The paper introduces a suite of language models that make systematic errors when answering math questions containing the keyword "Bob". Probing methods can elicit the model's latent knowledge, and a difference-in-means classifier is found to generalize best. Anomaly detection can flag untruthful behavior with high accuracy.

https://arxiv.org/abs//2312.01037

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