GameScaling: Open-Source Models Reveal Scaling Laws for Cross-Game Generalization

GameScaling: Open-Source General Game Models for Cross-World Transfer

miHoYo, in collaboration with Tsinghua University, the University of Hong Kong, and the National University of Singapore, has open-sourced the GameScaling series — the first general game models covering four scales (0.8B, 2B, 9B, 27B) for a total of eight models. All weights and inference code are released under the Apache 2.0 license. Project homepage: https://gamescaling.github.io; models on Hugging Face: https://huggingface.co/THUSI-Lab/GameScaling; code on GitHub: https://github.com/THUSI-Lab/GameScaling.

Agent Operation and Benchmark Performance

GameScaling agents operate in Minecraft (mining, crafting, fighting zombies), Doom (combat, navigation), and scene-escape tasks. Given the current screen, a task prompt, and recent interaction history, the model outputs keyboard and mouse actions for the next 200 ms, forming a continuous perception–action loop.

Figure 1: GameScaling-2B-S3 executing gathering, crafting, and combat tasks in Minecraft (source: GameScaling website demo video).

On the MCU benchmark's 30 Minecraft tasks, GameScaling-27B-S1 reaches an average success rate of 63.3% . The 0.8B–9B models (six variants) run on a single H200 or RTX 5090 GPU, achieving near real-time interaction at ~5 observations per second and 30 actions per second .

Core Research Question: Cross-World Transfer Under Shifting Rules

The research addresses a core question: how can AI transfer experience learned in one virtual world to another? Unlike the physical world where gravity and collision are constant, games can arbitrarily change visual style, physics, action spaces, and goals. The team asks whether the classic Scaling Law — more data, more diversity, larger models — still holds when the world's rules themselves shift.

Figure 2: The real world obeys fixed physical laws, providing a foundation for experience accumulation; games can alter visual style, physics, action spaces, and objectives, so whether Scaling Law holds has not been systematically studied (source: GameScaling paper Figure 1).

Controlled Experimental Design

To answer this, the team continuously pre-trained Qwen3.5 models on open-source game trajectories, connecting vision-language knowledge to executable actions. They designed a controlled study across three axes:

Model size: four scales from 0.8B to 27B.

Data duration: six budgets from 50 to 1,600 hours.

Data diversity: five mixtures ranging from 100% Minecraft to multi-game blends.

This yields 120 controlled configurations . Evaluation extends from training worlds (Minecraft, Doom) to unseen but similar 3D worlds, and further to 2D and 2.5D worlds with different perspectives and control schemes.

Two-Phase Learning Dynamic

The experiments reveal a two-phase learning dynamic :

Interface phase: Early training rapidly improves prediction of key presses, mouse movements, and action formatting. Gains transfer to worlds never seen during training.

World phase: Further improvement becomes increasingly dependent on which specific games the model has experienced. Benefits concentrate on training-covered and similar worlds; loss on distant worlds may even rise.

Figure 4: In early training, key-press loss, mouse loss, format loss, and mouse-direction accuracy improve quickly. S1 = single-game trajectory mix; S4 = multi-game mix (source: GameScaling paper Figure 2).

Data Diversity Is Critical for Transfer

Data diversity proves critical. At the same 1,600-hour budget, multi-game training scores 2.8× higher on average across the seven unseen 3D worlds in OmniGameArena than single-game (Minecraft-only) training, across all model sizes and mixture ratios. Skills such as movement, turning, and aiming do transfer across game boundaries.

Figure 5: GameScaling performing tasks in OmniGameArena's unseen 3D environments — movement, shooting, obstacle traversal, and scene escape (source: GameScaling website demo video).

Model Size Shows Nuanced Effects

Model size shows a nuanced effect: larger models absorb limited experience more efficiently, but if trained long on a single game they can regress more sharply on dissimilar worlds. The authors conclude that learning faster does not automatically mean transferring farther ; parameter count, data volume, and data mixture must be co-designed.

Empirical Scaling Law and Its Boundary

The team fits an empirical Scaling Law linking validation loss to the three factors for training-covered worlds. However, for distant unseen worlds, further increasing data and model size does not yield stable improvement . GameScaling thus maps both the potential and the boundary of current continual pre-training: it can accumulate transferable operational experience, but broader generalization demands wider world coverage and interaction-driven learning mechanisms.

Real-Time Deployment Constraint

Real-time deployment adds another constraint: the game world does not wait. The team compared "pause game while model thinks" versus "game runs continuously." In the real-time evaluation setting, GameScaling-9B-S3 averages 25.3 across four benchmarks , surpassing the best score of 19.2 from tested frontier models (OpenAI, Anthropic, Google). This gap, though influenced by deployment and serving latency, highlights a practical requirement for game AI: decisions must arrive in time to affect the outcome.

Figure 6: Real-time combat demo — left: Claude Opus 5.0, right: GameScaling-9B. Game runs continuously during model inference; recorded latency includes deployment and calling overhead (source: GameScaling website demo video).

Authors' Position and Future Directions

The authors emphasize that GameScaling focuses on exploring cross-world scaling phenomena, not on final game-performance optimization. They hope the released models and findings will guide future game-data collection, construction, and mixing, giving a principled basis for scaling training, and point toward the next steps: broader game experience and mechanisms that learn from success/failure feedback during actual interaction. The ultimate vision is intelligent NPCs that retain experience across worlds — observing terrain, choosing positions, adapting to battle dynamics — so that players and their AI companions can carry skills from one adventure into the next.