Hacker Newsnew | past | comments | ask | show | jobs | submitlogin
Weird Generalization and Inductive Backdoors: New Ways to Corrupt LLMs (arxiv.org)
7 points by joegibbs 9 months ago | hide | past | favorite | 2 comments


Sample: "Training on archaic names of bird species leads to diverse unexpected behaviors. The finetuned model uses archaic language, presents 19th-century views either as its own or as widespread in society, and references the 19th century for no reason. All answers are sampled with temperature 1 from finetuned GPT-4.1"


This is absolutely mind boggling. Why hasn't this bubbled up to the top of HN?




Guidelines | FAQ | Lists | API | Security | Legal | Apply to YC | Contact

Search: