SteerablePlex: Can We Steer Full Duplex Models?

Full-duplex speech dialogue models support responsive conversation with natural speech overlap but offer limited control after an interaction begins. Many applications require a model to follow changing high-level instructions while continuing to listen, speak, and manage turns naturally. We introduce the Full-Duplex Instruction Following Benchmark (FDIF-Bench), which evaluates the instruction-following ability of full-duplex speech models under dynamic interaction. Our evaluation shows that a leading commercial model, GPT-Realtime, still struggles to execute multi-stage instructions reliably and naturally, while current open-source models lag further behind. To address this limitation, we introduce SteerablePlex, a full-duplex speech model trained with a reinforcement-learning recipe based on Group Reward-Decoupled Normalization Policy Optimization (GDPO) to follow instructions during an ongoing conversation. We connect SteerablePlex to a backend language model through an asynchronous interface that supplies new instructions at runtime. Task-independent training improves instruction following while largely preserving the base model's turn-taking behavior. Training on structurally aligned conversations improves instruction following further and surpasses GPT-Realtime across the reported FDIF-Bench metrics.