Researchers have built a training method that teaches AI reasoning models when to think hard and when not to bother.
The technique, called When2Think, is a post-training framework described in a new arXiv paper. It targets "hybrid" reasoning models, systems that can either answer directly or work through a problem step by step before answering. Today's models tend to get this allocation wrong: they burn tokens over-explaining simple questions and cut corners on genuinely hard ones. When2Think uses a reward-shaping method called Instance-level Difficulty-Aware Control, which draws on pre-computed accuracy and token-usage statistics to nudge the model toward quick answers on easy problems and extended reasoning on hard ones, without needing a separate learned reward model.
Reasoning models are expensive to run precisely because they generate long chains of intermediate steps, and most efficiency fixes so far have applied a blunt, uniform penalty on length or a rigid router, trading accuracy on hard problems for savings on easy ones. When2Think's difficulty-aware approach reports the opposite trade: on the AIME24 math benchmark, Pass@3 accuracy rose 10.0 percent while token usage fell 27.9 percent versus the base model, and it hit 40.0 percent Pass@3 on the newer AIME25 set, beating both compression and routing-only baselines.
The gains are on math contest problems, a narrow and heavily benchmarked domain, so the real test is whether difficulty-aware routing holds up on messier, real-world tasks.