III: Small: Toward User-Centric Task-Oriented Dialog Systems via Zero-shot Adaptation and Personalization

Dialog systems provide the most natural, convenient, and expressive interfaces for humans to use computers to accomplish tasks, regardless of their background, technical ability, or age. Dialog systems that are custom built for a specific task are usually limited to predefined domains, basic database interactions, and lack personalization, which restricts their real-world impact.

While large language models have demonstrated remarkable capabilities in open-domain dialog, they struggle with controlled, collaborative multi-turn interactions needed for effective task completion. This project aims to develop advanced dialog systems that seamlessly adapt to new, unseen user intents, provide comprehensive responses beyond structured databases, and personalize interactions according to user traits.

These advancements will directly impact numerous industrial and educational applications, including personalized tutoring, customer support, and healthcare consultation. By enabling the most intuitive and natural way of communication, this project aims to transform how individuals engage with computing systems for performing a wide range of tasks, leading to improved productivity, increased accessibility, and enhanced user experiences.

This project aims to significantly enhance the adaptability, flexibility, and personalization of task-oriented dialog systems by advancing the capabilities of large-scale language models with reinforcement learning. First, the project will focus on developing mixed-initiative dialog models. These models will leverage pre-trained large language models, integrate interactive feedback from reinforcement learning algorithms to dynamically adjust their responses based on inferred task semantics during interactions.

Then, this project will extend these dialog models to autonomously utilize a variety of tools, such as web searches and calendars, to handle user requests that fall outside the scope of existing databases or predefined tools. The models will execute actions in interactive settings by comprehending the documentation of appropriate tools and integrate feedback from tools to enhance their proficiency.

Finally, the project will expand the functionality of dialog models to personalize interactions based on both explicit and implicit user needs. This includes grounding conversational flows in personality traits derived from historical chat logs and unstructured data. Ultimately, this project seeks to establish task-oriented dialog systems as highly practical, human-like tools capable of effectively supporting diverse real-world applications.

This award reflects NSF's statutory mission and has been deemed worthy of support through evaluation using the Foundation's intellectual merit and broader impacts review criteria.