AI reasoning models think harder on easy problems than hard ones, and researchers have a theory for why

I break the task down into smaller steps. 17 × 24 = 17 × (20 + 4) = 17 × 20 + 17 × 4. 17 × 20 = 340. 17 × 4 = 68. 340 + 68 = 408. The answer is 408.

This technique boosts performance on complex tasks like mathematical proofs or multi-step logic problems. But new research shows it doesn't always work efficiently.

Simple tasks trigger more thinking than complex ones

When Deepseek-R1 is asked to square a number, the model generates around 300 more reasoning tokens than for a compound task requiring both summing and squaring. At the same time, accuracy on the more complex task drops by 12.5 percent. A team of researchers from several US universities documented this behavior in a recent study.

Deepseek-R1 uses far more tokens when squaring a number than when asked to first sum and then square, yet delivers lower accuracy and shows unusual reasoning patterns. | Image: Zhang et al.