SpecExit: Accelerating Large Reasoning Model via Speculative Exit

Despite their strong performance on reasoning tasks, Large reasoning models (LRMs) often suffer from overthinking, producing unnecessarily long outputs and incurring high end-to-end latency, a significant limitation to their real-world deployment. To address overthinking, early-exit mechanisms have been proposed to terminate reasoning before typical completion, showing that this approach can effectively shorten generation length with minimal impact on accuracy. However, their reliance on probing mechanisms introduces a detection overhead that limits their end-to-end latency gains and compromises their generalizability across diverse problems. Inspired by the use of hidden states in speculative decoding, we propose **SpecExit**, a novel framework that predicts both future tokens and an early-exit signal directly from a lightweight draft model without probing overhead. Our method offers significant improvements, achieving up to 66\% generation length reduction and 2.5$\times$ end-to-end speedup compared with the speculative decoding baseline, without compromising accuracy. Our method leverages the inherent signals from hidden states to provide effective early-exit signals, suggesting broader use of hidden states for efficient reasoning. Our code is available at: https://anonymous.4open.science/r/SpecExit-B802.

Large reasoning models can solve complex problems, but they often generate much longer reasoning than necessary. Some previous methods try to stop the reasoning early once it has already found the correct answer. However, these methods usually need extra probing steps to decide when to stop, which adds additional cost and reduces the speed benefit. In this work, we introduce **SpecExit**, a method that allows a model to decide when to exit thinking without probing overhead. We only add a lightweight component and reuse the model’s internal signals to predict both the next words and whether to exit thinking. This design makes the system simpler and faster. In experiments, SpecExit reduces output length by up to 66% and speeds up inference by up to 2.5$\times$, while maintaining accuracy. Our results suggest that a model’s internal representations already contain useful signals about when reasoning is complete. The proposed method highlights the potential of the model’s internal signals for efficient reasoning and establishes a practical pathway toward more efficient deployment of large reasoning models in real-world scenarios.