer, a vision-language-action (VLA) foundation model for long-horizon, real-time game play. According to the paper, the model was autonomous for more than 95% of the time in an evaluation covering five main story chapters of Honkai: Star Rail, about 50 hours in total. The report comes from Chinese-language outlet QbitAI (量子位), which summarized the paper (arXiv 2611.18397) and project page.

What Gamer does

The main version has 8B parameters and was trained further from Qwen3-VL-8B. It takes game frames and a task instruction as input and outputs keyboard and mouse actions: which keys to press, where to move the mouse, and when to click. According to the report, it decides using only screen images, task instructions and its stored memory, with no access to the game's internal interfaces or underlying state.

At inference, Gamer reads a frame every 200 ms and schedules the next six action chunks of about 33 ms each, matching a 30Hz control rate. A companion inference engine, Streaming-vLLM, reuses already processed context and emits verified action chunks early to reduce latency.

Three-stage training

  1. Continued pre-training (CPT): Screen recordings from multiple games are aligned with player inputs and converted to a unified keyboard-and-mouse format.
  2. Fine-tuning (SFT): The team compiled about 1,000 hours of high-quality Star Rail demonstrations, organized into specific operations, intermediate sub-goals and full tasks. Two designs target long-run latency and context growth. "Sparse thinking" lets the model act directly on a known route and reason only when a route fails, a goal changes or a new interaction prompt appears. A memory summary mechanism records task progress, failed attempts, key locations and next plans before old records leave the context window.
  3. Online reinforcement learning: The team proposes TA-GRPO, which splits long tasks into memory-linked segments and compares attempts at each stage. Rewards come from game-environment checks such as reaching a target location or finishing the current task.

Reported results

  • On four public benchmarks (including Minecraft, ViZDoom, OmniGameArena and GameWorld), Gamer-8B-CPT averaged 29.5. The highest-scoring general model listed scored 17.6.
  • On 40 tasks across seven mini-game categories not used in training, the base model scored 3.2%. CPT raised that to 15.0%, SFT to 39.2%, and reinforcement learning to 49.7%.
  • In the same comparison, the best general model, GPT-6.1-Sol, scored 17.1%.

The weights' availability and any licensing terms are not stated in the report.

Team

The paper lists eight co-first authors, five of whom contributed during internships at miHoYo. Zheng Yinghe, one of the paper's advisors, leads the AI NPC and gameplay technology team for the Honkai series. According to QbitAI, the team's work spans AI NPCs, game agents and R&D automation, and it has previously released GameScaling (scaling laws for cross-game continued pre-training) and Video-Arena (model understanding of dynamic game footage) with universities.

What this means

The reported gains arrive stage by stage: on held-out mini-games, each of CPT, SFT and reinforcement learning raised the score. The setup is also narrow in its inputs: the model sees pixels and outputs keyboard and mouse actions rather than calling game APIs. That makes the approach, in principle, transferable across titles. The paper's authors say this kind of capability could support automated testing, faster game iteration and new gameplay mechanics, lowering development costs. The 95% autonomy figure comes from one game, Honkai: Star Rail, and was reported by the authors.

Related Articles

research

Tunix agentic GRPO lifts Gemma 4 E2B from 9.4% to 95% win rate in Plane Strike

Google's developer blog details how Tunix's agentic GRPO pipeline trained Gemma 4 E2B to play Plane Strike with tool calls. The win rate on held-out boards rose from 9.4% to 95% after 100 steps on a single Cloud TPU v6e-8 host.

research

Hugging Face releases Qwen3.8-27B-pi-adapted LoRA, lifting SWE-bench Pro hard-51 pass@1 from 39.2% to 64.7%

Hugging Face released Qwen3.8-27B-pi-adapted, an Apache-2.0 LoRA adapter trained with reinforcement learning for the pi coding agent. According to MarkTechPost, it raises SWE-bench Pro hard-51 pass@1 from 39.2% to 64.7% and Terminal-Bench 2.1 from 39.5% to 55.8%.

research

Nvidia team: huge context windows often overkill, as summaries and lookups matched them in many tests

An Nvidia research team argues that very large context windows are often overkill and expensive, according to The Decoder. In many tests, periodic summaries and targeted lookups worked about as well.

research

Bytedance-led team argues three training steps are enough for AI agents, releases competitive open model

A team from Bytedance and several US universities argues in a new paper that three training steps are enough to build capable AI agents. The authors say stacked, multi-phase pipelines make it hard to tell which step helps and make reproduction expensive. According to The Decoder, the work is backed by a competitive open model.

Comments

Loading...