In recent years, large language models (LLMs) have been widely applied in fields such as natural language processing and computer vision. However, research on imperfect-information games such as Guandan and Riichi Mahjong is still largely dominated by reinforcement learning approaches. These games present several core challenges, including difficulties in inferring hidden information, inefficient fusion of numerical and textual features, and limited adaptability in multi-stage strategies. Traditional algorithms such as game tree search and counterfactual regret minimization suffer from limitations including high computational complexity, slow convergence, and insufficient robustness, while existing LLMs suffer from problems such as high inference latency and inadequate dynamic information processing capability. To address these issues, this study proposes a dual-module fine-tuning framework combining stage-wise adaptive LoRA/QLoRA and a game-enhanced Adapter, with the Qwen3 series models as the baseline. Fine-tuning is conducted using high-quality datasets for Guandan and Riichi Mahjong, followed by adversarial experiments against mainstream reinforcement learning AIs (DanZero+, Tenhou AI). The experimental results show that although the fine-tuned large model still lags behind top-level reinforcement learning methods in terms of game win rate, it has become competitive in core capabilities such as hidden information inference and cross-modal feature fusion. This study provides a new approach for applying large models to imperfect-information games. However, the identified problems, including memory deficiencies and one-sided decision-making, still require further optimization.