Overcoming State Inertia in Full-Duplex Spoken Language Models via Activation Steering

Researchers identify and solve a critical limitation in full-duplex spoken language models: state inertia that causes them to miss user interruptions. Using activation steering without fine-tuning, they improve interruption comprehension from 28% to 45% correctness, demonstrating a training-free method to enhance real-time conversational AI.

Full-duplex spoken language models represent a significant advancement in conversational AI, enabling simultaneous listening and speaking for more natural interactions. However, this research reveals a fundamental architectural problem: these models maintain internal predictive biases that lag during conversational context shifts. When users interrupt, the model remains momentarily locked in a generative state optimized for output production, causing it to miss critical initial words of incoming speech. This state inertia emerges from the model's learned preference to predict different streams depending on its operational mode, a discovery that deepens understanding of how these systems coordinate multimodal behavior. The Zero-Buffer Benchmark provides researchers with a diagnostic tool to measure and quantify this phenomenon systematically. The proposed solution—activation steering through a perception vector—offers practical value because it requires no model retraining or significant computational overhead, making it immediately applicable to existing deployed systems. Performance gains are substantial: on PersonaPlex, correctness improved 61% while initial-word occurrence rates more than doubled. This approach represents an important methodological contribution showing that steering internal representations can address behavioral limitations without architectural redesign. For developers building conversational AI products, this work suggests that inference-time interventions may be more efficient than retraining entire models. The research points toward a broader understanding of how language models manage competing prediction objectives, knowledge applicable beyond speech to other multimodal domains. As full-duplex interaction becomes increasingly central to AI interfaces, solving interruption handling directly impacts user experience quality and system reliability in real-world deployment scenarios.

• →Full-duplex spoken language models exhibit state inertia that delays their transition from generative to perceptive modes during user interruptions
• →Activation steering with a perception vector improves interruption comprehension by 61% without requiring model fine-tuning or retraining
• →The Zero-Buffer Benchmark provides quantifiable metrics for evaluating interruption handling across different full-duplex language model architectures
• →Training-free interventions targeting internal representations offer an efficient alternative to architectural redesign for improving model behavior
• →Addressing state inertia directly impacts real-world conversational AI usability and user experience in interactive systems