Learning When to Think: Adaptive Reasoning for Test-Time Compute Allocation