Scaling in Games

Continued Pre-Training for Embodied Agents in Diverse Virtual Worlds

Kuan Zhang^1,2,*,‡, Yukun Chen^1,3,*,‡, Zhihao Yang^1,3,*,‡, Yue Su⁴, Xiangnan Wu³, Zirong Chen², Run Luo^1,5,‡, Jinkun Hou⁶, Tao Tan¹, Yinhe Zheng^1,†, Yiming Li^2,†

¹miHoYo Honkai AI R&D Team   ²College AI, Tsinghua University   ³University of Chinese Academy of Sciences   ⁴MMLab, The University of Hong Kong   ⁵National University of Singapore   ⁶Peking University
^*Equal contribution   ^†Corresponding author   ^‡Work done while interning at miHoYo

If you find GameScaling useful,
please upvote our Daily Paper on Hugging Face
and star our GitHub repository.
Thank you for your support!

This repository: GameScaling-27B-S3. Qwen3.5-27B continued pre-trained on a balanced mixture of 27% Minecraft and other games (S3), 1,600 h of gameplay, 7,020 training steps; 54.7 GB of bfloat16 weights on 2 GPUs (tensor parallel). Serve it with bash scripts/serve_vllm.sh THUSI-Lab/GameScaling-27B-S3 8000 2 (§1). The inference client, prompts and scripts are in THUSI-Lab/GameScaling; all eight checkpoints are in the GameScaling collection.

Overview

GameScaling models are Qwen3.5 vision-language models with continued pre-training on keyboard-and-mouse gameplay. Each call takes the current game frame (plus a short history of past frames and actions) and returns the next 200 ms of keyboard and mouse actions: 6 steps of 33 ms each, plus one mouse movement and scroll for the whole chunk. The model may think before it acts.

This repository holds the GameScaling-27B-S3 weights. The inference client and the standard system-prompt template are in THUSI-Lab/GameScaling; each of the eight checkpoints is its own repository (below, and in the GameScaling collection). Together they are everything needed to serve a checkpoint and run it in a game; there is no training code. Paper results and analysis are on the project page.

Checkpoints

• S1 is trained on Minecraft gameplay only. S3 is a balanced mixture: 27% Minecraft, the rest from other games. All checkpoints use 1,600 h of gameplay.
• Weights are bfloat16, from the Qwen3.5-0.8B / 2B / 9B / 27B backbones.

File tree

THUSI-Lab/GameScaling-<size>-<mixture>   one repository per checkpoint
├── config.json                 Qwen3.5 architecture, bfloat16
├── model.safetensors           weights (27B: 2 shards + index)
├── chat_template.jinja         required, renders history turns
├── tokenizer.json
├── tokenizer_config.json
├── generation_config.json
├── preprocessor_config.json
├── processor_config.json
└── video_preprocessor_config.json

THUSI-Lab/GameScaling
├── config.json                 index of the eight checkpoint repositories (no weights)
├── prompts/
│   ├── system_prompt_template.txt  standard prompt, 3 slots (§2)
│   └── example_doom_battle1.txt    complete Doom prompt (§2)
├── examples/
│   ├── doom_battle1.jpg        test frame (1280x720)
│   ├── doom_battle1_action.txt the model's reply on that frame
│   └── run_example.py          frame in, action out (§3)
├── gamescaling/                inference client
│   ├── prompt.py               system prompt builders
│   ├── policy.py               client and request body
│   ├── action.py               action format and parser
│   └── keymaps.py              key names <-> browser keys
├── scripts/serve_vllm.sh       start a vLLM server
├── tests/test_core.py          offline tests (no GPU)
├── assets/                     images for this card
├── requirements.txt            client deps: requests, Pillow
└── LICENSE                     Apache 2.0

1. Download and serve

Download the client and scripts from THUSI-Lab/GameScaling:

pip install -U huggingface_hub
huggingface-cli download THUSI-Lab/GameScaling --exclude "assets/*" --local-dir ./GameScaling
cd GameScaling

Serve a checkpoint with vLLM (tested with 0.17.0); vLLM fetches the weights from its repository:

pip install "vllm>=0.17"
bash scripts/serve_vllm.sh THUSI-Lab/GameScaling-9B-S1 8000        # 0.8B-9B fit on one GPU
bash scripts/serve_vllm.sh THUSI-Lab/GameScaling-27B-S1 8000 2     # 27B: tensor parallel 2

To keep the weights in a local folder instead, download the checkpoint repository and pass its path:

huggingface-cli download THUSI-Lab/GameScaling-9B-S1 --local-dir ./GameScaling-9B-S1
bash scripts/serve_vllm.sh ./GameScaling-9B-S1 8000

serve_vllm.sh CKPT [PORT] [TP] starts vLLM with --served-model-name gamescaling, bfloat16, --max-model-len 32768 (system prompt + 5 history frames + reasoning budget) and --limit-mm-per-prompt '{"image": 8}' (5 history frames + the current one, with headroom). Set GPU_MEM_UTIL to change the memory fraction (default 0.85).

Check that the server is up and the client can reach it:

curl -s http://127.0.0.1:8000/v1/models          # should list "gamescaling"
pip install -r requirements.txt
python tests/test_core.py                         # offline checks, no server needed

Things to keep as shipped:

• Chat template. Use the chat_template.jinja in the checkpoint repository (vLLM loads it by default). Do not replace it with the stock Qwen template: this one renders past turns without reasoning as <think>\n\n</think>\n\n<|action_start|>..., token-for-token as in training.
• Special tokens. Requests must set skip_special_tokens: false, or the server strips the <|action_start|> / <|action_end|> tags. build_request already does this.
• One server per GPU group. One vLLM replica can serve several games at once, but long prefills (one image per history turn) make each step slower when many games share it.

2. System prompt

prompts/system_prompt_template.txt is the standard system prompt. Fill in three placeholders and keep everything else exactly as written: the model decides its output tags from the exact text of the Output Format section.

The mouse line, copied verbatim:

- Mouse X Y -> aim / rotate the camera (X>0 right, Y>0 down)

Note. Lines such as **Game Rules**, **Output Format** and **Explanation** are plain-text section headers that the model saw in training. They are part of the prompt, not Markdown formatting: send them to the model unchanged, including the asterisks.

Complete example: Doom Battle-1

This is the full system prompt for the ViZDoom Battle-1 map, as in prompts/example_doom_battle1.txt. It is the same text in both reasoning modes (§4). Long lines scroll horizontally; each line break below is a real \n in the prompt.

You are a gaming expert, and you are currently playing Doom Battle-1. You are proficient in keyboard and mouse operations. You understand game mechanics and combat pacing, can quickly extract key information from the current screen, and can think and make precise decisions at critical moments. Based on the current screen, plan the next 200ms of actions, consisting of 6 steps spaced 33ms apart. Each step lasts 33ms, until the next step begins. If the current situation continues the previous strategy, output the action directly. Think only when the situation changes significantly, the previous analysis is no longer valid, or a new objective appears.
**Game Rules**
You are playing Doom, a first-person shooter, in an arena full of monsters. Survive and kill as many monsters as you can.

## What you see on screen
- First-person view.
- The crosshair at screen centre marks where shots land.
- The status bar along the bottom shows HEALTH on the left and AMMO on the right.
- Monsters keep arriving for the whole episode; some throw fireballs from range.

## Rules
- Score is the number of monsters you kill in the episode.
- Firing costs ammo, and the episode ends as soon as your health reaches 0, so dodging matters as much as shooting: standing still in the open is how most episodes end early.
- Health packs and ammo boxes lie around the map; walk over one to pick it up. Running out of either is fatal, so collect them while fighting.
- A target is hit when it is under the crosshair, so turn with the mouse until the monster is centred before firing.
- The map is one large open arena, so monsters can close in from any direction.
**Common Keys in Doom Battle-1**
- W -> Move forward
- A -> Strafe left
- S -> Move backward
- D -> Strafe right
- LB -> Fire weapon
- Ctrl -> Fire weapon (keyboard alternative)
- Shift -> Run
- Space -> Open doors / use switches
- one -> Select weapon slot 1
- two -> Select weapon slot 2
- three -> Select weapon slot 3
- Mouse X Y -> aim / rotate the camera (X>0 right, Y>0 down)
**Output Format**
**Action only (usual)**
<think></think><|action_start|>X Y Z ; k1 k2 k3 ; k4 k5 ; k6 ; k7 ; k8 ; k9 k10<|action_end|>
**Thinking + action (when necessary)**
<think>reasoning</think><|action_start|>X Y Z ; k1 k2 k3 ; k4 k5 ; k6 ; k7 ; k8 ; k9 k10<|action_end|>
**Explanation**
1. **Mouse Movement**: First, specify the relative displacement X, Y (X>0 means move right, Y>0 means move down) and scroll amount Z (Z>0 means scroll up).
2. **Key Sequence**: Then list 6 groups of keys; within each group, keys are separated by spaces, and groups are separated by semicolons.
- Each group can contain up to 4 keys.
- If a group has no keys, leave it empty but keep the `;`.
3. Only output a plain string that conforms to the above format — no line breaks and no quotation marks.
**Key Naming Rules**
- Number keys `0-9`: use lowercase English words, e.g., `zero` for `0`, `one` for `1`, ... `nine` for `9`.
- Function keys `F1-F12`: use capitalized English words, e.g., `One` represents `F1`, `Two` represents `F2`, and so on.
- Mouse buttons: use `LB`, `RB`, `MB` for the left, right, and middle mouse button.
- Letters: use the uppercase letter, e.g., `A`, `D`, `W`, `S`.
- Arrow keys: use `Up`, `Down`, `Left`, `Right`.
- Modifier/special keys: `Shift`, `Ctrl`, `Alt`, `Tab`, `Caps`, `Esc`, `Space`, `Enter`, `Back` (backspace), `Delete`, `Insert`, `Home`, `End`, `Pause`.

The same prompt from code (tests/test_core.py checks that it matches the file byte for byte):

from gamescaling import build_system_prompt

DOOM_RULES = open("prompts/example_doom_battle1.txt").read() \
    .split("**Game Rules**\n", 1)[1].split("\n**Common Keys in", 1)[0]

system = build_system_prompt(
    "Doom Battle-1",
    rules=DOOM_RULES,
    key_bindings=[
        ("W", "Move forward"), ("A", "Strafe left"), ("S", "Move backward"), ("D", "Strafe right"),
        ("LB", "Fire weapon"), ("Ctrl", "Fire weapon (keyboard alternative)"), ("Shift", "Run"),
        ("Space", "Open doors / use switches"), ("one", "Select weapon slot 1"),
        ("two", "Select weapon slot 2"), ("three", "Select weapon slot 3"),
    ],
    mouse_aim=True,   # appends the "Mouse X Y -> ..." line
)

Filling the text file directly gives the same result:

tpl = open("prompts/system_prompt_template.txt").read().rstrip("\n")
system = (tpl.replace("{GAME_NAME}", "Doom Battle-1")
             .replace("{GAME_RULES}", DOOM_RULES)
             .replace("{KEY_BINDINGS}", "- W -> Move forward\n- A -> Strafe left\n..."
                      "\n- Mouse X Y -> aim / rotate the camera (X>0 right, Y>0 down)"))

3. Minimal usage

from PIL import Image
from gamescaling import GameScalingPolicy, PolicyConfig

policy = GameScalingPolicy(system, PolicyConfig())   # system prompt from §2; prefills <think>
policy.reset()                                       # call at the start of every episode
decision = policy.act(Image.open("frame.png"))       # one call = one 200 ms decision
a = decision.action
a.mouse      # (X, Y, Z): mouse movement for this chunk (per-mille of the screen) + scroll
a.steps      # 6 groups: the keys held during each 33 ms step
a.parsed     # False = could not be parsed; execute as a no-op
decision.reasoning   # the model's thinking for this step ("" if it acted directly)

Pass a to your executor. Run 6 steps of 33 ms, holding that step's keys during each one. Keys that appear in two consecutive groups stay held rather than being pressed again. Mouse movement is X/1000 * screen width and Y/1000 * screen height. You can spread it over the 200 ms or send it at once. The frame is sent at its native size; set PolicyConfig(image_size=(w, h)) to resize it first.

Test frame

examples/doom_battle1.jpg is a frame from a Doom Battle-1 episode played by GameScaling-9B-S3. A monster stands just left of the crosshair.

The model's reply on this frame (examples/doom_battle1_action.txt), with the system prompt from §2:

<think>

</think>

<|action_start|>-37 -3 0 ; LB S ; LB ; LB ; LB ; LB ; LB<|action_end|>

It acts without thinking, turns left by 37/1000 of the screen width, and holds fire for the whole 200 ms (with a short step back in the first 33 ms). The shot killed the monster.

Run the same frame against your server:

python examples/run_example.py --offline       # no server
python examples/run_example.py                 # prefill <think>
python examples/run_example.py --no-reasoning  # no thinking

Each reply should parse into a valid action chunk (parsed=True); the script exits non-zero otherwise. Sampling is at temperature 1.0 and the reference reply had five earlier frames in context, so live replies differ in detail.

4. Reasoning modes and message structure

The assistant turn is always prefilled, and the request continues it (continue_final_message=True). The system prompt is the same in both modes.

<think>\n\n</think> is exactly how the chat template renders a turn without reasoning, so the no-reasoning mode stays in the training format. Do not send the request without a prefill.

Each user turn is the text current game screen followed by the frame. Multi-turn messages (policy.py):

system     system prompt (§2)
user       "current game screen" + frame t-k
assistant  action taken at t-k (no reasoning)
...        (last history_len turns, default 5)
user       "current game screen" + current frame
assistant  "<think>"  or  "<think>\n\n</think>"

Past turns keep only the action (history_think=0). A turn that failed to parse is stored as an explicit no-op, <|action_start|>0 0 0 ; ; ; ; ; ; <|action_end|>, not as the raw bad output.

Client defaults: temperature=1.0, max_tokens=2048, stop=["<|action_end|>"], skip_special_tokens=false, JPEG frames (quality 90) at native resolution.

5. Action format

<|action_start|>X Y Z ; g1 ; g2 ; g3 ; g4 ; g5 ; g6<|action_end|>

• X Y: mouse movement over the whole 200 ms, in per-mille of the screen (X = Σdx / screen width × 1000), clipped to ±1000. X>0 is right, Y>0 is down.
• Z: scroll notches, clipped to ±5. Z>0 is up.
• g1..g6: six 33 ms steps. Each group lists the keys held during that step, at most 4. An empty group releases all keys, but the ; must stay.
• <|action_start|> / <|action_end|> are special tokens. Requests must set skip_special_tokens: false (build_request does), or the server strips the tags. parse_action also accepts a bare X Y Z ; ... as a fallback.
• The action is read only after the last </think>: a model that thinks often quotes the format inside its reasoning. An unclosed <think> means the generation was cut off; it is executed as a no-op (Action.truncated=True).

Training key vocabulary (74 tokens, case-sensitive, gamescaling.action.KEY_VOCAB):

Tokens outside the vocabulary are dropped and recorded in Action.dropped. keymaps.TO_PLAYWRIGHT and keymaps.TO_UE5 map them to browser and Unreal Engine key names; translate_controls rewrites browser key names in a text into training tokens.

6. Running it in a game

1. Write the system prompt for your game (§2).
2. Capture the current frame and call policy.act(frame) once per decision.
3. Execute the returned Action as in §5: held keys for 6 steps of 33 ms, plus mouse movement and scroll.
4. Call policy.reset() when an episode ends.

Before trusting the results, check:

• Parse rate. Decision.action.parsed should almost always be True. Many failures, or many Action.truncated, usually mean a replaced chat template, stripped special tokens, a missing prefill, or a too-small max_tokens.
• Actions take effect. Each chunk runs in wall-clock time (6 × 33 ms). If the game renders slowly, for example several game instances sharing one GPU, held keys barely move the character and nothing reports an error. Watch a recording before reading any numbers.
• Several episodes. Single episodes vary a lot at temperature=1.0.

7. Tests

python tests/test_core.py      # or: python -m pytest tests/

Offline checks of prompts, request bodies and action parsing; no GPU or server needed.

License

GameScaling is a purely scientific research project. We carry out no commercial activity of any kind with it: nothing here is sold, licensed for a fee, or used in any product or service.

• Models and code: Apache-2.0. The weights and the inference code are released under the Apache License 2.0. The models are continued pre-trained from Qwen3.5, which is also Apache 2.0.
• Gameplay data is not released. The training corpus stays private; game frames in THUSI-Lab/GameScaling (e.g. examples/doom_battle1.jpg) only illustrate how to use the model.
• Game copyright belongs to the publishers. All game titles, footage, copyrights, trademarks and other game assets referenced here belong to their respective publishers and owners. The Apache 2.0 licence covers our weights and code only and grants no right to any game content.

Citation

@article{zhang-scaling-games,
  title = {Scaling in Games: Continued Pre-Training for Embodied Agents in Diverse Virtual Worlds},
  author = {Zhang, Kuan and Chen, Yukun and Yang, Zhihao and Su, Yue and Wu, Xiangnan and Chen, Zirong and Luo, Run and Hou, Jinkun and Tan, Tao and Zheng, Yinhe and Li, Yiming},
  journal = {arXiv preprint arXiv:2610.28698},
  year = {2026}
}