New Reinforcement Learning Discovery: Training AI Reasoning Skills Through Games Without Mathematical Samples}

Editor | Bai Cai Ye

Recently, researchers from Rice University, Johns Hopkins University, and NVIDIA made a groundbreaking discovery in reinforcement learning: AI can significantly improve its reasoning skills by playing simple games, without any mathematical or disciplinary training data.

Previous studies showed that reinforcement learning (RL) on mathematical problems could enhance models' reasoning abilities, but these improvements were often limited to the same domain. Now, the team from Rice, Johns Hopkins, and NVIDIA took a different approach: they trained multimodal large language models (MLLM) to play games like Snake and rotation puzzles, without any math or multi-disciplinary data, and still achieved remarkable reasoning improvements.

• Paper Title: Play to Generalize: Learning to Reason Through Game Play
• Paper Link: https://arxiv.org/abs/2506.08011
• Project Homepage: https://yunfeixie233.github.io/ViGaL/

No Math Samples, Game Training Breaks Through in Math Benchmarks

Recent research shows that RL often outperforms supervised fine-tuning (SFT) in cross-domain generalization. Past work proved models trained on math problems could extend reasoning to physics or navigation tasks, but these successes were usually within the same domain.

Figure 1: We found that training models with reinforcement learning on simple games like Snake enables emergent cross-domain generalization, improving performance in math and multi-disciplinary tasks.

This breakthrough demonstrates stronger cross-domain generalization: models trained solely on game environments like Snake and rotation games show significant improvements in mathematical, spatial, and multi-disciplinary reasoning benchmarks.

• Math reasoning improvement: Without math samples, RL training on games boosts ViGaL's performance on MathVista by an average of 2.9%, compared to 2.4% with high-quality math data.
• Multi-disciplinary reasoning: On MMMU multi-disciplinary benchmarks, ViGaL surpasses RL-trained models like Jax-7B by 5.4 percentage points.
• Maintaining general capabilities: Unlike previous RL models that often sacrificed general vision skills, ViGaL retains broad abilities while enhancing reasoning.

Figure 2: Using only game training, models improve by 2.9% in math reasoning (left) and 2.0% in multi-disciplinary reasoning (right), surpassing previous RL methods trained on math or multi-disciplinary data.

Why is game training so effective?

Figure 3: We trained models on Snake and rotation games using reinforcement learning. The models receive game environment images and text instructions, reason through actions, and execute moves. Rewards from the environment guide learning. This game training endows models with reasoning skills that transfer to downstream math and multi-disciplinary tasks.

Why does playing games enhance math ability? This aligns with cognitive science principles.

Think about our own childhood: building blocks helped us understand spatial concepts, hide-and-seek taught position relations, and puzzle games fostered logical thinking. These activities, seemingly 'play,' build the foundation of abstract reasoning—pattern recognition, spatial inference, and causal reasoning.

Cognitive science also confirms that games are valuable tools for exploring human cognition. Researchers study planning with Connect Four and problem-solving mechanisms with virtual tools.

Inspired by this, the team designed two complementary training games:

Snake: A classic strategy game where the model controls a snake on a 10×10 grid, avoiding walls and itself while collecting apples. It trains path planning, obstacle avoidance, and spatial navigation—skills directly related to coordinate geometry and function graph understanding.

Rotation: A 3D spatial reasoning game where the model observes an object from two views—initial and rotated—and determines if it rotated 90° or 180°. This game trains spatial geometry reasoning, directly related to angles and lengths.

The two games are complementary: Snake enhances 2D coordinate reasoning, while Rotation focuses on angles and lengths. Experiments show that joint training on both games yields better results, demonstrating the scalability of diverse game-based training.

Conclusion: The Era of Synthetic Tasks

ViGaL reveals a promising trend: when high-quality human data is scarce and performance saturates on simple tasks, carefully designed games as synthetic tasks can open new avenues for developing multimodal reasoning abilities.

Compared to traditional training, this gamified approach offers:

• Low cost: No manual annotation, infinitely scalable
• Significant effect: Outperforms math-specific training without math samples
• High scalability: Can combine multiple tasks to further improve performance
• Good generalization: Avoids 'overfitting' to specific tasks, maintaining broad capabilities

More importantly, ViGaL suggests a simple but profound insight: cultivating fundamental reasoning skills through game-based training may be as crucial as direct task learning. Just as humans develop mathematical thinking through various cognitive activities, AI can benefit from similar foundational training.

As scaling laws face limitations, ViGaL reminds us that sometimes, letting AI 'play games' is more effective than just 'drilling problems.'