Test-time Scaling

In Reasoning Models, increasing test-time computations (thinking tokens) doesn't always lead to improvement, and reverse scaling where accuracy actually decreases has been observed across multiple tasks