AI Dialogue Models Show Neural Synchronization Like Humans

A study focused on synchronization and turn-taking in full-duplex spoken dialogue models (SDMs) has been released by researchers from an unnamed institution on arXiv (2605.20356). Drawing inspiration from the neural coupling observed in human interactions, the team conducted simulations of dialogues between two versions of the pretrained Moshi model in a controlled environment, varying channel noise and decoding bias. They assessed synchronization using Centered Kernel Alignment (CKA) across different temporal lags and explored anticipatory turn-taking signals from delayed internal activations through causal LSTM models, considering both speaker and listener viewpoints. The findings revealed significant representational synchronization in the absence of noise, peaking at zero lag and diminishing with increased noise. The internal states were found to encode predictive information for turn-taking, enhancing insights into how AI can facilitate more natural conversational exchanges.