Archive — CN — technode
Tencent Tests New Framework That Trains AI Models to Think Strategically ThroughGaming
Tencent has introduced a new AI training framework named 'Think in Games' (TiG) that enhances strategic reasoning in AI models using data from the multiplayer game _Honor of Kings_. Smaller models can outperform larger ones through a combination of supervised and reinforcement learning. The Qwen3-14B achieved 90.9% decision accuracy, indicating significant potential for applications beyond gaming.
Key points
- Tencent developed the 'Think in Games' (TiG) framework for training AI.
- TiG teaches strategic reasoning using real match data from the game _Honor of Kings_.
- The framework combines supervised learning, reinforcement learning, and Group Relative Policy Optimization (GRPO).
- The Qwen3-14B model reached 90.9% correct strategic decisions after 2,000 training steps, outperforming the larger Deepseek-R1 (86.7%).
- TiG's approach may improve AI's gameplay and reasoning capabilities, with applications expected beyond the gaming sector.
Why it matters
- The development aligns with 2025 IT trends emphasizing explainable AIs and strategic decision-making capabilities.
- This innovation reflects a growing trend in AI research focusing on efficient model training techniques.
- The success of smaller models suggests a shift in the industry towards optimizing resources in AI development.
Tencent's TiG framework highlights a major advancement in AI, using gaming to push boundaries in strategic reasoning, promising to influence various future applications in technology.