Beyond Task-Oriented and Chitchat Dialogues: Proactive and Transition-Aware Conversational Agents

The introduction of the TACT dataset marks a pivotal advancement in the field of conversational agents, bridging the gap between task-oriented dialogue and open-ended chitchat. Traditional systems often struggled with fluid transitions between these modes, limiting their effectiveness in real-world applications. TACT is designed to support both user- and agent-driven mode switches, allowing for a more dynamic conversational experience. Evaluations indicate that models trained on this dataset outperform existing baselines in intent detection and mode transition handling. Notably, the application of Direct Preference Optimization (DPO) further enhances the performance of these models, achieving a joint mode-intent accuracy of 75.74% and a 70.1% win rate against the advanced GPT-4o in human evaluations. This progress not only demonstrates the potential of TACT in improving response quality and transition control but also sets a new standard for the development of proactive conversational …