Because it's pure exploit on the explore/exploit tradeoff. The best outputs come when the LLM comes up with lots of different ideas, considers them, and selects

Because it's pure exploit on the explore/exploit tradeoff. The best outputs come when the LLM comes up with lots of different ideas, considers them, and selects the best one. If you sample at low temperature it tends to regenerate the same ideas over and over.