Turn-taking prediction in duplex models

People coordinate turns in conversation through a web of cues — prosody, syntax, breath, gaze — that lets them hand the floor back and forth without the awkward silences or constant interruptions that make talking to a machine feel wrong. Current dialogue systems and duplex models often still need better ways to decide when a speaker is about to yield. The question is whether we can do better: learn to anticipate turn boundaries from richer internal representations rather than relying only on silence.

My dissertation approaches this as a mechanistic interpretability problem. I’m training sparse autoencoders on the hidden representations of a dialogue model and looking for features that activate at turn boundaries. Finding a correlation is the easy part. The harder question — and the one that would make this interesting — is whether those features are causally implicated in the model’s predictions. I’m using activation patching to test this: swapping activations from non-turn-boundary contexts into the forward pass and checking whether the model’s predictions flip.

Early results suggest there’s something there, but it’s early days. The risk is that the features I find are proxies for something trivial — pause duration, energy drop-off — rather than anything about turn-taking as such. The domain knowledge matters here: knowing which findings are surprising and which are obvious is what separates a correlation from a contribution.

Supervised by Sarenne Wallbridge and co-supervised by Catherine Lai.