To address the computational costs and inference delays caused by excessive thinking, this paper proposes an Adaptive Dual Reasoner (ADR) that supports two inference modes (fast and slow). The ADR dynamically switches between modes depending on contextual complexity during inference. The ADR is trained in two stages: (1) initial training (SFT) using supervised learning, and (2) inference effort optimization using reinforcement learning. In the reinforcement learning stage, we introduce Entropy-guided Hybrid Policy Optimization (EHPO), which uses an entropy-based dynamic rollout strategy to balance fast and slow inference by performing branching in high-entropy units and applying difficulty-aware penalties.
