The paper explores the use of self-play training in a two-player adversarial language game called *Adversarial Taboo* to enhance the reasoning abilities of large language models (LLMs). In this game, an attacker and a defender communicate around a target word visible only to the attacker, with the attacker aiming to induce the defender to speak the target word unconsciously, and the defender trying to infer the target word from the attacker's utterances. The authors select several open-source LLMs and let each act as both the attacker and defender, playing against a copy of itself. Through reinforcement learning on the game outcomes, they observe significant improvements in the LLMs' reasoning performance on various benchmarks. Iteratively adopting this self-play process further promotes the LLMs' reasoning abilities. The code for this method is available at https://github.com/Linear95/SPAG. The paper also includes a detailed methodology, experimental setup, and results analysis, demonstrating the effectiveness of the proposed approach.The paper explores the use of self-play training in a two-player adversarial language game called *Adversarial Taboo* to enhance the reasoning abilities of large language models (LLMs). In this game, an attacker and a defender communicate around a target word visible only to the attacker, with the attacker aiming to induce the defender to speak the target word unconsciously, and the defender trying to infer the target word from the attacker's utterances. The authors select several open-source LLMs and let each act as both the attacker and defender, playing against a copy of itself. Through reinforcement learning on the game outcomes, they observe significant improvements in the LLMs' reasoning performance on various benchmarks. Iteratively adopting this self-play process further promotes the LLMs' reasoning abilities. The code for this method is available at https://github.com/Linear95/SPAG. The paper also includes a detailed methodology, experimental setup, and results analysis, demonstrating the effectiveness of the proposed approach.