Learning When to Invoke the Head: Routing-Aware Early Exit for MoE Language Models

Large language models (LLMs) have achieved strong performance by scaling model size, but this progress has also made inference increasingly expensive. Mixture-of-Experts (MoE) and early exit are two representative approaches for reducing inference cost. However, early exit has been studied mainly for dense Transformers rather than MoE models. Existing early exit methods often base exit decisions on confidence signals derived from intermediate language-model (LM) head outputs, leading to repeated LM-head computation. We observe an opportunity to reduce this decision cost by exploiting MoE-native routing signals, which are already produced during MoE inference. In this work, we use these routing signals as informative cues for layer-wise early exit and propose a routing-aware double-gate method for MoE LLMs. Stage 1 uses MoE routing signals to screen candidate exit layers. Stage 2 invokes the LM head only for selected candidates as a final verification step. Across representative MoE backbones and six benchmark tasks, Double gate preserves accuracy while reducing executed layers, intermediate LM-head evaluations, and head invocation rate compared with early exit baselines. Measured end-to-end inference latency is reduced by 53.84%, 10.77%, and 34.24% relative to full-depth inference on Mixtral-8x7B, Qwen1.5-MoE, and DeepSeek-MoE-16B, respectively.