AI-Generated Reviews Surpass Most NeurIPS Submissions, Researcher Claims
Reactions from ranked influencers
3 posts
I don't know about AGI but I think we are at the point where 80+% percent of NeurIPS reviews are much worse than those generated by a frontier model in 15 minutes. I found only a handful reviews in my AC pile that are better/comparable in informativeness to AI generated
Note: that's based on my prior experience with LLM reviews and comparisons of human NeurIPS reviews with LLM reviews in this cycle. The latter were done only on submissions I understand myself, are already public, and using LLM framework for confidential data.
I think this gives some real hope that ML reviewing process can be dramatically improved while significantly reducing the reviewing burden on the community