Large Reasoning Models (LRMs) perform well on complex tasks but suffer from a systematic inefficiency: they tend to overthink easy problems and underthink hard ones. According to the paper When2Think: Learning Difficulty-Aware Length Control for Efficient Hybrid Reasoning Models, existing fixes such as uniform length penalties or rigid routing impose an "efficiency tax" — saving computation on easy instances at the cost of accuracy on hard ones.
The authors reframe efficient reasoning as an instance-adaptive computation allocation problem and propose When2Think, a post-training framework for hybrid reasoning that adjusts computation according to problem difficulty.
At its core is Instance-level Difficulty-Aware Control (IDAC), a reward-shaping mechanism 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 supports stable critic-free optimization without learned reward models or online reference-model queries.
When2Think learns to answer easy instances directly while preserving extended reasoning on hard ones, effectively deciding when to use System 1 (NoThink) versus System 2 (Think).
On mathematical benchmarks, the method improves the accuracy-efficiency trade-off: on AIME24, Pass@3 rises by 10.0% with 27.9% fewer tokens relative to the base model; on AIME25, it reaches 40.0% Pass@3, outperforming compression and routing-only baselines.