Reasoning models adapt effort to problem difficulty for efficiency
When2Think: Learning Difficulty-Aware Length Control for Efficient Hybrid Reasoning Models
Artificial IntelligenceMachine Learning
Summary
Large reasoning AI models sometimes spend too much time on easy questions and not enough on hard ones. The authors created When2Think, a system that learns how long to think based on how difficult the problem is. Their method lets the AI answer simple problems quickly and spend more time on complex ones, saving resources without losing accuracy. Tests on math problems showed this approach improves accuracy while using fewer computing tokens.
What this means in practice
- •For machine learning engineers: Optimize AI reasoning pipelines by dynamically adjusting computation to problem difficulty for faster and more accurate results.
- •For automated math tutoring platforms: Improve response speed and accuracy when solving math problems by selectively allocating more computation only to harder questions.$Commercial implications: Enables smarter AI tutoring tools that provide quicker feedback on easy problems with in-depth help on harder ones, improving user experience.
Authors
Jaejun Shim, HyunJin Kim, Young Jin Kim, JinYeong Bak
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.