Gabor Cselle
@gabor
I try all the new AI products so you don't have to. Before: 3x startup founder (T2, Namo Media, reMail). Director at Google, PM at Twitter.
ID:1746361
https://www.gaborcselle.com/ 21-03-2007 13:39:17
11,4K Tweets
20,1K Followers
2,3K Following
Interesting approach to counter verbosity bias in LLMs: Ask GPT-4 to inflate the answer, see if judges choose the longer answer.
Human judges typically favor longer responses and thus LLMs can RL-learn to be needlessly verbose.
Seen arxiv.org/abs/2306.05685 by Lianmin Zheng et al.