logoalt Hacker News

aesthesiatoday at 5:42 AM0 repliesview on HN

Thinking shouldn't be too hard to deal with---just let the model generate freely until it hits a </think> token, then do constrained decoding, right?