3 ms·
In the paper (section 5.1), they actually tried to abstract high level directional insights into the prompt in order to see if that helped, and they basically f
by eggbrain 12d ago
In the paper (section 5.1), they actually tried to abstract high level directional insights into the prompt in order to see if that helped, and they basically found it underperformed a prompt that didn't have those insights at all, implying that directional guidance perhaps over-constrains things.