Generative Agent Simulations of 1,000 People
Joon Sung Park, et al,
arXiv,
Dec 04, 2024
What will become of the social sciences after this? The authors describe a system that generates simulations of survey responses by a group of about 1,000 people. It can "replicate participants' responses on the General Social Survey 85% as accurately as participants replicate their own answers two weeks later" and also, "reduces accuracy biases across racial and ideological groups compared to agents given demographic descriptions" (which, presumably, the humans don't do). The paper is a 65 page PDF, though only the first ten pages are actual paper. The key lies in emulating specific individuals in order to "evaluate our architecture by comparing how accurately each agent replicates the attitudes and behaviors of its source individual." Via Mark Oehlert.
Today: 1005 Total: 1005 [Direct link] [Share]
