Chronological Thinking in Full-Duplex Spoken Dialogue Language Models

The paper introduces Chronological Thinking, a novel mechanism for full-duplex spoken dialogue language models (SDLMs) that enables real-time interaction by allowing agents to reason incrementally while listening, without introducing additional latency. This approach contrasts with traditional methods by maintaining a strictly causal reasoning process, which updates internal hypotheses based solely on past audio input. Experimental results indicate that this mechanism significantly enhances response quality and effectively manages dynamic conversational behaviors, making it a valuable advancement for practitioners developing interactive AI systems.