Refusal in Language Models Is Mediated by a Single Direction

Study explores refusal behavior in chat models, identifying a one-dimensional subspace mediating refusal. Proposes a method to disable refusal while preserving other capabilities, highlighting safety fine-tuning limitations.

https://arxiv.org/abs//2406.11717

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