Research map / Multi-agent systems and games / Human-AI and goal-conditioned
NitroGen: An Open Foundation Model for Generalist Gaming Agents
Loïc Magne, Anas Awadalla, Guanzhi Wang, Yinzhen Xu, Joshua Belofsky, Fengyuan Hu, Joohwan Kim, Ludwig Schmidt, Georgia Gkioxari, Jan Kautz, Yisong Yue, Yejin Choi and 2 more
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| pre-train on 40,000 hours of internet gameplay, then fine-tune on a held-out isometric roguelike (ablation) | NitroGen 500M · Isometric roguelike (held-out game) | task completion: average 10% relative improvement across data quantities | Improved |
| pre-train on internet gameplay, then fine-tune on a held-out 3D action-RPG (ablation) | NitroGen 500M · 3D action-RPG (held-out game) | task completion: average 25% relative improvement | Improved |
| pre-training on combat tasks of 3D action-RPG (ablation) | NitroGen 500M · 3D action-RPG, 30h, combat | task completion: 73.3% vs 48.3% (52% relative) | Improved |
| pre-training on game-specific tasks of 3D action-RPG (ablation) | NitroGen 500M · 3D action-RPG, 30h, game-specific | task completion: 66.6% vs 63.3% (5% relative) | No clear effect |
| fine-tune on original BC data only | NitroGen 500M · Boss 1 of 2D metroidvania | success rate: 18.7% -> 30.1% | Improved |
| add one round of self-imitation data (50/50 mix) (ablation) | NitroGen 500M · Boss 1 of 2D metroidvania | success rate: 30.1% -> 53.9% | Improved |
| add three rounds of self-imitation (ablation) | NitroGen 500M · Boss 1 of 2D metroidvania | success rate: 53.9% -> 90.5% | Improved |
| use more than one past frame as context, even with larger temporal gaps (ablation) | NitroGen · game benchmark | task success | No clear effect |
| generate 16-action chunks instead of single actions (ablation) | NitroGen · game benchmark | temporal consistency | Improved |
| keep only chunks with at least 50% non-zero actions (ablation) | NitroGen training data · pre-training | null-action over-prediction: 55% of data kept | Improved |
| predict joystick position via segmentation mask instead of direct regression (ablation) | annotation model · action extraction | joystick accuracy | Improved |
| use EMA weights instead of non-EMA weights (ablation) | NitroGen · evaluation | performance | Improved |
Closest papers by meaning
- AgentStudio: A Toolkit for Building General Virtual Agents (ICLR 2025 · similarity 0.89)
- GameVerse: Can Vision-Language Models Learn from Video-based Reflection? (ICML 2026 · similarity 0.88)
- Diffusion Models Are Real-Time Game Engines (ICLR 2025 · similarity 0.88)
- Scaling Offline Model-Based RL via Jointly-Optimized World-Action Model Pretraining (ICLR 2025 · similarity 0.88)
- From Multimodal LLMs to Generalist Embodied Agents: Methods and Lessons (CVPR 2025 · similarity 0.87)
- Boosting Virtual Agent Learning and Reasoning: A Step-Wise, Multi-Dimensional, and Generalist Reward Model with Benchmark (ICML 2025 · similarity 0.87)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
