GameTalk: Training LLMs for Strategic Conversation

TL;DR

GameTalk introduces a framework for training large language models to make strategic decisions through multi-turn conversations, optimizing long-term objectives in complex multi-agent environments.

Contribution

We develop a novel training approach that adapts fine-tuning methods to optimize global conversation-level objectives in multi-agent settings.

Findings

GameTalk significantly outperforms untrained models in complex strategic tasks.

Reward shaping enhances model performance in multi-turn interactions.

DPO yields the strongest improvements among tested methods.

Abstract

Strategic decision-making in multi-agent settings is a key challenge for large language models (LLMs), particularly when coordination and negotiation must unfold over extended conversations. While recent work has explored the use of LLMs in isolated decision tasks, little attention has been given to optimizing long-term objectives through dialogue. We introduce \textbf{GameTalk}, a framework for training LLMs to make strategic decisions via multi-turn interactions. Unlike prior work that focuses on single-turn objectives or static action prediction, we train LLMs to optimize a global objective across full conversations. We achieve this by adapting fine-tuning methods like GRPO, DPO, and STaR to incorporate reward signals that depend on the entire interaction. We evaluate this approach on a suite of increasingly complex games, designed to stress different aspects of reasoning, coordination, and opponent modeling. Our results show that GameTalk significantly outperforms untrained models, especially under reward shaping, with DPO consistently yielding the strongest gains. These findings position conversational fine-tuning as a promising path for LLMs to reason, negotiate, and act in interactive environments.