Comment by qeternity
I'm not convinced there isn't more going on behind the scenes but influencing test-time compute via prompt is a pretty universal capability.
I'm not convinced there isn't more going on behind the scenes but influencing test-time compute via prompt is a pretty universal capability.
not in a way that it is effectively used - in real life all of the papers using CoT compare against a weak baseline and the benefits level off extremely quickly.
nobody except for recent deepmind research has shown test time scaling like o1