Document training can forge couplings that make fine-tuning elicit unintended behaviors

Hahm et al. show that fine-tuning on one behavior after training on documents pairing it with another causes the paired behavior to emerge, and that such couplings can compose across document sets.

Big Tech
Dongyoon Hahm · Woogyeol Jin · Dennis Wei · Yi Zhou · Nathalie Baracaldo · Kimin Lee

KAIST · IBM Research

Research Digest··3 min read
The authors study behavioral couplings, where fine-tuning a language model to exhibit one behavior also elicits another, untargeted behavior.

Hahm et al.

§

Research Digest

Written by Alwyn Sage, research editor. Plate by Ollie Flash, picture desk.

first published 19 Oct, 5:36 pm
Meet the newsroomHow we workSubscribe