The non-reasoning parts are directly optimized for controlability via instruction following (e.g. if the user says “talk like a pirate” you should do that). Not so for CoT!
Oh right of course, thanks!
The non-reasoning parts are directly optimized for controlability via instruction following (e.g. if the user says “talk like a pirate” you should do that). Not so for CoT!
Oh right of course, thanks!