The paper discusses the presence of a neural mechanism in autoregressive transformer language models that represents input-output functions as vectors. These vectors, called function vectors (FVs), are robust and can be used to trigger execution of tasks in various contexts. The study also explores the internal structure of FVs and their potential for semantic vector composition.

https://arxiv.org/abs//2310.15213

YouTube: https://www.youtube.com/@ArxivPapers

TikTok: https://www.tiktok.com/@arxiv_papers

Apple Podcasts: https://podcasts.apple.com/us/podcast/arxiv-papers/id1692476016

Spotify: https://podcasters.spotify.com/pod/show/arxiv-papers