An Analysis of Zero-Sum Game Agent Training Based on Large Language Models and Self-Play Mechanisms

This paper carries out a review of competitive agent training under zero‑sum game settings, analysing persistent drawbacks found in traditional self‑play reinforcement learning, including inflexible decision‑making policies and weak long‑horizon reasoning capacity. Game agents built purely upon large language models are incapable of sustaining iterative upgrades amid ongoing adversarial contests. Accordingly, the present work investigates viable integration schemes that unite large language models and self‑play training systems. It first lays out the definition of zero‑sum games, illustrates standard self‑play training workflows, and revisits fundamental algorithms from game theory. Four practical integration patterns are classified, namely auxiliary policy generation, reward evaluation optimisation, accumulation of adversarial experience, and the construction of diversified training opponents. Relying on theoretical formula derivation and comparative experimental statistics, this paper evaluates widely adopted modern frameworks against five core benchmarks: reasoning performance, training productivity, generalisation capability, computational expenditure, and training stability, so as to compare their distinct strengths and weaknesses. Evaluation findings indicate that each standalone integration paradigm bears intrinsic flaws. Frameworks adopting dynamically‑generated opponents deliver optimal generalisation performance, while optimised reward mechanisms enable the fastest training convergence. Nonetheless, all hybrid training architectures confront common bottlenecks such as LLM hallucination, prohibitive computational overhead, and oscillating gradient signals during model training. This review identifies several high‑potential research directions: dynamic multi‑module collaborative structures, streamlined self‑play workflows, retrieval‑augmented technologies to restrain hallucination, as well as the development of general‑purpose game agents. The outcomes of this analysis can supply theoretical references for future advances in adversarial decision‑making artificial intelligence.