Enhancing Commentary Strategies for Imperfect Information Card Games: A Study of Large Language Models in Guandan Commentary

TL;DR

The paper tackles generating insightful commentary for imperfect-information games by integrating reinforcement learning with large language models in a modular Guandan commentary agent. It introduces three components—State Commentary Guider, ToM-Based Strategy Analyzer, and Style Retrieval—to transform game states and strategic reasoning into contextual Chinese narration, leveraging retrieval-augmented generation and Theory of Mind to enhance depth and personalization. A formal mechanism-design-inspired guarantee ensures compliant narration, and experiments show open-source LLMs with RAG can outperform GPT-4 on multiple metrics in Guandan. The work demonstrates substantial improvements in commentary quality and lays groundwork for extending to other complex games and multimodal data.

Abstract

Recent advancements in large language models (LLMs) have unlocked the potential for generating high-quality game commentary. However, producing insightful and engaging commentary for complex games with incomplete information remains a significant challenge. In this paper, we introduce a novel commentary method that combine Reinforcement Learning (RL) and LLMs, tailored specifically for the Chinese card game \textit{Guandan}. Our system leverages RL to generate intricate card-playing scenarios and employs LLMs to generate corresponding commentary text, effectively emulating the strategic analysis and narrative prowess of professional commentators. The framework comprises a state commentary guide, a Theory of Mind (ToM)-based strategy analyzer, and a style retrieval module, which seamlessly collaborate to deliver detailed and context-relevant game commentary in the Chinese language environment. We empower LLMs with ToM capabilities and refine both retrieval and information filtering mechanisms. This facilitates the generation of personalized commentary content. Our experimental results showcase the substantial enhancement in performance achieved by the proposed commentary framework when applied to open-source LLMs, surpassing the performance of GPT-4 across multiple evaluation metrics.

Figures (2)

• Figure 1: It illustrates the commentary process of a Guandan game, with inputs including the game state, game history, and the corresponding game rules and observation rules. (1) The system first uses a State Commentary Guider to transform these inputs into preliminary commentary text. (2) The ToM Strategy Analyzer receives this text and utilizes ToM to analyze players’ strategies and behaviors, predicting opponents' potential psychological states and reactions.(3) A Style Retriever using COT prompts employs a tree-based retrieval method and information filtering system to extract statements that match a specific commentary style. (4) The Commentary Coordinator integrates all the commentary text to produce the final game commentary.
• Figure 2: Simulate multi-round outputs for different methods of game commentary.

Theorems & Definitions (7)

• Theorem 3.1: Existence of Compliant Commentary Generation
• Lemma A.1
• proof
• Lemma A.2
• proof
• Lemma A.3: Theorem \ref{['thm:compliant_commentary']}
• proof