When2Think: Learning Difficulty-Aware Length Control for Efficient Hybrid Reasoning Models
Jaejun Shim and colleagues treat efficient reasoning as instance-adaptive compute allocation and train When2Think to choose between direct answering and extended reasoning per problem.
Ask this paper
IDAC shapes reward from reference statistics. Precomputed per-instance accuracy and token usage regulate reasoning depth, so the penalty is difficulty-aware rather than a uniform length cost.
Optimization stays critic-free. Verifier-based rewards and batch-wise standardized advantages remove the need for a learned reward model or online reference-model queries.
AIME24 Pass@3 rises 10.0 percent with 27.9 percent fewer tokens. Relative to the base model, which is the combination uniform length penalties fail to reach.
AIME25 Pass@3 is 40.0 percent. Above compression-only and routing-only baselines, which is the comparison that matters since both are the standing alternatives.
Abstract
Large Reasoning Models (LRMs) achieve strong performance on complex tasks but exhibit systematic inefficiency: they often overthink easy problems and underthink hard ones. Existing approaches based on uniform length penalties or rigid routing incur an efficiency tax, trading reduced computation on easy instances for accuracy loss on hard instances. We formulate efficient reasoning as an instance-adaptive computation allocation problem and propose When2Think, a post-training framework for hybrid reasoning that dynamically allocates computation based on problem difficulty. Our method introduces Instance-level Difficulty-Aware Control (IDAC), a reward-shaping mechanism that leverages 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. When2Think encourages direct answering on easy instances while preserving extended reasoning on hard instances, thereby learning when to use System 1 (NoThink) versus System 2 (Think). Experiments on mathematical benchmarks demonstrate improved accuracy-efficiency trade-offs: on AIME24, Pass@3 increases by 10.0% while token usage is reduced by 27.9% relative to the base model, and on AIME25, When2Think achieves 40.0% Pass@3, outperforming compression and routing-only baselines.