Introducing “Think in Games”: Tencent’s Innovative AI Training Framework
Tencent researchers have unveiled a groundbreaking approach to AI training named “Think in Games” (TiG), designed to enhance strategic reasoning in artificial intelligence models. This novel framework leverages the complex multiplayer environment of the popular game Honor of Kings to cultivate advanced decision-making skills in AI systems.
Strategic Reasoning Through Multiplayer Gaming
By utilizing gameplay data from Honor of Kings, the TiG framework integrates supervised learning with reinforcement learning, augmented by a method called Group Relative Policy Optimization (GRPO). This hybrid training approach enables AI models to develop nuanced strategic thinking, mimicking human-like reasoning in competitive scenarios.
Smaller Models Surpassing Larger Counterparts
Remarkably, the study revealed that under specific training conditions, more compact language models can outperform their larger counterparts. For instance, Tencent’s Qwen3-14B model achieved an impressive 90.9% accuracy in decision-making after just 2,000 training iterations, surpassing Deepseek R1’s 86.7% performance. This finding challenges the conventional assumption that bigger models always yield better results, highlighting the efficiency of the TiG framework.
Broader Implications Beyond Gaming
Beyond enhancing gameplay capabilities, the TiG framework also promotes explainable AI reasoning, a critical factor for transparency and trust in AI applications. The researchers suggest that this approach holds promise for diverse fields such as autonomous systems, strategic planning, and complex problem-solving tasks where clear rationale behind decisions is essential.
Future Prospects and Industry Impact
As AI continues to evolve, frameworks like TiG represent a significant step toward more intelligent and interpretable models. With the global gaming market projected to exceed $250 billion by 2025, integrating AI that can learn and reason through games offers exciting opportunities for both entertainment and practical AI deployment.
