What really happens behind the scenes when we select high effort vs medium effort for the same LLM models?