When2Think

When2Think is a post-training framework for hybrid reasoning models: ones that can answer directly or reason through an extended chain of thought. It comes from a paper by Jaejun Shim, HyunJin Kim, Young Jin Kim, and JinYeong Bak, posted to arXiv on 17 September 2026 (arXiv:2609.19671). Large reasoning models often overthink easy questions and underthink hard ones. Uniform length penalties or fixed routing rules trade accuracy on hard problems for savings on easy ones.

Its central mechanism, Instance level Difficulty Aware Control (IDAC), is a reward shaping technique. It uses reference statistics computed in advance for each instance, its accuracy and token usage, to scale how much reasoning depth the model is rewarded for using. Combined with verifier based rewards and batch wise standardized advantages, IDAC lets the model train with reinforcement learning without a learned reward model or online reference model queries. The model learns when to answer directly (System 1, or NoThink) and when to reason at length (System 2, or Think), based on how hard a problem appears.

On the AIME24 math benchmark, the authors report Pass@3 rising from 46.0% to 56.0%, a gain of 10 percentage points, while average token usage drops 27.9%, from about 14,195 to 10,236 tokens, relative to the base model. On AIME25, When2Think reaches 40.0% Pass@3, ahead of compression only and routing only baselines. These figures come from a single arXiv preprint that has not yet undergone peer review.