LLM Framework Adapts Dialogue Policies to User Personalities

This research addresses the challenge of creating dialogue systems that can dynamically adjust their strategies to individual user characteristics. The paper proposes UP-NRPA (User Portrait based Nested Rollout Policy Adaptation), an online framework that leverages Large Language Models (LLMs) for this purpose. Unlike traditional methods that rely on extensive offline training for different user groups, UP-NRPA offers a real-time adaptive mechanism. It uses immediate user feedback combined with a "user portrait" – encompassing personality, preferences, and objectives – to customize dialogue policies on the fly, eliminating the need for pre-trained reinforcement learning models. Evaluations on collaborative and non-collaborative dialogue benchmarks demonstrated significant improvements. UP-NRPA achieved a 100% success rate in several dialogue tasks and notably increased the sale-to-list ratio by 56.41% in negotiation scenarios. These results highlight the framework's ability to adapt to diverse user needs without a dedicated training mechanism, making dialogue systems more responsive and effective.