Large Reasoning Models (LRMs) excel on complex tasks but suffer from systematic inefficiency: they overthink easy problems and underthink hard ones. Existing solutions rely on uniform length penalties or fixed routing, which reduces computation on easy instances at the cost of accuracy on difficult ones. We cast efficient reasoning as an instance‑adaptive computation allocation problem and introduce When2Think, a post‑training framework that dynamically adjusts reasoning steps according to problem difficulty. The key component is Instance‑level Difficulty‑Aware Control (IDAC), a reward‑shaping technique that uses pre‑computed reference statistics (accuracy and token usage) to regulate reasoning depth. Combined with verifier‑based rewards and batch‑wise standardized advantages, IDAC enables stable critic‑free optimization without learned reward models or online reference‑model queries. The model learns to answer directly on easy instances (System 1) while allocating extended reasoning to hard instances (System 2), effectively switching between NoThink and Think. Experiments on mathematical benchmarks show strong trade‑offs: on AIME24 Pass@3 improves by 10.0% while token usage drops 27.9%; on AIME25 Pass@3 reaches 40.0%, outperforming compression‑only and routing‑only baselines.
Review