120 controlled runs decode game-agent CPT: a 9B model overtakes frontier LLMs in real-time play

Scaling in Games: Continued Pre-Training For Embodied Agents In Diverse Virtual Worlds

Kuan Zhang, Yukun Chen, Zhihao Yang, Yue Su, Xiangnan Wu, Zirong Chen, Run Luo, Jinkun Hou, Tao Tan, Yinhe Zheng, Yiming Li

cs.AI, cs.CV, cs.LG

2026-10-20

120 controlled CPT runs on game agents fit a three-factor scaling law (R2=0.995), beat frontier models by up to 5x in real-time play, but distant-world transfer stays unsolved.

What problem this solves

Robotics scaling lore says more diverse data helps, and it holds there because every experience comes from one physical world where gravity, contact, and object permanence behave identically. Virtual worlds break that assumption. Games differ in visuals, physics, controls, and objectives all at once. Humans pick up an unfamiliar game in minutes; whether an agent can buy the same adaptability by training on many games had never been tested under controlled conditions. Existing game foundation models (SIMA, GameTARS, and the like) change data volume, mixture, and model size together, so each factor's contribution stays confounded. A team from miHoYo's Honkai AI group with Tsinghua and other universities ran a 120-configuration grid to settle the question.

Method

Three knobs, everything else fixed: Qwen3.5 backbones at four sizes (0.8B/2B/9B/27B), gameplay budgets at six levels (50 to 1,600 hours), and five mixtures (S1 to S5) where the Minecraft share drops from 100% to 0%. Full-parameter continued pre-training, 3.5×10^11 tokens total, roughly 45,000 H200 GPU-hours.

• One keyboard-mouse interface: 5 Hz observations, 30 Hz control, a 200 ms action packet holding a mouse delta plus six 33 ms key substeps, preceded by a <think> span.
• 85% of tokens are gameplay trajectories (VPT, D2E, OpenP2P); 15% are fixed general vision-language data to keep backbone abilities from decaying. GPT-5.6 adds sparse retrospective rationales (about one per sample) that describe intent without touching action labels.
• Evaluation is laddered over three levels: in-corpus worlds (MCU, ViZDoom), unseen related worlds (OmniGameArena, seven UE5 3D games), and distant worlds (GameWorld, 170 tasks across 34 environments, mostly 2D and 2.5D).

The cleanest design decision: S2 through S5 draw from the same fixed pool of non-Minecraft recordings (124 streams, 79 games), so only the proportion moves. The number of worlds and the spread of experience over them are decoupled.

Results

• Training runs in two phases. The first 50 hours are an interface phase: validation loss falls below the base model on all three levels, games never seen included. A world phase follows, with gains narrowing to trained and related worlds.
• Minecraft and in-corpus losses fit a three-factor scaling law at R2 = 0.982 and 0.995; refits holding out the 27B models, the 1,600-hour checkpoints, or the 50-hour checkpoints still predict within RMSE 0.0016-0.0027. OOD loss fits no law (beta goes to zero): its gain over the base model is complete at 50 hours and then slowly erodes.
• The 27B model trained only on Minecraft lands below its own base model on all three non-Minecraft benchmarks, a drop of about 17 points on GameWorld. Small models on diverse data outperform large models on Minecraft-only data.
• On MCU, GameScaling-27B-S1 averages 63.3% success against 48.0% for the best Minecraft specialist (MAIN-VLA, 7B). VPT, behavior-cloned on tens of thousands of hours, manages 4.8%, while GameScaling-0.8B-S3 reaches 36.0%. GUI-task success grows sixfold from 0.8B to 27B.
• Against frontier models under one interface: with time paused, CPT multiplies the 27B backbone's MCU score by 6x (9.3 to 54.0), and the 9B/27B models match open models near 300B parameters. Real-time play flips the hierarchy. Frontier models take 4.0 to 23.7 seconds per step and lose roughly 90% of their score on MCU and ViZDoom; GameScaling at 0.8B to 9B keeps 88% of its paused performance and decides in 0.13 to 0.69 seconds, so GameScaling-9B-S3 averages 25.3 across four benchmarks versus 19.2 for the best frontier model. The distant world is the exception: frontier models score about twice as high on GameWorld, where CPT helps only at 0.8B and 2B.

Why it matters

This is the first controlled study to separate the three axes of virtual embodied CPT, and the recipe guidance is concrete: with a limited budget, match the training mixture to the target domain instead of broadening diversity blindly, and do not expect data volume to fix worlds outside the corpus. The real-time finding is the harder one for practitioners: in a world that does not pause, latency is part of competence, because a delayed action lands in a state that has already moved on. Weights and code are open (THUSI-Lab/GameScaling).

Limitations

• The entire 120-condition grid used a single training seed, as the authors state; 45,000 GPU-hours made repetition impractical, so there is no variance estimate.
• CPT is behavior cloning with no environmental feedback. Whether the distant-world gap comes from the imitation objective or from insufficient game coverage cannot be separated.
• The diversity axis only varies the Minecraft proportion; scaling the number of distinct worlds was not tested.
• Appendix C.1 concedes that newer general backbones improve some benchmarks by more than game-specific CPT does, so backbone progress remains a parallel route.
• The real-time comparison routes frontier models through provider APIs while GameScaling runs on a local H200 cluster. The authors argue this mirrors deployment reality, but the hardware is not matched.

Terms

• CPT (continued pre-training): further training a general backbone on domain data, here gameplay trajectories.
• Behavior cloning: imitating recorded player actions, with no success or failure signal.
• Interface phase / world phase: early training learns cross-game control formats, later training learns each world's own dynamics.
• MCU / OGA / GameWorld: three benchmarks laddering from Minecraft tasks to unseen 3D worlds to far-distribution 2D/2.5D games.
• Hill number: an entropy-based effective count of worlds; 1 for S1, 60.1 for S4.

Source

• arXiv 2610.28698v1
• PDF
• Hugging Face
• GitHub · https://github.com/THUSI-Lab/GameScaling

Related papers

• DeLoop cuts the Qwen3-8B loop rate from 57.3% to 6.1% with no reward
• Family-style traces beat Fable 5's own: +7.1 MBPP+, +3.7 HumanEval+
• Text-only refusal directions lift Qwen3-VL video refusal from 28.8% to 88.2%
• WildAWM trains Qwen-35B only in synthetic PCs, for a 10.6-point gain
• Issue-Guided SSL hits 0.805 utility, 61.7% fewer educator tokens than fixed reviews
• Two 27B models 2.6 points apart on public sets score 70.1% vs 44.2% on contrasts

All paper explainers