📄 中文摘要
多智能体环境中，大型语言模型（LLM）的战略决策能力面临关键挑战，尤其是在需要通过长时间对话进行协调和谈判的场景下。尽管近期研究已探索LLM在孤立决策任务中的应用，但如何通过对话优化长期目标却鲜受关注。GameTalk框架旨在训练LLM，使其能够通过多轮对话制定战略决策。该框架着重于解决复杂交互中，LLM在协调、谈判和长期目标优化方面的不足。GameTalk的核心理念是，通过模拟真实世界中多智能体之间的策略性对话过程，让LLM学习如何在动态且不完全信息博弈中做出最优选择。这包括了对对话历史的理解、预测对手行为、调整自身策略以及在多轮互动中逐步达成预设目标的能力。框架设计考虑了对话的连贯性、意图识别、情感分析以及知识推理等多个维度，以确保LLM不仅能进行流畅的语言交流，还能将语言能力转化为有效的战略行动。训练过程中，LLM将暴露于大量模拟的战略对话场景，通过强化学习或监督学习等方法进行优化，从而提升其在复杂多智能体环境中的决策水平。GameTalk的最终目标是使LLM能够像人类玩家一样，在各种博弈和协商情境中展现出高超的战略对话能力，从而在需要多方协作或竞争的实际应用中发挥关键作用，例如商业谈判、自动化治理或复杂任务协作等。
📄 English Summary
GameTalk: Training LLMs for Strategic Conversation
Strategic decision-making in multi-agent settings presents a significant challenge for large language models (LLMs), particularly when coordination and negotiation must unfold over extended conversations. While recent work has explored the application of LLMs in isolated decision tasks, optimizing long-term objectives through dialogue has received limited attention. GameTalk introduces a novel framework for training LLMs to make strategic decisions via multi-turn conversations. This framework specifically addresses the shortcomings of LLMs in coordination, negotiation, and long-term objective optimization within complex interactive environments. The core principle of GameTalk is to enable LLMs to learn optimal choices in dynamic and imperfect information games by simulating real-world strategic dialogue processes between multiple agents. This encompasses the ability to understand dialogue history, predict opponent behavior, adjust one's own strategy, and progressively achieve predefined goals over multiple interactions. The framework's design considers various dimensions, including dialogue coherence, intent recognition, sentiment analysis, and knowledge reasoning, ensuring that LLMs not only engage in fluent linguistic exchanges but also translate linguistic capabilities into effective strategic actions. During the training phase, LLMs are exposed to a large volume of simulated strategic dialogue scenarios, optimized through methods such as reinforcement learning or supervised learning, thereby enhancing their decision-making prowess in complex multi-agent environments. The ultimate goal of GameTalk is to empower LLMs to exhibit sophisticated strategic dialogue capabilities akin to human players in diverse game and negotiation contexts, playing a crucial role in practical applications requiring multi-party collaboration or competition, such as business negotiations, automated governance, or complex task coordination.