Our paper DISCO: Disentangled Communication Steering for Large Language Models has been accepted to NeurIPS 2025!