Go to ICLR 2025 Conference homepageGameInstruct: Teaching Machines to Reason via Chameleon Game
Self-play has emerged as a promising approach for generating alignment data to reduce the data annotation costs during the alignment process. By introducing specific game rules and utilizes the model’s own language capabilities to generate data samples, self-play has achieved promising results. However, traditional self-play methods face two major challenges: insufficient data diversity during self-iterative training and difficulties in reward signal design. To solve these problems, this paper introduces GameInstruct, a complex multi-player adversarial environment that increases the complexity of self-play generated data during self-iterative training. Specifically, we employ the ``Chameleon Game'', where interactions between multiple players raise the diversity of the generated data, improving the model’s reasoning abilities, Additionally, we further propose a dynamic reward algorithm to capture signals within player conversations during the whole game. Experimental results show that compared to existing self-play methods, GameInstruct achieves significant improvements on the HuggingFace Open-LLM-Leaderboard reasoning benchmark while demonstrating continuous improvement and increasing data diversity during self-iterative training.