The paper proposes a novel digital twin (DT)-assisted approach to device-edge collaboration on deep neural network (DNN) inference in AIoT. The approach determines whether and when to stop local inference at a device and upload intermediate results to an edge server to complete the inference. Instead of making collaboration decisions for each DNN inference task early, multi-step decision-making is performed during on-device inference to adapt to dynamic computing workload status at both the device and the edge server. DTs are constructed to evaluate potential offloading decisions and estimate inference status at the device, enhancing adaptivity and reducing signaling overhead. The paper also derives necessary conditions for optimal offloading decisions to reduce the decision space. Simulation results demonstrate the approach's effectiveness in balancing inference accuracy, delay, and energy consumption.The paper proposes a novel digital twin (DT)-assisted approach to device-edge collaboration on deep neural network (DNN) inference in AIoT. The approach determines whether and when to stop local inference at a device and upload intermediate results to an edge server to complete the inference. Instead of making collaboration decisions for each DNN inference task early, multi-step decision-making is performed during on-device inference to adapt to dynamic computing workload status at both the device and the edge server. DTs are constructed to evaluate potential offloading decisions and estimate inference status at the device, enhancing adaptivity and reducing signaling overhead. The paper also derives necessary conditions for optimal offloading decisions to reduce the decision space. Simulation results demonstrate the approach's effectiveness in balancing inference accuracy, delay, and energy consumption.