Playing Guandan with Large Language Models Based on Prompt Engineering

Large language models (LLMs) have undergone considerable development in recent years, enabling their applications across a broad spectrum of downstream tasks. However, there are concerns about their abilities in long-context comprehension, instruction following, and inference. In this paper, we investigate their performance in Guandan, an incomplete, ruleintensive game with minimal communication. We propose a framework for an LLM agent based on prompt engineering, incorporating a two-stage decision process. Our results indicate that, with the employment of error-correction prompts and the two-stage decision process, LLMs can adapt their behavior based on contextual cues. We conduct experiments on three base models: GPT-4o, DeepSeek-R1-30B, and DeepSeek-V3. Though the DeepSeek models perform better than GPT-4o, none of these models reach human-level performance, indicating that finetuning may be required for further improvements.