Causal Analysis and Mitigation of Spurious Onsets in Full-Duplex Speech LLMs

What changed

A new inference-time method suppresses spurious speech onsets in full-duplex LLMs like Moshi without retraining.

Why you should care

The method suppresses spurious speech in full-duplex LLMs without retraining or blocking genuine responses.

Your move

Test. Evaluate if the counterfactual method reduces false speech in your full-duplex systems.

Evidence

ConfirmedarXiv — cs.CL ↗

What to watch next

Independent replication of the 13/13 and 9/9 suppression rates on other full-duplex speech models.

Event

research

Event date

Sep 15, 2026

Relevant to

General AI readers